of human problem solving

• Bipolar Disorder
• Therapy Center
• When To See a Therapist
• Types of Therapy
• Best Online Therapy
• Best Couples Therapy
• Best Family Therapy
• Managing Stress
• Sleep and Dreaming
• Understanding Emotions
• Self-Improvement
• Healthy Relationships
• Student Resources
• Personality Types
• Guided Meditations
• Verywell Mind Insights
• 2023 Verywell Mind 25
• Mental Health in the Classroom
• Editorial Process
• Meet Our Review Board
• Crisis Support

Overview of the Problem-Solving Mental Process

Kendra Cherry, MS, is a psychosocial rehabilitation specialist, psychology educator, and author of the "Everything Psychology Book."

Rachel Goldman, PhD FTOS, is a licensed psychologist, clinical assistant professor, speaker, wellness expert specializing in eating behaviors, stress management, and health behavior change.

• Identify the Problem
• Define the Problem
• Form a Strategy
• Organize Information
• Allocate Resources
• Monitor Progress
• Evaluate the Results

Frequently Asked Questions

Problem-solving is a mental process that involves discovering, analyzing, and solving problems. The ultimate goal of problem-solving is to overcome obstacles and find a solution that best resolves the issue.

The best strategy for solving a problem depends largely on the unique situation. In some cases, people are better off learning everything they can about the issue and then using factual knowledge to come up with a solution. In other instances, creativity and insight are the best options.

It is not necessary to follow problem-solving steps sequentially, It is common to skip steps or even go back through steps multiple times until the desired solution is reached.

In order to correctly solve a problem, it is often important to follow a series of steps. Researchers sometimes refer to this as the problem-solving cycle. While this cycle is portrayed sequentially, people rarely follow a rigid series of steps to find a solution.

The following steps include developing strategies and organizing knowledge.

1. Identifying the Problem

While it may seem like an obvious step, identifying the problem is not always as simple as it sounds. In some cases, people might mistakenly identify the wrong source of a problem, which will make attempts to solve it inefficient or even useless.

Some strategies that you might use to figure out the source of a problem include :

• Asking questions about the problem
• Breaking the problem down into smaller pieces
• Looking at the problem from different perspectives
• Conducting research to figure out what relationships exist between different variables

2. Defining the Problem

After the problem has been identified, it is important to fully define the problem so that it can be solved. You can define a problem by operationally defining each aspect of the problem and setting goals for what aspects of the problem you will address

At this point, you should focus on figuring out which aspects of the problems are facts and which are opinions. State the problem clearly and identify the scope of the solution.

3. Forming a Strategy

After the problem has been identified, it is time to start brainstorming potential solutions. This step usually involves generating as many ideas as possible without judging their quality. Once several possibilities have been generated, they can be evaluated and narrowed down.

The next step is to develop a strategy to solve the problem. The approach used will vary depending upon the situation and the individual's unique preferences. Common problem-solving strategies include heuristics and algorithms.

• Heuristics are mental shortcuts that are often based on solutions that have worked in the past. They can work well if the problem is similar to something you have encountered before and are often the best choice if you need a fast solution.
• Algorithms are step-by-step strategies that are guaranteed to produce a correct result. While this approach is great for accuracy, it can also consume time and resources.

Heuristics are often best used when time is of the essence, while algorithms are a better choice when a decision needs to be as accurate as possible.

4. Organizing Information

Before coming up with a solution, you need to first organize the available information. What do you know about the problem? What do you not know? The more information that is available the better prepared you will be to come up with an accurate solution.

When approaching a problem, it is important to make sure that you have all the data you need. Making a decision without adequate information can lead to biased or inaccurate results.

5. Allocating Resources

Of course, we don't always have unlimited money, time, and other resources to solve a problem. Before you begin to solve a problem, you need to determine how high priority it is.

If it is an important problem, it is probably worth allocating more resources to solving it. If, however, it is a fairly unimportant problem, then you do not want to spend too much of your available resources on coming up with a solution.

At this stage, it is important to consider all of the factors that might affect the problem at hand. This includes looking at the available resources, deadlines that need to be met, and any possible risks involved in each solution. After careful evaluation, a decision can be made about which solution to pursue.

6. Monitoring Progress

After selecting a problem-solving strategy, it is time to put the plan into action and see if it works. This step might involve trying out different solutions to see which one is the most effective.

It is also important to monitor the situation after implementing a solution to ensure that the problem has been solved and that no new problems have arisen as a result of the proposed solution.

Effective problem-solvers tend to monitor their progress as they work towards a solution. If they are not making good progress toward reaching their goal, they will reevaluate their approach or look for new strategies .

7. Evaluating the Results

After a solution has been reached, it is important to evaluate the results to determine if it is the best possible solution to the problem. This evaluation might be immediate, such as checking the results of a math problem to ensure the answer is correct, or it can be delayed, such as evaluating the success of a therapy program after several months of treatment.

Once a problem has been solved, it is important to take some time to reflect on the process that was used and evaluate the results. This will help you to improve your problem-solving skills and become more efficient at solving future problems.

A Word From Verywell​

It is important to remember that there are many different problem-solving processes with different steps, and this is just one example. Problem-solving in real-world situations requires a great deal of resourcefulness, flexibility, resilience, and continuous interaction with the environment.

Get Advice From The Verywell Mind Podcast

Hosted by therapist Amy Morin, LCSW, this episode of The Verywell Mind Podcast shares how you can stop dwelling in a negative mindset.

Follow Now : Apple Podcasts / Spotify / Google Podcasts

You can become a better problem solving by:

• Practicing brainstorming and coming up with multiple potential solutions to problems
• Being open-minded and considering all possible options before making a decision
• Breaking down problems into smaller, more manageable pieces
• Asking for help when needed
• Researching different problem-solving techniques and trying out new ones
• Learning from mistakes and using them as opportunities to grow

It's important to communicate openly and honestly with your partner about what's going on. Try to see things from their perspective as well as your own. Work together to find a resolution that works for both of you. Be willing to compromise and accept that there may not be a perfect solution.

Take breaks if things are getting too heated, and come back to the problem when you feel calm and collected. Don't try to fix every problem on your own—consider asking a therapist or counselor for help and insight.

If you've tried everything and there doesn't seem to be a way to fix the problem, you may have to learn to accept it. This can be difficult, but try to focus on the positive aspects of your life and remember that every situation is temporary. Don't dwell on what's going wrong—instead, think about what's going right. Find support by talking to friends or family. Seek professional help if you're having trouble coping.

Davidson JE, Sternberg RJ, editors.  The Psychology of Problem Solving .  Cambridge University Press; 2003. doi:10.1017/CBO9780511615771

Sarathy V. Real world problem-solving .  Front Hum Neurosci . 2018;12:261. Published 2018 Jun 26. doi:10.3389/fnhum.2018.00261

By Kendra Cherry, MSEd Kendra Cherry, MS, is a psychosocial rehabilitation specialist, psychology educator, and author of the "Everything Psychology Book."

• Search Menu
• Browse content in Arts and Humanities
• Browse content in Archaeology
• Anglo-Saxon and Medieval Archaeology
• Archaeological Methodology and Techniques
• Archaeology by Region
• Archaeology of Religion
• Archaeology of Trade and Exchange
• Biblical Archaeology
• Contemporary and Public Archaeology
• Environmental Archaeology
• Historical Archaeology
• History and Theory of Archaeology
• Industrial Archaeology
• Landscape Archaeology
• Mortuary Archaeology
• Prehistoric Archaeology
• Underwater Archaeology
• Urban Archaeology
• Zooarchaeology
• Browse content in Architecture
• Architectural Structure and Design
• History of Architecture
• Residential and Domestic Buildings
• Theory of Architecture
• Browse content in Art
• Art Subjects and Themes
• History of Art
• Industrial and Commercial Art
• Theory of Art
• Biographical Studies
• Byzantine Studies
• Browse content in Classical Studies
• Classical Literature
• Classical Reception
• Classical History
• Classical Philosophy
• Classical Mythology
• Classical Art and Architecture
• Classical Oratory and Rhetoric
• Greek and Roman Archaeology
• Greek and Roman Papyrology
• Greek and Roman Epigraphy
• Greek and Roman Law
• Late Antiquity
• Religion in the Ancient World
• Digital Humanities
• Browse content in History
• Colonialism and Imperialism
• Diplomatic History
• Environmental History
• Genealogy, Heraldry, Names, and Honours
• Genocide and Ethnic Cleansing
• Historical Geography
• History by Period
• History of Agriculture
• History of Education
• History of Emotions
• History of Gender and Sexuality
• Industrial History
• Intellectual History
• International History
• Labour History
• Legal and Constitutional History
• Local and Family History
• Maritime History
• Military History
• National Liberation and Post-Colonialism
• Oral History
• Political History
• Public History
• Regional and National History
• Revolutions and Rebellions
• Slavery and Abolition of Slavery
• Social and Cultural History
• Theory, Methods, and Historiography
• Urban History
• World History
• Browse content in Language Teaching and Learning
• Language Learning (Specific Skills)
• Language Teaching Theory and Methods
• Browse content in Linguistics
• Applied Linguistics
• Cognitive Linguistics
• Computational Linguistics
• Forensic Linguistics
• Grammar, Syntax and Morphology
• Historical and Diachronic Linguistics
• History of English
• Language Variation
• Language Families
• Language Evolution
• Language Reference
• Language Acquisition
• Lexicography
• Linguistic Theories
• Linguistic Typology
• Linguistic Anthropology
• Phonetics and Phonology
• Psycholinguistics
• Sociolinguistics
• Translation and Interpretation
• Writing Systems
• Browse content in Literature
• Bibliography
• Children's Literature Studies
• Literary Studies (Modernism)
• Literary Studies (Romanticism)
• Literary Studies (American)
• Literary Studies (Asian)
• Literary Studies (European)
• Literary Studies (Eco-criticism)
• Literary Studies - World
• Literary Studies (1500 to 1800)
• Literary Studies (19th Century)
• Literary Studies (20th Century onwards)
• Literary Studies (African American Literature)
• Literary Studies (British and Irish)
• Literary Studies (Early and Medieval)
• Literary Studies (Fiction, Novelists, and Prose Writers)
• Literary Studies (Gender Studies)
• Literary Studies (Graphic Novels)
• Literary Studies (History of the Book)
• Literary Studies (Plays and Playwrights)
• Literary Studies (Poetry and Poets)
• Literary Studies (Postcolonial Literature)
• Literary Studies (Queer Studies)
• Literary Studies (Science Fiction)
• Literary Studies (Travel Literature)
• Literary Studies (War Literature)
• Literary Studies (Women's Writing)
• Literary Theory and Cultural Studies
• Mythology and Folklore
• Shakespeare Studies and Criticism
• Browse content in Media Studies
• Browse content in Music
• Applied Music
• Dance and Music
• Ethics in Music
• Ethnomusicology
• Gender and Sexuality in Music
• Medicine and Music
• Music Cultures
• Music and Culture
• Music and Media
• Music and Religion
• Music Education and Pedagogy
• Music Theory and Analysis
• Musical Scores, Lyrics, and Libretti
• Musical Structures, Styles, and Techniques
• Musicology and Music History
• Performance Practice and Studies
• Race and Ethnicity in Music
• Sound Studies
• Browse content in Performing Arts
• Browse content in Philosophy
• Aesthetics and Philosophy of Art
• Epistemology
• Feminist Philosophy
• History of Western Philosophy
• Metaphysics
• Moral Philosophy
• Non-Western Philosophy
• Philosophy of Action
• Philosophy of Law
• Philosophy of Religion
• Philosophy of Language
• Philosophy of Mind
• Philosophy of Perception
• Philosophy of Science
• Philosophy of Mathematics and Logic
• Practical Ethics
• Social and Political Philosophy
• Browse content in Religion
• Biblical Studies
• Christianity
• East Asian Religions
• History of Religion
• Judaism and Jewish Studies
• Qumran Studies
• Religion and Education
• Religion and Health
• Religion and Politics
• Religion and Science
• Religion and Law
• Religion and Art, Literature, and Music
• Religious Studies
• Browse content in Society and Culture
• Cookery, Food, and Drink
• Cultural Studies
• Customs and Traditions
• Ethical Issues and Debates
• Hobbies, Games, Arts and Crafts
• Lifestyle, Home, and Garden
• Natural world, Country Life, and Pets
• Popular Beliefs and Controversial Knowledge
• Sports and Outdoor Recreation
• Technology and Society
• Travel and Holiday
• Visual Culture
• Browse content in Law
• Arbitration
• Browse content in Company and Commercial Law
• Commercial Law
• Company Law
• Browse content in Comparative Law
• Systems of Law
• Competition Law
• Browse content in Constitutional and Administrative Law
• Government Powers
• Judicial Review
• Local Government Law
• Military and Defence Law
• Parliamentary and Legislative Practice
• Construction Law
• Contract Law
• Browse content in Criminal Law
• Criminal Procedure
• Criminal Evidence Law
• Sentencing and Punishment
• Employment and Labour Law
• Environment and Energy Law
• Browse content in Financial Law
• Banking Law
• Insolvency Law
• History of Law
• Human Rights and Immigration
• Intellectual Property Law
• Browse content in International Law
• Private International Law and Conflict of Laws
• Public International Law
• IT and Communications Law
• Jurisprudence and Philosophy of Law
• Law and Society
• Law and Politics
• Browse content in Legal System and Practice
• Courts and Procedure
• Legal Skills and Practice
• Primary Sources of Law
• Regulation of Legal Profession
• Medical and Healthcare Law
• Browse content in Policing
• Criminal Investigation and Detection
• Police and Security Services
• Police Procedure and Law
• Police Regional Planning
• Browse content in Property Law
• Personal Property Law
• Study and Revision
• Terrorism and National Security Law
• Browse content in Trusts Law
• Wills and Probate or Succession
• Browse content in Medicine and Health
• Browse content in Allied Health Professions
• Arts Therapies
• Clinical Science
• Dietetics and Nutrition
• Occupational Therapy
• Operating Department Practice
• Physiotherapy
• Radiography
• Speech and Language Therapy
• Browse content in Anaesthetics
• General Anaesthesia
• Neuroanaesthesia
• Clinical Neuroscience
• Browse content in Clinical Medicine
• Acute Medicine
• Cardiovascular Medicine
• Clinical Genetics
• Clinical Pharmacology and Therapeutics
• Dermatology
• Endocrinology and Diabetes
• Gastroenterology
• Genito-urinary Medicine
• Geriatric Medicine
• Infectious Diseases
• Medical Oncology
• Medical Toxicology
• Pain Medicine
• Palliative Medicine
• Rehabilitation Medicine
• Respiratory Medicine and Pulmonology
• Rheumatology
• Sleep Medicine
• Sports and Exercise Medicine
• Community Medical Services
• Critical Care
• Emergency Medicine
• Forensic Medicine
• Haematology
• History of Medicine
• Medical Ethics
• Browse content in Medical Skills
• Clinical Skills
• Communication Skills
• Nursing Skills
• Surgical Skills
• Browse content in Medical Dentistry
• Oral and Maxillofacial Surgery
• Paediatric Dentistry
• Restorative Dentistry and Orthodontics
• Surgical Dentistry
• Medical Statistics and Methodology
• Browse content in Neurology
• Clinical Neurophysiology
• Neuropathology
• Nursing Studies
• Browse content in Obstetrics and Gynaecology
• Gynaecology
• Occupational Medicine
• Ophthalmology
• Otolaryngology (ENT)
• Browse content in Paediatrics
• Neonatology
• Browse content in Pathology
• Chemical Pathology
• Clinical Cytogenetics and Molecular Genetics
• Histopathology
• Medical Microbiology and Virology
• Patient Education and Information
• Browse content in Pharmacology
• Psychopharmacology
• Browse content in Popular Health
• Caring for Others
• Complementary and Alternative Medicine
• Self-help and Personal Development
• Browse content in Preclinical Medicine
• Cell Biology
• Molecular Biology and Genetics
• Reproduction, Growth and Development
• Primary Care
• Professional Development in Medicine
• Browse content in Psychiatry
• Addiction Medicine
• Child and Adolescent Psychiatry
• Forensic Psychiatry
• Learning Disabilities
• Old Age Psychiatry
• Psychotherapy
• Browse content in Public Health and Epidemiology
• Epidemiology
• Public Health
• Browse content in Radiology
• Clinical Radiology
• Interventional Radiology
• Nuclear Medicine
• Radiation Oncology
• Reproductive Medicine
• Browse content in Surgery
• Cardiothoracic Surgery
• Gastro-intestinal and Colorectal Surgery
• General Surgery
• Neurosurgery
• Paediatric Surgery
• Peri-operative Care
• Plastic and Reconstructive Surgery
• Surgical Oncology
• Transplant Surgery
• Trauma and Orthopaedic Surgery
• Vascular Surgery
• Browse content in Science and Mathematics
• Browse content in Biological Sciences
• Aquatic Biology
• Biochemistry
• Bioinformatics and Computational Biology
• Developmental Biology
• Ecology and Conservation
• Evolutionary Biology
• Genetics and Genomics
• Microbiology
• Molecular and Cell Biology
• Natural History
• Plant Sciences and Forestry
• Research Methods in Life Sciences
• Structural Biology
• Systems Biology
• Zoology and Animal Sciences
• Browse content in Chemistry
• Analytical Chemistry
• Computational Chemistry
• Crystallography
• Environmental Chemistry
• Industrial Chemistry
• Inorganic Chemistry
• Materials Chemistry
• Medicinal Chemistry
• Mineralogy and Gems
• Organic Chemistry
• Physical Chemistry
• Polymer Chemistry
• Study and Communication Skills in Chemistry
• Theoretical Chemistry
• Browse content in Computer Science
• Artificial Intelligence
• Computer Architecture and Logic Design
• Game Studies
• Human-Computer Interaction
• Mathematical Theory of Computation
• Programming Languages
• Software Engineering
• Systems Analysis and Design
• Virtual Reality
• Browse content in Computing
• Business Applications
• Computer Games
• Computer Security
• Computer Networking and Communications
• Digital Lifestyle
• Graphical and Digital Media Applications
• Operating Systems
• Browse content in Earth Sciences and Geography
• Atmospheric Sciences
• Environmental Geography
• Geology and the Lithosphere
• Maps and Map-making
• Meteorology and Climatology
• Oceanography and Hydrology
• Palaeontology
• Physical Geography and Topography
• Regional Geography
• Soil Science
• Urban Geography
• Browse content in Engineering and Technology
• Agriculture and Farming
• Biological Engineering
• Civil Engineering, Surveying, and Building
• Electronics and Communications Engineering
• Energy Technology
• Engineering (General)
• Environmental Science, Engineering, and Technology
• History of Engineering and Technology
• Mechanical Engineering and Materials
• Technology of Industrial Chemistry
• Transport Technology and Trades
• Browse content in Environmental Science
• Applied Ecology (Environmental Science)
• Conservation of the Environment (Environmental Science)
• Environmental Sustainability
• Environmentalist Thought and Ideology (Environmental Science)
• Management of Land and Natural Resources (Environmental Science)
• Natural Disasters (Environmental Science)
• Nuclear Issues (Environmental Science)
• Pollution and Threats to the Environment (Environmental Science)
• Social Impact of Environmental Issues (Environmental Science)
• History of Science and Technology
• Browse content in Materials Science
• Ceramics and Glasses
• Composite Materials
• Metals, Alloying, and Corrosion
• Nanotechnology
• Browse content in Mathematics
• Applied Mathematics
• Biomathematics and Statistics
• History of Mathematics
• Mathematical Education
• Mathematical Finance
• Mathematical Analysis
• Numerical and Computational Mathematics
• Probability and Statistics
• Pure Mathematics
• Browse content in Neuroscience
• Cognition and Behavioural Neuroscience
• Development of the Nervous System
• Disorders of the Nervous System
• History of Neuroscience
• Invertebrate Neurobiology
• Molecular and Cellular Systems
• Neuroendocrinology and Autonomic Nervous System
• Neuroscientific Techniques
• Sensory and Motor Systems
• Browse content in Physics
• Astronomy and Astrophysics
• Atomic, Molecular, and Optical Physics
• Biological and Medical Physics
• Classical Mechanics
• Computational Physics
• Condensed Matter Physics
• Electromagnetism, Optics, and Acoustics
• History of Physics
• Mathematical and Statistical Physics
• Measurement Science
• Nuclear Physics
• Particles and Fields
• Plasma Physics
• Quantum Physics
• Relativity and Gravitation
• Semiconductor and Mesoscopic Physics
• Browse content in Psychology
• Affective Sciences
• Clinical Psychology
• Cognitive Neuroscience
• Cognitive Psychology
• Criminal and Forensic Psychology
• Developmental Psychology
• Educational Psychology
• Evolutionary Psychology
• Health Psychology
• History and Systems in Psychology
• Music Psychology
• Neuropsychology
• Organizational Psychology
• Psychological Assessment and Testing
• Psychology of Human-Technology Interaction
• Psychology Professional Development and Training
• Research Methods in Psychology
• Social Psychology
• Browse content in Social Sciences
• Browse content in Anthropology
• Anthropology of Religion
• Human Evolution
• Medical Anthropology
• Physical Anthropology
• Regional Anthropology
• Social and Cultural Anthropology
• Theory and Practice of Anthropology
• Browse content in Business and Management
• Business History
• Business Ethics
• Business Strategy
• Business and Technology
• Business and Government
• Business and the Environment
• Comparative Management
• Corporate Governance
• Corporate Social Responsibility
• Entrepreneurship
• Health Management
• Human Resource Management
• Industrial and Employment Relations
• Industry Studies
• Information and Communication Technologies
• International Business
• Knowledge Management
• Management and Management Techniques
• Operations Management
• Organizational Theory and Behaviour
• Pensions and Pension Management
• Public and Nonprofit Management
• Strategic Management
• Supply Chain Management
• Browse content in Criminology and Criminal Justice
• Criminal Justice
• Criminology
• Forms of Crime
• International and Comparative Criminology
• Youth Violence and Juvenile Justice
• Development Studies
• Browse content in Economics
• Agricultural, Environmental, and Natural Resource Economics
• Asian Economics
• Behavioural Finance
• Behavioural Economics and Neuroeconomics
• Econometrics and Mathematical Economics
• Economic Methodology
• Economic History
• Economic Systems
• Economic Development and Growth
• Financial Markets
• Financial Institutions and Services
• General Economics and Teaching
• Health, Education, and Welfare
• History of Economic Thought
• International Economics
• Labour and Demographic Economics
• Law and Economics
• Macroeconomics and Monetary Economics
• Microeconomics
• Public Economics
• Urban, Rural, and Regional Economics
• Welfare Economics
• Browse content in Education
• Adult Education and Continuous Learning
• Care and Counselling of Students
• Early Childhood and Elementary Education
• Educational Equipment and Technology
• Educational Strategies and Policy
• Higher and Further Education
• Organization and Management of Education
• Philosophy and Theory of Education
• Schools Studies
• Secondary Education
• Teaching of a Specific Subject
• Teaching of Specific Groups and Special Educational Needs
• Teaching Skills and Techniques
• Browse content in Environment
• Applied Ecology (Social Science)
• Climate Change
• Conservation of the Environment (Social Science)
• Environmentalist Thought and Ideology (Social Science)
• Natural Disasters (Environment)
• Social Impact of Environmental Issues (Social Science)
• Browse content in Human Geography
• Cultural Geography
• Economic Geography
• Political Geography
• Browse content in Interdisciplinary Studies
• Communication Studies
• Museums, Libraries, and Information Sciences
• Browse content in Politics
• African Politics
• Asian Politics
• Chinese Politics
• Comparative Politics
• Conflict Politics
• Elections and Electoral Studies
• Environmental Politics
• European Union
• Foreign Policy
• Gender and Politics
• Human Rights and Politics
• Indian Politics
• International Relations
• International Organization (Politics)
• International Political Economy
• Irish Politics
• Latin American Politics
• Middle Eastern Politics
• Political Theory
• Political Behaviour
• Political Economy
• Political Institutions
• Political Methodology
• Political Communication
• Political Philosophy
• Political Sociology
• Politics and Law
• Public Policy
• Public Administration
• Quantitative Political Methodology
• Regional Political Studies
• Russian Politics
• Security Studies
• State and Local Government
• UK Politics
• US Politics
• Browse content in Regional and Area Studies
• African Studies
• Asian Studies
• East Asian Studies
• Japanese Studies
• Latin American Studies
• Middle Eastern Studies
• Native American Studies
• Scottish Studies
• Browse content in Research and Information
• Research Methods
• Browse content in Social Work
• Addictions and Substance Misuse
• Adoption and Fostering
• Care of the Elderly
• Child and Adolescent Social Work
• Couple and Family Social Work
• Developmental and Physical Disabilities Social Work
• Direct Practice and Clinical Social Work
• Emergency Services
• Human Behaviour and the Social Environment
• International and Global Issues in Social Work
• Mental and Behavioural Health
• Social Justice and Human Rights
• Social Policy and Advocacy
• Social Work and Crime and Justice
• Social Work Macro Practice
• Social Work Practice Settings
• Social Work Research and Evidence-based Practice
• Welfare and Benefit Systems
• Browse content in Sociology
• Childhood Studies
• Community Development
• Comparative and Historical Sociology
• Economic Sociology
• Gender and Sexuality
• Gerontology and Ageing
• Health, Illness, and Medicine
• Marriage and the Family
• Migration Studies
• Occupations, Professions, and Work
• Organizations
• Population and Demography
• Race and Ethnicity
• Social Theory
• Social Movements and Social Change
• Social Research and Statistics
• Social Stratification, Inequality, and Mobility
• Sociology of Religion
• Sociology of Education
• Sport and Leisure
• Urban and Rural Studies
• Browse content in Warfare and Defence
• Defence Strategy, Planning, and Research
• Land Forces and Warfare
• Military Administration
• Military Life and Institutions
• Naval Forces and Warfare
• Other Warfare and Defence Issues
• Peace Studies and Conflict Resolution
• Weapons and Equipment

• < Previous chapter
• Next chapter >

48 Problem Solving

Department of Psychological and Brain Sciences, University of California, Santa Barbara

• Published: 03 June 2013
• Cite Icon Cite
• Permissions Icon Permissions

Problem solving refers to cognitive processing directed at achieving a goal when the problem solver does not initially know a solution method. A problem exists when someone has a goal but does not know how to achieve it. Problems can be classified as routine or nonroutine, and as well defined or ill defined. The major cognitive processes in problem solving are representing, planning, executing, and monitoring. The major kinds of knowledge required for problem solving are facts, concepts, procedures, strategies, and beliefs. Classic theoretical approaches to the study of problem solving are associationism, Gestalt, and information processing. Current issues and suggested future issues include decision making, intelligence and creativity, teaching of thinking skills, expert problem solving, analogical reasoning, mathematical and scientific thinking, everyday thinking, and the cognitive neuroscience of problem solving. Common themes concern the domain specificity of problem solving and a focus on problem solving in authentic contexts.

The study of problem solving begins with defining problem solving, problem, and problem types. This introduction to problem solving is rounded out with an examination of cognitive processes in problem solving, the role of knowledge in problem solving, and historical approaches to the study of problem solving.

Definition of Problem Solving

Problem solving refers to cognitive processing directed at achieving a goal for which the problem solver does not initially know a solution method. This definition consists of four major elements (Mayer, 1992 ; Mayer & Wittrock, 2006 ):

Cognitive —Problem solving occurs within the problem solver’s cognitive system and can only be inferred indirectly from the problem solver’s behavior (including biological changes, introspections, and actions during problem solving). Process —Problem solving involves mental computations in which some operation is applied to a mental representation, sometimes resulting in the creation of a new mental representation. Directed —Problem solving is aimed at achieving a goal. Personal —Problem solving depends on the existing knowledge of the problem solver so that what is a problem for one problem solver may not be a problem for someone who already knows a solution method.

The definition is broad enough to include a wide array of cognitive activities such as deciding which apartment to rent, figuring out how to use a cell phone interface, playing a game of chess, making a medical diagnosis, finding the answer to an arithmetic word problem, or writing a chapter for a handbook. Problem solving is pervasive in human life and is crucial for human survival. Although this chapter focuses on problem solving in humans, problem solving also occurs in nonhuman animals and in intelligent machines.

How is problem solving related to other forms of high-level cognition processing, such as thinking and reasoning? Thinking refers to cognitive processing in individuals but includes both directed thinking (which corresponds to the definition of problem solving) and undirected thinking such as daydreaming (which does not correspond to the definition of problem solving). Thus, problem solving is a type of thinking (i.e., directed thinking).

Reasoning refers to problem solving within specific classes of problems, such as deductive reasoning or inductive reasoning. In deductive reasoning, the reasoner is given premises and must derive a conclusion by applying the rules of logic. For example, given that “A is greater than B” and “B is greater than C,” a reasoner can conclude that “A is greater than C.” In inductive reasoning, the reasoner is given (or has experienced) a collection of examples or instances and must infer a rule. For example, given that X, C, and V are in the “yes” group and x, c, and v are in the “no” group, the reasoning may conclude that B is in “yes” group because it is in uppercase format. Thus, reasoning is a type of problem solving.

Definition of Problem

A problem occurs when someone has a goal but does not know to achieve it. This definition is consistent with how the Gestalt psychologist Karl Duncker ( 1945 , p. 1) defined a problem in his classic monograph, On Problem Solving : “A problem arises when a living creature has a goal but does not know how this goal is to be reached.” However, today researchers recognize that the definition should be extended to include problem solving by intelligent machines. This definition can be clarified using an information processing approach by noting that a problem occurs when a situation is in the given state, the problem solver wants the situation to be in the goal state, and there is no obvious way to move from the given state to the goal state (Newell & Simon, 1972 ). Accordingly, the three main elements in describing a problem are the given state (i.e., the current state of the situation), the goal state (i.e., the desired state of the situation), and the set of allowable operators (i.e., the actions the problem solver is allowed to take). The definition of “problem” is broad enough to include the situation confronting a physician who wishes to make a diagnosis on the basis of preliminary tests and a patient examination, as well as a beginning physics student trying to solve a complex physics problem.

Types of Problems

It is customary in the problem-solving literature to make a distinction between routine and nonroutine problems. Routine problems are problems that are so familiar to the problem solver that the problem solver knows a solution method. For example, for most adults, “What is 365 divided by 12?” is a routine problem because they already know the procedure for long division. Nonroutine problems are so unfamiliar to the problem solver that the problem solver does not know a solution method. For example, figuring out the best way to set up a funding campaign for a nonprofit charity is a nonroutine problem for most volunteers. Technically, routine problems do not meet the definition of problem because the problem solver has a goal but knows how to achieve it. Much research on problem solving has focused on routine problems, although most interesting problems in life are nonroutine.

Another customary distinction is between well-defined and ill-defined problems. Well-defined problems have a clearly specified given state, goal state, and legal operators. Examples include arithmetic computation problems or games such as checkers or tic-tac-toe. Ill-defined problems have a poorly specified given state, goal state, or legal operators, or a combination of poorly defined features. Examples include solving the problem of global warming or finding a life partner. Although, ill-defined problems are more challenging, much research in problem solving has focused on well-defined problems.

Cognitive Processes in Problem Solving

The process of problem solving can be broken down into two main phases: problem representation , in which the problem solver builds a mental representation of the problem situation, and problem solution , in which the problem solver works to produce a solution. The major subprocess in problem representation is representing , which involves building a situation model —that is, a mental representation of the situation described in the problem. The major subprocesses in problem solution are planning , which involves devising a plan for how to solve the problem; executing , which involves carrying out the plan; and monitoring , which involves evaluating and adjusting one’s problem solving.

For example, given an arithmetic word problem such as “Alice has three marbles. Sarah has two more marbles than Alice. How many marbles does Sarah have?” the process of representing involves building a situation model in which Alice has a set of marbles, there is set of marbles for the difference between the two girls, and Sarah has a set of marbles that consists of Alice’s marbles and the difference set. In the planning process, the problem solver sets a goal of adding 3 and 2. In the executing process, the problem solver carries out the computation, yielding an answer of 5. In the monitoring process, the problem solver looks over what was done and concludes that 5 is a reasonable answer. In most complex problem-solving episodes, the four cognitive processes may not occur in linear order, but rather may interact with one another. Although some research focuses mainly on the execution process, problem solvers may tend to have more difficulty with the processes of representing, planning, and monitoring.

Knowledge for Problem Solving

An important theme in problem-solving research is that problem-solving proficiency on any task depends on the learner’s knowledge (Anderson et al., 2001 ; Mayer, 1992 ). Five kinds of knowledge are as follows:

Facts —factual knowledge about the characteristics of elements in the world, such as “Sacramento is the capital of California” Concepts —conceptual knowledge, including categories, schemas, or models, such as knowing the difference between plants and animals or knowing how a battery works Procedures —procedural knowledge of step-by-step processes, such as how to carry out long-division computations Strategies —strategic knowledge of general methods such as breaking a problem into parts or thinking of a related problem Beliefs —attitudinal knowledge about how one’s cognitive processing works such as thinking, “I’m good at this”

Although some research focuses mainly on the role of facts and procedures in problem solving, complex problem solving also depends on the problem solver’s concepts, strategies, and beliefs (Mayer, 1992 ).

Historical Approaches to Problem Solving

Psychological research on problem solving began in the early 1900s, as an outgrowth of mental philosophy (Humphrey, 1963 ; Mandler & Mandler, 1964 ). Throughout the 20th century four theoretical approaches developed: early conceptions, associationism, Gestalt psychology, and information processing.

Early Conceptions

The start of psychology as a science can be set at 1879—the year Wilhelm Wundt opened the first world’s psychology laboratory in Leipzig, Germany, and sought to train the world’s first cohort of experimental psychologists. Instead of relying solely on philosophical speculations about how the human mind works, Wundt sought to apply the methods of experimental science to issues addressed in mental philosophy. His theoretical approach became structuralism —the analysis of consciousness into its basic elements.

Wundt’s main contribution to the study of problem solving, however, was to call for its banishment. According to Wundt, complex cognitive processing was too complicated to be studied by experimental methods, so “nothing can be discovered in such experiments” (Wundt, 1911/1973 ). Despite his admonishments, however, a group of his former students began studying thinking mainly in Wurzburg, Germany. Using the method of introspection, subjects were asked to describe their thought process as they solved word association problems, such as finding the superordinate of “newspaper” (e.g., an answer is “publication”). Although the Wurzburg group—as they came to be called—did not produce a new theoretical approach, they found empirical evidence that challenged some of the key assumptions of mental philosophy. For example, Aristotle had proclaimed that all thinking involves mental imagery, but the Wurzburg group was able to find empirical evidence for imageless thought .

Associationism

The first major theoretical approach to take hold in the scientific study of problem solving was associationism —the idea that the cognitive representations in the mind consist of ideas and links between them and that cognitive processing in the mind involves following a chain of associations from one idea to the next (Mandler & Mandler, 1964 ; Mayer, 1992 ). For example, in a classic study, E. L. Thorndike ( 1911 ) placed a hungry cat in what he called a puzzle box—a wooden crate in which pulling a loop of string that hung from overhead would open a trap door to allow the cat to escape to a bowl of food outside the crate. Thorndike placed the cat in the puzzle box once a day for several weeks. On the first day, the cat engaged in many extraneous behaviors such as pouncing against the wall, pushing its paws through the slats, and meowing, but on successive days the number of extraneous behaviors tended to decrease. Overall, the time required to get out of the puzzle box decreased over the course of the experiment, indicating the cat was learning how to escape.

Thorndike’s explanation for how the cat learned to solve the puzzle box problem is based on an associationist view: The cat begins with a habit family hierarchy —a set of potential responses (e.g., pouncing, thrusting, meowing, etc.) all associated with the same stimulus (i.e., being hungry and confined) and ordered in terms of strength of association. When placed in the puzzle box, the cat executes its strongest response (e.g., perhaps pouncing against the wall), but when it fails, the strength of the association is weakened, and so on for each unsuccessful action. Eventually, the cat gets down to what was initially a weak response—waving its paw in the air—but when that response leads to accidentally pulling the string and getting out, it is strengthened. Over the course of many trials, the ineffective responses become weak and the successful response becomes strong. Thorndike refers to this process as the law of effect : Responses that lead to dissatisfaction become less associated with the situation and responses that lead to satisfaction become more associated with the situation. According to Thorndike’s associationist view, solving a problem is simply a matter of trial and error and accidental success. A major challenge to assocationist theory concerns the nature of transfer—that is, where does a problem solver find a creative solution that has never been performed before? Associationist conceptions of cognition can be seen in current research, including neural networks, connectionist models, and parallel distributed processing models (Rogers & McClelland, 2004 ).

Gestalt Psychology

The Gestalt approach to problem solving developed in the 1930s and 1940s as a counterbalance to the associationist approach. According to the Gestalt approach, cognitive representations consist of coherent structures (rather than individual associations) and the cognitive process of problem solving involves building a coherent structure (rather than strengthening and weakening of associations). For example, in a classic study, Kohler ( 1925 ) placed a hungry ape in a play yard that contained several empty shipping crates and a banana attached overhead but out of reach. Based on observing the ape in this situation, Kohler noted that the ape did not randomly try responses until one worked—as suggested by Thorndike’s associationist view. Instead, the ape stood under the banana, looked up at it, looked at the crates, and then in a flash of insight stacked the crates under the bananas as a ladder, and walked up the steps in order to reach the banana.

According to Kohler, the ape experienced a sudden visual reorganization in which the elements in the situation fit together in a way to solve the problem; that is, the crates could become a ladder that reduces the distance to the banana. Kohler referred to the underlying mechanism as insight —literally seeing into the structure of the situation. A major challenge of Gestalt theory is its lack of precision; for example, naming a process (i.e., insight) is not the same as explaining how it works. Gestalt conceptions can be seen in modern research on mental models and schemas (Gentner & Stevens, 1983 ).

Information Processing

The information processing approach to problem solving developed in the 1960s and 1970s and was based on the influence of the computer metaphor—the idea that humans are processors of information (Mayer, 2009 ). According to the information processing approach, problem solving involves a series of mental computations—each of which consists of applying a process to a mental representation (such as comparing two elements to determine whether they differ).

In their classic book, Human Problem Solving , Newell and Simon ( 1972 ) proposed that problem solving involved a problem space and search heuristics . A problem space is a mental representation of the initial state of the problem, the goal state of the problem, and all possible intervening states (based on applying allowable operators). Search heuristics are strategies for moving through the problem space from the given to the goal state. Newell and Simon focused on means-ends analysis , in which the problem solver continually sets goals and finds moves to accomplish goals.

Newell and Simon used computer simulation as a research method to test their conception of human problem solving. First, they asked human problem solvers to think aloud as they solved various problems such as logic problems, chess, and cryptarithmetic problems. Then, based on an information processing analysis, Newell and Simon created computer programs that solved these problems. In comparing the solution behavior of humans and computers, they found high similarity, suggesting that the computer programs were solving problems using the same thought processes as humans.

An important advantage of the information processing approach is that problem solving can be described with great clarity—as a computer program. An important limitation of the information processing approach is that it is most useful for describing problem solving for well-defined problems rather than ill-defined problems. The information processing conception of cognition lives on as a keystone of today’s cognitive science (Mayer, 2009 ).

Classic Issues in Problem Solving

Three classic issues in research on problem solving concern the nature of transfer (suggested by the associationist approach), the nature of insight (suggested by the Gestalt approach), and the role of problem-solving heuristics (suggested by the information processing approach).

Transfer refers to the effects of prior learning on new learning (or new problem solving). Positive transfer occurs when learning A helps someone learn B. Negative transfer occurs when learning A hinders someone from learning B. Neutral transfer occurs when learning A has no effect on learning B. Positive transfer is a central goal of education, but research shows that people often do not transfer what they learned to solving problems in new contexts (Mayer, 1992 ; Singley & Anderson, 1989 ).

Three conceptions of the mechanisms underlying transfer are specific transfer , general transfer , and specific transfer of general principles . Specific transfer refers to the idea that learning A will help someone learn B only if A and B have specific elements in common. For example, learning Spanish may help someone learn Latin because some of the vocabulary words are similar and the verb conjugation rules are similar. General transfer refers to the idea that learning A can help someone learn B even they have nothing specifically in common but A helps improve the learner’s mind in general. For example, learning Latin may help people learn “proper habits of mind” so they are better able to learn completely unrelated subjects as well. Specific transfer of general principles is the idea that learning A will help someone learn B if the same general principle or solution method is required for both even if the specific elements are different.

In a classic study, Thorndike and Woodworth ( 1901 ) found that students who learned Latin did not subsequently learn bookkeeping any better than students who had not learned Latin. They interpreted this finding as evidence for specific transfer—learning A did not transfer to learning B because A and B did not have specific elements in common. Modern research on problem-solving transfer continues to show that people often do not demonstrate general transfer (Mayer, 1992 ). However, it is possible to teach people a general strategy for solving a problem, so that when they see a new problem in a different context they are able to apply the strategy to the new problem (Judd, 1908 ; Mayer, 2008 )—so there is also research support for the idea of specific transfer of general principles.

Insight refers to a change in a problem solver’s mind from not knowing how to solve a problem to knowing how to solve it (Mayer, 1995 ; Metcalfe & Wiebe, 1987 ). In short, where does the idea for a creative solution come from? A central goal of problem-solving research is to determine the mechanisms underlying insight.

The search for insight has led to five major (but not mutually exclusive) explanatory mechanisms—insight as completing a schema, insight as suddenly reorganizing visual information, insight as reformulation of a problem, insight as removing mental blocks, and insight as finding a problem analog (Mayer, 1995 ). Completing a schema is exemplified in a study by Selz (Fridja & de Groot, 1982 ), in which people were asked to think aloud as they solved word association problems such as “What is the superordinate for newspaper?” To solve the problem, people sometimes thought of a coordinate, such as “magazine,” and then searched for a superordinate category that subsumed both terms, such as “publication.” According to Selz, finding a solution involved building a schema that consisted of a superordinate and two subordinate categories.

Reorganizing visual information is reflected in Kohler’s ( 1925 ) study described in a previous section in which a hungry ape figured out how to stack boxes as a ladder to reach a banana hanging above. According to Kohler, the ape looked around the yard and found the solution in a flash of insight by mentally seeing how the parts could be rearranged to accomplish the goal.

Reformulating a problem is reflected in a classic study by Duncker ( 1945 ) in which people are asked to think aloud as they solve the tumor problem—how can you destroy a tumor in a patient without destroying surrounding healthy tissue by using rays that at sufficient intensity will destroy any tissue in their path? In analyzing the thinking-aloud protocols—that is, transcripts of what the problem solvers said—Duncker concluded that people reformulated the goal in various ways (e.g., avoid contact with healthy tissue, immunize healthy tissue, have ray be weak in healthy tissue) until they hit upon a productive formulation that led to the solution (i.e., concentrating many weak rays on the tumor).

Removing mental blocks is reflected in classic studies by Duncker ( 1945 ) in which solving a problem involved thinking of a novel use for an object, and by Luchins ( 1942 ) in which solving a problem involved not using a procedure that had worked well on previous problems. Finding a problem analog is reflected in classic research by Wertheimer ( 1959 ) in which learning to find the area of a parallelogram is supported by the insight that one could cut off the triangle on one side and place it on the other side to form a rectangle—so a parallelogram is really a rectangle in disguise. The search for insight along each of these five lines continues in current problem-solving research.

Heuristics are problem-solving strategies, that is, general approaches to how to solve problems. Newell and Simon ( 1972 ) suggested three general problem-solving heuristics for moving from a given state to a goal state: random trial and error , hill climbing , and means-ends analysis . Random trial and error involves randomly selecting a legal move and applying it to create a new problem state, and repeating that process until the goal state is reached. Random trial and error may work for simple problems but is not efficient for complex ones. Hill climbing involves selecting the legal move that moves the problem solver closer to the goal state. Hill climbing will not work for problems in which the problem solver must take a move that temporarily moves away from the goal as is required in many problems.

Means-ends analysis involves creating goals and seeking moves that can accomplish the goal. If a goal cannot be directly accomplished, a subgoal is created to remove one or more obstacles. Newell and Simon ( 1972 ) successfully used means-ends analysis as the search heuristic in a computer program aimed at general problem solving, that is, solving a diverse collection of problems. However, people may also use specific heuristics that are designed to work for specific problem-solving situations (Gigerenzer, Todd, & ABC Research Group, 1999 ; Kahneman & Tversky, 1984 ).

Current and Future Issues in Problem Solving

Eight current issues in problem solving involve decision making, intelligence and creativity, teaching of thinking skills, expert problem solving, analogical reasoning, mathematical and scientific problem solving, everyday thinking, and the cognitive neuroscience of problem solving.

Decision Making

Decision making refers to the cognitive processing involved in choosing between two or more alternatives (Baron, 2000 ; Markman & Medin, 2002 ). For example, a decision-making task may involve choosing between getting $240 for sure or having a 25% change of getting $1000. According to economic theories such as expected value theory, people should chose the second option, which is worth $250 (i.e., .25 x $1000) rather than the first option, which is worth $240 (1.00 x $240), but psychological research shows that most people prefer the first option (Kahneman & Tversky, 1984 ).

Research on decision making has generated three classes of theories (Markman & Medin, 2002 ): descriptive theories, such as prospect theory (Kahneman & Tversky), which are based on the ideas that people prefer to overweight the cost of a loss and tend to overestimate small probabilities; heuristic theories, which are based on the idea that people use a collection of short-cut strategies such as the availability heuristic (Gigerenzer et al., 1999 ; Kahneman & Tversky, 2000 ); and constructive theories, such as mental accounting (Kahneman & Tversky, 2000 ), in which people build a narrative to justify their choices to themselves. Future research is needed to examine decision making in more realistic settings.

Intelligence and Creativity

Although researchers do not have complete consensus on the definition of intelligence (Sternberg, 1990 ), it is reasonable to view intelligence as the ability to learn or adapt to new situations. Fluid intelligence refers to the potential to solve problems without any relevant knowledge, whereas crystallized intelligence refers to the potential to solve problems based on relevant prior knowledge (Sternberg & Gregorenko, 2003 ). As people gain more experience in a field, their problem-solving performance depends more on crystallized intelligence (i.e., domain knowledge) than on fluid intelligence (i.e., general ability) (Sternberg & Gregorenko, 2003 ). The ability to monitor and manage one’s cognitive processing during problem solving—which can be called metacognition —is an important aspect of intelligence (Sternberg, 1990 ). Research is needed to pinpoint the knowledge that is needed to support intelligent performance on problem-solving tasks.

Creativity refers to the ability to generate ideas that are original (i.e., other people do not think of the same idea) and functional (i.e., the idea works; Sternberg, 1999 ). Creativity is often measured using tests of divergent thinking —that is, generating as many solutions as possible for a problem (Guilford, 1967 ). For example, the uses test asks people to list as many uses as they can think of for a brick. Creativity is different from intelligence, and it is at the heart of creative problem solving—generating a novel solution to a problem that the problem solver has never seen before. An important research question concerns whether creative problem solving depends on specific knowledge or creativity ability in general.

Teaching of Thinking Skills

How can people learn to be better problem solvers? Mayer ( 2008 ) proposes four questions concerning teaching of thinking skills:

What to teach —Successful programs attempt to teach small component skills (such as how to generate and evaluate hypotheses) rather than improve the mind as a single monolithic skill (Covington, Crutchfield, Davies, & Olton, 1974 ). How to teach —Successful programs focus on modeling the process of problem solving rather than solely reinforcing the product of problem solving (Bloom & Broder, 1950 ). Where to teach —Successful programs teach problem-solving skills within the specific context they will be used rather than within a general course on how to solve problems (Nickerson, 1999 ). When to teach —Successful programs teaching higher order skills early rather than waiting until lower order skills are completely mastered (Tharp & Gallimore, 1988 ).

Overall, research on teaching of thinking skills points to the domain specificity of problem solving; that is, successful problem solving depends on the problem solver having domain knowledge that is relevant to the problem-solving task.

Expert Problem Solving

Research on expertise is concerned with differences between how experts and novices solve problems (Ericsson, Feltovich, & Hoffman, 2006 ). Expertise can be defined in terms of time (e.g., 10 years of concentrated experience in a field), performance (e.g., earning a perfect score on an assessment), or recognition (e.g., receiving a Nobel Prize or becoming Grand Master in chess). For example, in classic research conducted in the 1940s, de Groot ( 1965 ) found that chess experts did not have better general memory than chess novices, but they did have better domain-specific memory for the arrangement of chess pieces on the board. Chase and Simon ( 1973 ) replicated this result in a better controlled experiment. An explanation is that experts have developed schemas that allow them to chunk collections of pieces into a single configuration.

In another landmark study, Larkin et al. ( 1980 ) compared how experts (e.g., physics professors) and novices (e.g., first-year physics students) solved textbook physics problems about motion. Experts tended to work forward from the given information to the goal, whereas novices tended to work backward from the goal to the givens using a means-ends analysis strategy. Experts tended to store their knowledge in an integrated way, whereas novices tended to store their knowledge in isolated fragments. In another study, Chi, Feltovich, and Glaser ( 1981 ) found that experts tended to focus on the underlying physics concepts (such as conservation of energy), whereas novices tended to focus on the surface features of the problem (such as inclined planes or springs). Overall, research on expertise is useful in pinpointing what experts know that is different from what novices know. An important theme is that experts rely on domain-specific knowledge rather than solely general cognitive ability.

Analogical Reasoning

Analogical reasoning occurs when people solve one problem by using their knowledge about another problem (Holyoak, 2005 ). For example, suppose a problem solver learns how to solve a problem in one context using one solution method and then is given a problem in another context that requires the same solution method. In this case, the problem solver must recognize that the new problem has structural similarity to the old problem (i.e., it may be solved by the same method), even though they do not have surface similarity (i.e., the cover stories are different). Three steps in analogical reasoning are recognizing —seeing that a new problem is similar to a previously solved problem; abstracting —finding the general method used to solve the old problem; and mapping —using that general method to solve the new problem.

Research on analogical reasoning shows that people often do not recognize that a new problem can be solved by the same method as a previously solved problem (Holyoak, 2005 ). However, research also shows that successful analogical transfer to a new problem is more likely when the problem solver has experience with two old problems that have the same underlying structural features (i.e., they are solved by the same principle) but different surface features (i.e., they have different cover stories) (Holyoak, 2005 ). This finding is consistent with the idea of specific transfer of general principles as described in the section on “Transfer.”

Mathematical and Scientific Problem Solving

Research on mathematical problem solving suggests that five kinds of knowledge are needed to solve arithmetic word problems (Mayer, 2008 ):

Factual knowledge —knowledge about the characteristics of problem elements, such as knowing that there are 100 cents in a dollar Schematic knowledge —knowledge of problem types, such as being able to recognize time-rate-distance problems Strategic knowledge —knowledge of general methods, such as how to break a problem into parts Procedural knowledge —knowledge of processes, such as how to carry our arithmetic operations Attitudinal knowledge —beliefs about one’s mathematical problem-solving ability, such as thinking, “I am good at this”

People generally possess adequate procedural knowledge but may have difficulty in solving mathematics problems because they lack factual, schematic, strategic, or attitudinal knowledge (Mayer, 2008 ). Research is needed to pinpoint the role of domain knowledge in mathematical problem solving.

Research on scientific problem solving shows that people harbor misconceptions, such as believing that a force is needed to keep an object in motion (McCloskey, 1983 ). Learning to solve science problems involves conceptual change, in which the problem solver comes to recognize that previous conceptions are wrong (Mayer, 2008 ). Students can be taught to engage in scientific reasoning such as hypothesis testing through direct instruction in how to control for variables (Chen & Klahr, 1999 ). A central theme of research on scientific problem solving concerns the role of domain knowledge.

Everyday Thinking

Everyday thinking refers to problem solving in the context of one’s life outside of school. For example, children who are street vendors tend to use different procedures for solving arithmetic problems when they are working on the streets than when they are in school (Nunes, Schlieman, & Carraher, 1993 ). This line of research highlights the role of situated cognition —the idea that thinking always is shaped by the physical and social context in which it occurs (Robbins & Aydede, 2009 ). Research is needed to determine how people solve problems in authentic contexts.

Cognitive Neuroscience of Problem Solving

The cognitive neuroscience of problem solving is concerned with the brain activity that occurs during problem solving. For example, using fMRI brain imaging methodology, Goel ( 2005 ) found that people used the language areas of the brain to solve logical reasoning problems presented in sentences (e.g., “All dogs are pets…”) and used the spatial areas of the brain to solve logical reasoning problems presented in abstract letters (e.g., “All D are P…”). Cognitive neuroscience holds the potential to make unique contributions to the study of problem solving.

Problem solving has always been a topic at the fringe of cognitive psychology—too complicated to study intensively but too important to completely ignore. Problem solving—especially in realistic environments—is messy in comparison to studying elementary processes in cognition. The field remains fragmented in the sense that topics such as decision making, reasoning, intelligence, expertise, mathematical problem solving, everyday thinking, and the like are considered to be separate topics, each with its own separate literature. Yet some recurring themes are the role of domain-specific knowledge in problem solving and the advantages of studying problem solving in authentic contexts.

Future Directions

Some important issues for future research include the three classic issues examined in this chapter—the nature of problem-solving transfer (i.e., How are people able to use what they know about previous problem solving to help them in new problem solving?), the nature of insight (e.g., What is the mechanism by which a creative solution is constructed?), and heuristics (e.g., What are some teachable strategies for problem solving?). In addition, future research in problem solving should continue to pinpoint the role of domain-specific knowledge in problem solving, the nature of cognitive ability in problem solving, how to help people develop proficiency in solving problems, and how to provide aids for problem solving.

Anderson L. W. , Krathwohl D. R. , Airasian P. W. , Cruikshank K. A. , Mayer R. E. , Pintrich P. R. , Raths, J., & Wittrock M. C. ( 2001 ). A taxonomy for learning, teaching, and assessing: A revision of Bloom’s taxonomy of educational objectives. New York : Longman.

Baron J. ( 2000 ). Thinking and deciding (3rd ed.). New York : Cambridge University Press.

Google Scholar

Google Preview

Bloom B. S. , & Broder B. J. ( 1950 ). Problem-solving processes of college students: An exploratory investigation. Chicago : University of Chicago Press.

Chase W. G. , & Simon H. A. ( 1973 ). Perception in chess.   Cognitive Psychology, 4, 55–81.

Chen Z. , & Klahr D. ( 1999 ). All other things being equal: Acquisition and transfer of the control of variable strategy . Child Development, 70, 1098–1120.

Chi M. T. H. , Feltovich P. J. , & Glaser R. ( 1981 ). Categorization and representation of physics problems by experts and novices. Cognitive Science, 5, 121–152.

Covington M. V. , Crutchfield R. S. , Davies L. B. , & Olton R. M. ( 1974 ). The productive thinking program. Columbus, OH : Merrill.

de Groot A. D. ( 1965 ). Thought and choice in chess. The Hague, The Netherlands : Mouton.

Duncker K. ( 1945 ). On problem solving.   Psychological Monographs, 58 (3) (Whole No. 270).

Ericsson K. A. , Feltovich P. J. , & Hoffman R. R. (Eds.). ( 2006 ). The Cambridge handbook of expertise and expert performance. New York : Cambridge University Press.

Fridja N. H. , & de Groot A. D. ( 1982 ). Otto Selz: His contribution to psychology. The Hague, The Netherlands : Mouton.

Gentner D. , & Stevens A. L. (Eds.). ( 1983 ). Mental models. Hillsdale, NJ : Erlbaum.

Gigerenzer G. , Todd P. M. , & ABC Research Group (Eds.). ( 1999 ). Simple heuristics that make us smart. Oxford, England : Oxford University Press.

Goel V. ( 2005 ). Cognitive neuroscience of deductive reasoning. In K. J. Holyoak & R. G. Morrison (Eds.), The Cambridge handbook of thinking and reasoning (pp. 475–492). New York : Cambridge University Press.

Guilford J. P. ( 1967 ). The nature of human intelligence. New York : McGraw-Hill.

Holyoak K. J. ( 2005 ). Analogy. In K. J. Holyoak & R. G. Morrison (Eds.), The Cambridge handbook of thinking and reasoning (pp. 117–142). New York : Cambridge University Press.

Humphrey G. ( 1963 ). Thinking: An introduction to experimental psychology. New York : Wiley.

Judd C. H. ( 1908 ). The relation of special training and general intelligence. Educational Review, 36, 28–42.

Kahneman D. , & Tversky A. ( 1984 ). Choices, values, and frames. American Psychologist, 39, 341–350.

Kahneman D. , & Tversky A. (Eds.). ( 2000 ). Choices, values, and frames. New York : Cambridge University Press.

Kohler W. ( 1925 ). The mentality of apes. New York : Liveright.

Larkin J. H. , McDermott J. , Simon D. P. , & Simon H. A. ( 1980 ). Expert and novice performance in solving physics problems. Science, 208, 1335–1342.

Luchins A. ( 1942 ). Mechanization in problem solving.   Psychological Monographs, 54 (6) (Whole No. 248).

Mandler J. M. , & Mandler G. ( 1964 ). Thinking from associationism to Gestalt. New York : Wiley.

Markman A. B. , & Medin D. L. ( 2002 ). Decision making. In D. Medin (Ed.), Stevens’ handbook of experimental psychology, Vol. 2. Memory and cognitive processes (2nd ed., pp. 413–466). New York : Wiley.

Mayer R. E. ( 1992 ). Thinking, problem solving, cognition (2nd ed). New York : Freeman.

Mayer R. E. ( 1995 ). The search for insight: Grappling with Gestalt psychology’s unanswered questions. In R. J. Sternberg & J. E. Davidson (Eds.), The nature of insight (pp. 3–32). Cambridge, MA : MIT Press.

Mayer R. E. ( 2008 ). Learning and instruction. Upper Saddle River, NJ : Merrill Prentice Hall.

Mayer R. E. ( 2009 ). Information processing. In T. L. Good (Ed.), 21st century education: A reference handbook (pp. 168–174). Thousand Oaks, CA : Sage.

Mayer R. E. , & Wittrock M. C. ( 2006 ). Problem solving. In P. A. Alexander & P. H. Winne (Eds.), Handbook of educational psychology (2nd ed., pp. 287–304). Mahwah, NJ : Erlbaum.

McCloskey M. ( 1983 ). Intuitive physics.   Scientific American, 248 (4), 122–130.

Metcalfe J. , & Wiebe D. ( 1987 ). Intuition in insight and non-insight problem solving. Memory and Cognition, 15, 238–246.

Newell A. , & Simon H. A. ( 1972 ). Human problem solving. Englewood Cliffs, NJ : Prentice-Hall.

Nickerson R. S. ( 1999 ). Enhancing creativity. In R. J. Sternberg (Ed.), Handbook of creativity (pp. 392–430). New York : Cambridge University Press.

Nunes T. , Schliemann A. D. , & Carraher D. W , ( 1993 ). Street mathematics and school mathematics. Cambridge, England : Cambridge University Press.

Robbins P. , & Aydede M. (Eds.). ( 2009 ). The Cambridge handbook of situated cognition. New York : Cambridge University Press.

Rogers T. T. , & McClelland J. L. ( 2004 ). Semantic cognition: A parallel distributed processing approach. Cambridge, MA : MIT Press.

Singley M. K. , & Anderson J. R. ( 1989 ). The transfer of cognitive skill. Cambridge, MA : Harvard University Press.

Sternberg R. J. ( 1990 ). Metaphors of mind: Conceptions of the nature of intelligence. New York : Cambridge University Press.

Sternberg R. J. ( 1999 ). Handbook of creativity. New York : Cambridge University Press.

Sternberg R. J. , & Gregorenko E. L. (Eds.). ( 2003 ). The psychology of abilities, competencies, and expertise. New York : Cambridge University Press.

Tharp R. G. , & Gallimore R. ( 1988 ). Rousing minds to life: Teaching, learning, and schooling in social context. New York : Cambridge University Press.

Thorndike E. L. ( 1911 ). Animal intelligence. New York: Hafner.

Thorndike E. L. , & Woodworth R. S. ( 1901 ). The influence of improvement in one mental function upon the efficiency of other functions. Psychological Review, 8, 247–261.

Wertheimer M. ( 1959 ). Productive thinking. New York : Harper and Collins.

Wundt W. ( 1973 ). An introduction to experimental psychology. New York : Arno Press. (Original work published in 1911).

Further Reading

Baron, J. ( 2008 ). Thinking and deciding (4th ed). New York: Cambridge University Press.

Duncker, K. ( 1945 ). On problem solving. Psychological Monographs , 58(3) (Whole No. 270).

Holyoak, K. J. , & Morrison, R. G. ( 2005 ). The Cambridge handbook of thinking and reasoning . New York: Cambridge University Press.

Mayer, R. E. , & Wittrock, M. C. ( 2006 ). Problem solving. In P. A. Alexander & P. H. Winne (Eds.), Handbook of educational psychology (2nd ed., pp. 287–304). Mahwah, NJ: Erlbaum.

Sternberg, R. J. , & Ben-Zeev, T. ( 2001 ). Complex cognition: The psychology of human thought . New York: Oxford University Press.

Weisberg, R. W. ( 2006 ). Creativity . New York: Wiley.

• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

Intelligent problem-solving as integrated hierarchical reinforcement learning

• Manfred Eppe   ORCID: orcid.org/0000-0002-5473-3221 1   nAff4 ,
• Christian Gumbsch   ORCID: orcid.org/0000-0003-2741-6551 2 , 3 ,
• Matthias Kerzel 1 ,
• Phuong D. H. Nguyen 1 ,
• Martin V. Butz   ORCID: orcid.org/0000-0002-8120-8537 2 &
• Stefan Wermter 1  

Nature Machine Intelligence volume  4 ,  pages 11–20 ( 2022 ) Cite this article

5204 Accesses

30 Citations

8 Altmetric

Metrics details

• Cognitive control
• Computational models
• Computer science
• Learning algorithms
• Problem solving

According to cognitive psychology and related disciplines, the development of complex problem-solving behaviour in biological agents depends on hierarchical cognitive mechanisms. Hierarchical reinforcement learning is a promising computational approach that may eventually yield comparable problem-solving behaviour in artificial agents and robots. However, so far, the problem-solving abilities of many human and non-human animals are clearly superior to those of artificial systems. Here we propose steps to integrate biologically inspired hierarchical mechanisms to enable advanced problem-solving skills in artificial agents. We first review the literature in cognitive psychology to highlight the importance of compositional abstraction and predictive processing. Then we relate the gained insights with contemporary hierarchical reinforcement learning methods. Interestingly, our results suggest that all identified cognitive mechanisms have been implemented individually in isolated computational architectures, raising the question of why there exists no single unifying architecture that integrates them. As our final contribution, we address this question by providing an integrative perspective on the computational challenges to develop such a unifying architecture. We expect our results to guide the development of more sophisticated cognitively inspired hierarchical machine learning architectures.

This is a preview of subscription content, access via your institution

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 12 digital issues and online access to articles

111,21 € per year

only 9,27 € per issue

Buy this article

• Purchase on Springer Link
• Instant access to full article PDF

Prices may be subject to local taxes which are calculated during checkout

Similar content being viewed by others

Phy-Q as a measure for physical reasoning intelligence

Cheng Xue, Vimukthini Pinto, … Jochen Renz

Hierarchical motor control in mammals and machines

Josh Merel, Matthew Botvinick & Greg Wayne

Hierarchical generative modelling for autonomous robots

Kai Yuan, Noor Sajid, … Zhibin Li

Gruber, R. et al. New Caledonian crows use mental representations to solve metatool problems. Curr. Biol. 29 , 686–692 (2019).

Article   Google Scholar  

Butz, M. V. & Kutter, E. F. How the Mind Comes into Being (Oxford Univ. Press, 2017).

Perkins, D. N. & Salomon, G. in International Encyclopedia of Education (eds. Husen T. & Postelwhite T. N.) 6452–6457 (Pergamon Press, 1992).

Botvinick, M. M., Niv, Y. & Barto, A. C. Hierarchically organized behavior and its neural foundations: a reinforcement learning perspective. Cognition 113 , 262–280 (2009).

Tomov, M. S., Yagati, S., Kumar, A., Yang, W. & Gershman, S. J. Discovery of hierarchical representations for efficient planning. PLoS Comput. Biol. 16 , e1007594 (2020).

Arulkumaran, K., Deisenroth, M. P., Brundage, M. & Bharath, A. A. Deep reinforcement learning: a brief survey. IEEE Signal Process. Mag. 34 , 26–38 (2017).

Li, Y. Deep reinforcement learning: an overview. Preprint at https://arxiv.org/abs/1701.07274 (2018).

Sutton, R. S. & Barto, A. G. Reinforcement Learning : An Introduction 2nd edn (MIT Press, 2018).

Neftci, E. O. & Averbeck, B. B. Reinforcement learning in artificial and biological systems. Nat. Mach. Intell. 1 , 133–143 (2019).

Eppe, M., Nguyen, P. D. H. & Wermter, S. From semantics to execution: integrating action planning with reinforcement learning for robotic causal problem-solving. Front. Robot. AI 6 , 123 (2019).

Oh, J., Singh, S., Lee, H. & Kohli, P. Zero-shot task generalization with multi-task deep reinforcement learning. In Proc. 34th International Conference on Machine Learning ( ICML ) (eds. Precup, D. & Teh, Y. W.) 2661–2670 (PMLR, 2017).

Sohn, S., Oh, J. & Lee, H. Hierarchical reinforcement learning for zero-shot generalization with subtask dependencies. In Proc. 32nd International Conference on Neural Information Processing Systems ( NeurIPS ) (eds Bengio S. et al.) Vol. 31, 7156–7166 (ACM, 2018).

Hegarty, M. Mechanical reasoning by mental simulation. Trends Cogn. Sci. 8 , 280–285 (2004).

Klauer, K. J. Teaching for analogical transfer as a means of improving problem-solving, thinking and learning. Instruct. Sci. 18 , 179–192 (1989).

Duncker, K. & Lees, L. S. On problem-solving. Psychol. Monographs 58, No.5 (whole No. 270), 85–101 https://doi.org/10.1037/h0093599 (1945).

Dayan, P. Goal-directed control and its antipodes. Neural Netw. 22 , 213–219 (2009).

Dolan, R. J. & Dayan, P. Goals and habits in the brain. Neuron 80 , 312–325 (2013).

O’Doherty, J. P., Cockburn, J. & Pauli, W. M. Learning, reward, and decision making. Annu. Rev. Psychol. 68 , 73–100 (2017).

Tolman, E. C. & Honzik, C. H. Introduction and removal of reward, and maze performance in rats. Univ. California Publ. Psychol. 4 , 257–275 (1930).

Google Scholar  

Butz, M. V. & Hoffmann, J. Anticipations control behavior: animal behavior in an anticipatory learning classifier system. Adaptive Behav. 10 , 75–96 (2002).

Miller, G. A., Galanter, E. & Pribram, K. H. Plans and the Structure of Behavior (Holt, Rinehart & Winston, 1960).

Botvinick, M. & Weinstein, A. Model-based hierarchical reinforcement learning and human action control. Philos. Trans. R. Soc. B Biol. Sci. 369 , 20130480 (2014).

Wiener, J. M. & Mallot, H. A. ’Fine-to-coarse’ route planning and navigation in regionalized environments. Spatial Cogn. Comput. 3 , 331–358 (2003).

Stock, A. & Stock, C. A short history of ideo-motor action. Psychol. Res. 68 , 176–188 (2004).

Hommel, B., Müsseler, J., Aschersleben, G. & Prinz, W. The theory of event coding (TEC): a framework for perception and action planning. Behav. Brain Sci. 24 , 849–878 (2001).

Hoffmann, J. in Anticipatory Behavior in Adaptive Learning Systems : Foundations , Theories and Systems (eds Butz, M. V. et al.) 44–65 (Springer, 2003).

Kunde, W., Elsner, K. & Kiesel, A. No anticipation-no action: the role of anticipation in action and perception. Cogn. Process. 8 , 71–78 (2007).

Barsalou, L. W. Grounded cognition. Annu. Rev. Psychol. 59 , 617–645 (2008).

Butz, M. V. Toward a unified sub-symbolic computational theory of cognition. Front. Psychol. 7 , 925 (2016).

Pulvermüller, F. Brain embodiment of syntax and grammar: discrete combinatorial mechanisms spelt out in neuronal circuits. Brain Lang. 112 , 167–179 (2010).

Sutton, R. S., Precup, D. & Singh, S. Between MDPs and semi-MDPs: a framework for temporal abstraction in reinforcement learning. Artif. Intell. 112 , 181–211 (1999).

Article   MathSciNet   MATH   Google Scholar  

Flash, T. & Hochner, B. Motor primitives in vertebrates and invertebrates. Curr. Opin. Neurobiol. 15 , 660–666 (2005).

Schaal, S. in Adaptive Motion of Animals and Machines (eds. Kimura, H. et al.) 261–280 (Springer, 2006).

Feldman, J., Dodge, E. & Bryant, J. in The Oxford Handbook of Linguistic Analysis (eds Heine, B. & Narrog, H.) 111–138 (Oxford Univ. Press, 2009).

Fodor, J. A. Language, thought and compositionality. Mind Lang. 16 , 1–15 (2001).

Frankland, S. M. & Greene, J. D. Concepts and compositionality: in search of the brain’s language of thought. Annu. Rev. Psychol. 71 , 273–303 (2020).

Hummel, J. E. Getting symbols out of a neural architecture. Connection Sci. 23 , 109–118 (2011).

Haynes, J. D., Wisniewski, D., Gorgen, K., Momennejad, I. & Reverberi, C. FMRI decoding of intentions: compositionality, hierarchy and prospective memory. In Proc. 3rd International Winter Conference on Brain-Computer Interface ( BCI ), 1-3 (IEEE, 2015).

Gärdenfors, P. The Geometry of Meaning : Semantics Based on Conceptual Spaces (MIT Press, 2014).

Book   MATH   Google Scholar  

Lakoff, G. & Johnson, M. Philosophy in the Flesh (Basic Books, 1999).

Eppe, M. et al. A computational framework for concept blending. Artif. Intell. 256 , 105–129 (2018).

Turner, M. The Origin of Ideas (Oxford Univ. Press, 2014).

Deci, E. L. & Ryan, R. M. Self-determination theory and the facilitation of intrinsic motivation. Am. Psychol. 55 , 68–78 (2000).

Friston, K. et al. Active inference and epistemic value. Cogn. Neurosci. 6 , 187–214 (2015).

Berlyne, D. E. Curiosity and exploration. Science 153 , 25–33 (1966).

Loewenstein, G. The psychology of curiosity: a review and reinterpretation. Psychol. Bull. 116 , 75–98 (1994).

Oudeyer, P.-Y., Kaplan, F. & Hafner, V. V. Intrinsic motivation systems for autonomous mental development. In IEEE Transactions on Evolutionary Computation (eds. Coello, C. A. C. et al.) Vol. 11, 265–286 (IEEE, 2007).

Pisula, W. Play and exploration in animals—a comparative analysis. Polish Psychol. Bull. 39 , 104–107 (2008).

Jeannerod, M. Mental imagery in the motor context. Neuropsychologia 33 , 1419–1432 (1995).

Kahnemann, D. & Tversky, A. in Judgement under Uncertainty : Heuristics and Biases (eds Kahneman, D. et al.) Ch. 14, 201–208 (Cambridge Univ. Press, 1982).

Wells, G. L. & Gavanski, I. Mental simulation of causality. J. Personal. Social Psychol. 56 , 161–169 (1989).

Taylor, S. E., Pham, L. B., Rivkin, I. D. & Armor, D. A. Harnessing the imagination: mental simulation, self-regulation and coping. Am. Psychol. 53 , 429–439 (1998).

Kaplan, F. & Oudeyer, P.-Y. in Embodied Artificial Intelligence , Lecture Notes in Computer Science Vol. 3139 (eds Iida, F. et al.) 259–270 (Springer, 2004).

Schmidhuber, J. Formal theory of creativity, fun, and intrinsic motivation. IEEE Trans. Auton. Mental Dev. 2 , 230–247 (2010).

Friston, K., Mattout, J. & Kilner, J. Action understanding and active inference. Biol. Cybern. 104 , 137–160 (2011).

Oudeyer, P.-Y. Computational theories of curiosity-driven learning. In The New Science of Curiosity (ed. Goren Gordon), 43-72 (Nova Science Publishers, 2018); https://arxiv.org/abs/1802.10546

Colombo, M. & Wright, C. First principles in the life sciences: the free-energy principle, organicism and mechanism. Synthese 198 , 3463–3488 (2021).

Article   MathSciNet   Google Scholar  

Huang, Y. & Rao, R. P. Predictive coding. WIREs Cogn. Sci. 2 , 580–593 (2011).

Friston, K. The free-energy principle: a unified brain theory? Nat. Rev. Neurosci. 11 , 127–138 (2010).

Knill, D. C. & Pouget, A. The Bayesian brain: the role of uncertainty in neural coding and computation. Trends Neurosci. 27 , 712–719 (2004).

Clark, A. Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behav. Brain Sci. 36 , 181–204 (2013).

Clark, A. Surfing Uncertainty : Prediction , Action and the Embodied Mind (Oxford Univ. Press, 2016).

Zacks, J. M., Speer, N. K., Swallow, K. M., Braver, T. S. & Reyonolds, J. R. Event perception: a mind/brain perspective. Psychol. Bull. 133 , 273–293 (2007).

Eysenbach, B., Ibarz, J., Gupta, A. & Levine, S. Diversity is all you need: learning skills without a reward function. In International Conference on Learning Representations (ICLR, 2019).

Frans, K., Ho, J., Chen, X., Abbeel, P. & Schulman, J. Meta learning shared hierarchies. In Proc. International Conference on Learning Representations https://openreview.net/pdf?id=SyX0IeWAW (ICLR, 2018).

Heess, N. et al. Learning and transfer of modulated locomotor controllers. Preprint at https://arxiv.org/abs/1610.05182 (2016).

Jiang, Y., Gu, S., Murphy, K. & Finn, C. Language as an abstraction for hierarchical deep reinforcement learning. In Neural Information Processing Systems ( NeurIPS ) (eds. Wallach, H. et al.) 9414–9426 (ACM, 2019).

Li, A. C., Florensa, C., Clavera, I. & Abbeel, P. Sub-policy adaptation for hierarchical reinforcement learning. In Proc. International Conference on Learning Representations https://openreview.net/forum?id=ByeWogStDS (ICLR, 2020).

Qureshi, A. H. et al. Composing task-agnostic policies with deep reinforcement learning. In Proc. International Conference on Learning Representations https://openreview.net/forum?id=H1ezFREtwH (ICLR, 2020).

Sharma, A., Gu, S., Levine, S., Kumar, V. & Hausman, K. Dynamics-aware unsupervised discovery of skills. In Proc. International Conference on Learning Representations https://openreview.net/forum?id=HJgLZR4KvH (ICLR, 2020).

Tessler, C., Givony, S., Zahavy, T., Mankowitz, D. J. & Mannor, S. A deep hierarchical approach to lifelong learning in minecraft. In Proc. 31st AAAI Conference on Artificial Intelligence 1553–1561 (AAAI, 2017).

Vezhnevets, A. et al. Strategic attentive writer for learning macro-actions. In Neural Information Processing Systems ( NIPS ) (eds. Lee, D. et al.) 3494–3502 (NIPS, 2016).

Devin, C., Gupta, A., Darrell, T., Abbeel, P. & Levine, S. Learning modular neural network policies for multi-task and multi-robot transfer. In Proc. International Conference on Robotics and Automation ( ICRA ) (eds. Okamura, A. et al.) 2169–2176 (IEEE, 2017).

Hejna, D. J., Abbeel, P. & Pinto, L. Hierarchically decoupled morphological transfer. In Proc. International Conference on Machine Learning ( ICML ) (eds. Daumé III, H. & Singh, A.) 11409–11420 (PMLR, 2020).

Hamrick, J. B. et al. On the role of planning in model-based deep reinforcement learning. In Proc. International Conference on Learning Representations https://openreview.net/pdf?id=IrM64DGB21 (ICLR, 2021).

Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proc. 7th International Conference on Machine Learning ( ICML ) (eds. Porter, B. W. & Mooney, R. J.) 216–224 (Morgan Kaufmann, 1990).

Nau, D. et al. SHOP2: an HTN planning system. J. Artif. Intell. Res. 20 , 379–404 (2003).

Article   MATH   Google Scholar  

Lyu, D., Yang, F., Liu, B. & Gustafson, S. SDRL: interpretable and data-efficient deep reinforcement learning leveraging symbolic planning. In Proc. AAAI Conference on Artificial Intelligence Vol. 33, 2970–2977 (AAAI, 2019).

Ma, A., Ouimet, M. & Cortés, J. Hierarchical reinforcement learning via dynamic subspace search for multi-agent planning. Auton. Robot. 44 , 485–503 (2020).

Bacon, P.-L., Harb, J. & Precup, D. The option-critic architecture. In Proc. 31st AAAI Conference on Artificial Intelligence 1726–1734 (AAAI, 2017).

Dietterich, T. G. State abstraction in MAXQ hierarchical reinforcement learning. In Advances in Neural Information Processing Systems ( NIPS ) (eds. Solla, S. et al.) Vol. 12, 994–1000 (NIPS, 1999).

Kulkarni, T. D., Narasimhan, K. R., Saeedi, A. & Tenenbaum, J. B. Hierarchical deep reinforcement learning: integrating temporal abstraction and intrinsic motivation. In Neural Information Processing Systems ( NIPS ) (eds. Lee, D. et al.) 3675–3683 (NIPS, 2016).

Shankar, T., Pinto, L., Tulsiani, S. & Gupta, A. Discovering motor programs by recomposing demonstrations. In Proc. International Conference on Learning Representations https://openreview.net/attachment?id=rkgHY0NYwr&name=original_pdf (ICLR, 2020).

Vezhnevets, A. S., Wu, Y. T., Eckstein, M., Leblond, R. & Leibo, J. Z. Options as responses: grounding behavioural hierarchies in multi-agent reinforcement learning. In Proc. International Conference on Machine Learning ( ICML ) (eds. Daumé III, H. & Singh, A.) 9733–9742 (PMLR, 2020).

Ghazanfari, B., Afghah, F. & Taylor, M. E. Sequential association rule mining for autonomously extracting hierarchical task structures in reinforcement learning. IEEE Access 8 , 11782–11799 (2020).

Levy, A., Konidaris, G., Platt, R. & Saenko, K. Learning multi-level hierarchies with hindsight. In Proc. International Conference on Learning Representations https://openreview.net/pdf?id=ryzECoAcY7 (ICLR, 2019).

Nachum, O., Gu, S., Lee, H. & Levine, S. Data-efficient hierarchical reinforcement learning. In Proc. 32nd International Conference on Neural Information Processing Systems (NIPS) (eds. Bengio, S. et al.) 3307–3317 (NIPS, 2018).

Rafati, J. & Noelle, D. C. Learning representations in model-free hierarchical reinforcement learning. In Proc. 33rd AAAI Conference on Artificial Intelligence 10009–10010 (AAAI, 2019).

Röder, F., Eppe, M., Nguyen, P. D. H. & Wermter, S. Curious hierarchical actor-critic reinforcement learning. In Proc. International Conference on Artificial Neural Networks ( ICANN ) (eds. Farkaš, I. et al.) 408–419 (Springer, 2020).

Zhang, T., Guo, S., Tan, T., Hu, X. & Chen, F. Generating adjacency-constrained subgoals in hierarchical reinforcement learning. In Neural Information Processing Systems ( NIPS ) (eds. Larochelle, H. et al.) 21579-21590 (NIPS, 2020).

Lample, G. & Chaplot, D. S. Playing FPS games with deep reinforcement learning. In Proc. 31st AAAI Conference on Artificial Intelligence 2140–2146 (AAAI, 2017).

Vezhnevets, A. S. et al. FeUdal networks for hierarchical reinforcement learning. In Proc. 34th International Conference on Machine Learning ( ICML ) (eds. Precup, D. & Teh, Y. W.) Vol. 70, 3540–3549 (PMLR, 2017).

Wulfmeier, M. et al. Compositional Transfer in Hierarchical Reinforcement Learning. In Robotics: Science and System XVI (RSS) (eds. Toussaint M. et al.) (Robotics: Science and Systems Foundation, 2020); https://arxiv.org/abs/1906.11228

Yang, Z., Merrick, K., Jin, L. & Abbass, H. A. Hierarchical deep reinforcement learning for continuous action control. IEEE Trans. Neural Netw. Learn. Syst. 29 , 5174–5184 (2018).

Toussaint, M., Allen, K. R., Smith, K. A. & Tenenbaum, J. B. Differentiable physics and stable modes for tool-use and manipulation planning. In Proc. Robotics : Science and Systems XIV ( RSS ) (eds. Kress-Gazit, H. et al.) https://ipvs.informatik.uni-stuttgart.de/mlr/papers/18-toussaint-RSS.pdf (Robotics: Science and Systems Foundation, 2018).

Akrour, R., Veiga, F., Peters, J. & Neumann, G. Regularizing reinforcement learning with state abstraction. In Proc. IEEE / RSJ International Conference on Intelligent Robots and Systems ( IROS ) 534–539 (IEEE, 2018).

Schaul, T. & Ring, M. Better generalization with forecasts. In Proc. 23rd International Joint Conference on Artificial Intelligence ( IJCAI ) (ed. Rossi, F.) 1656–1662 (AAAI, 2013).

Colas, C., Akakzia, A., Oudeyer, P.-Y., Chetouani, M. & Sigaud, O. Language-conditioned goal generation: a new approach to language grounding for RL. Preprint at https://arxiv.org/abs/2006.07043 (2020).

Blaes, S., Pogancic, M. V., Zhu, J. J. & Martius, G. Control what you can: intrinsically motivated task-planning agent. Neural Inf. Process. Syst. 32 , 12541–12552 (2019).

Haarnoja, T., Hartikainen, K., Abbeel, P. & Levine, S. Latent space policies for hierarchical reinforcement learning. In Proc. International Conference on Machine Learning ( ICML ) (eds. Dy, J. & Krause, A.) Vol. 4, 2965–2975 (PMLR, 2018).

Rasmussen, D., Voelker, A. & Eliasmith, C. A neural model of hierarchical reinforcement learning. PLoS ONE 12 , e0180234 (2017).

Riedmiller, M. et al. Learning by playing—solving sparse reward tasks from scratch. In Proc. International Conference on Machine Learning ( ICML ) (eds. Dy, J. & Krause, A.) Vol. 10, 6910–6919 (PMLR, 2018).

Yang, F., Lyu, D., Liu, B. & Gustafson, S. PEORL: integrating symbolic planning and hierarchical reinforcement learning for robust decision-making. In Proc. 27th International Joint Conference on Artificial Intelligence ( IJCAI ) (ed. Lang, J.) 4860–4866 (IJCAI, 2018).

Machado, M. C., Bellemare, M. G. & Bowling, M. A Laplacian framework for option discovery in reinforcement learning. In Proc. International Conference on Machine Learning (ICML) (eds. Precup, D. & Teh, Y. W.) Vol. 5, 3567–3582 (PMLR, 2017).

Pathak, D., Agrawal, P., Efros, A. A. & Darrell, T. Curiosity-driven exploration by self-supervised prediction. In Proc. 34th International Conference on Machine Learning ( ICML ) (eds. Precup, D. & Teh, Y. W.) 2778–2787 (PMLR, 2017).

Schillaci, G. et al. Intrinsic motivation and episodic memories for robot exploration of high-dimensional sensory spaces. Adaptive Behav. 29 549–566 (2020).

Colas, C., Fournier, P., Sigaud, O., Chetouani, M. & Oudeyer, P.-Y. CURIOUS: intrinsically motivated modular multi-goal reinforcement learning. In Proc. International Conference on Machine Learning ( ICML ) (eds. Chaudhuri, K. & Salakhutdinov, R.) 1331–1340 (PMLR, 2019).

Hafez, M. B., Weber, C., Kerzel, M. & Wermter, S. Improving robot dual-system motor learning with intrinsically motivated meta-control and latent-space experience imagination. Robot. Auton. Syst. 133 , 103630 (2020).

Yamamoto, K., Onishi, T. & Tsuruoka, Y. Hierarchical reinforcement learning with abductive planning. In Proc. ICML / IJCAI / AAMAS 2018 Workshop on Planning and Learning ( PAL-18 ) (2018).

Wu, B., Gupta, J. K. & Kochenderfer, M. J. Model primitive hierarchical lifelong reinforcement learning . In Proc. International Joint Conference on Autonomous Agents and Multiagent Systems ( AAMAS ) (eds. Agmon, N. et al.) Vol. 1, 34–42 (IFAAMAS, 2019).

Li, Z., Narayan, A. & Leong, T. Y. An efficient approach to model-based hierarchical reinforcement learning. In Proc. 31st AAAI Conference on Artificial Intelligence 3583–3589 (AAAI, 2017).

Hafner, D., Lillicrap, T. & Norouzi, M. Dream to control: learning behaviors by latent imagination. In Proc. International Conference on Learning Representations https://openreview.net/pdf?id=S1lOTC4tDS (ICLR, 2020).

Deisenroth, M. P., Rasmussen, C. E. & Fox, D. Learning to control a low-cost manipulator using data-efficient reinforcement learning. In Robotics : Science and Systems VII ( RSS ) (eds. Durrant-Whyte, H. et al.) 57–64 (Robotics: Science and Systems Foundation, 2011).

Ha, D. & Schmidhuber, J. Recurrent world models facilitate policy evolution. In Proc. 32nd International Conference on Neural Information Processing Systems (NeurIPS) (eds. Bengio, S. et al.) 2455–2467 (NIPS, 2018).

Battaglia, P. W. et al. Relational inductive biases, deep learning and graph networks. Preprint at https://arxiv.org/abs/1806.01261 (2018).

Andrychowicz, M. et al. Hindsight experience replay. In Proc. Neural Information Processing Systems ( NIPS ) (eds. Guyon I. et al.) 5048–5058 (NIPS, 2017); https://papers.nips.cc/paper/7090-hindsight-experience-replay.pdf

Schwartenbeck, P. et al. Computational mechanisms of curiosity and goal-directed exploration. eLife 8 , e41703 (2019).

Haarnoja, T., Zhou, A., Abbeel, P. & Levine, S. Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor. In Proc. International Conference on Machine Learning ( ICML ) (eds. Dy, J. & Krause, A.) 1861–1870 (PMLR, 2018).

Yu, A. J. & Dayan, P. Uncertainty, neuromodulation and attention. Neuron 46 , 681–692 (2005).

Baldwin, D. A. & Kosie, J. E. How does the mind render streaming experience as events? Top. Cogn. Sci. 13 , 79–105 (2021).

Download references

Acknowledgements

We acknowledge funding from the DFG (projects IDEAS, LeCAREbot, TRR169, SPP 2134, RTG 1808 and EXC 2064/1), the Humboldt Foundation and Max Planck Research School IMPRS-IS.

Author information

Manfred Eppe

Present address: Hamburg University of Technology, Hamburg, Germany

Authors and Affiliations

Universität Hamburg, Hamburg, Germany

Manfred Eppe, Matthias Kerzel, Phuong D. H. Nguyen & Stefan Wermter

University of Tübingen, Tübingen, Germany

Christian Gumbsch & Martin V. Butz

Max Planck Institute for Intelligent Systems, Tübingen, Germany

Christian Gumbsch

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Manfred Eppe .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary information.

Supplementary Boxes 1–6 and Table 1.

Rights and permissions

Reprints and permissions

About this article

Cite this article.

Eppe, M., Gumbsch, C., Kerzel, M. et al. Intelligent problem-solving as integrated hierarchical reinforcement learning. Nat Mach Intell 4 , 11–20 (2022). https://doi.org/10.1038/s42256-021-00433-9

Download citation

Received : 18 December 2020

Accepted : 07 December 2021

Published : 25 January 2022

Issue Date : January 2022

DOI : https://doi.org/10.1038/s42256-021-00433-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Efficient stacking and grasping in unstructured environments.

• Jinbiao Zhu

Journal of Intelligent & Robotic Systems (2024)

Four attributes of intelligence, a thousand questions

• Matthieu Bardal
• Eric Chalmers

Biological Cybernetics (2023)

An Alternative to Cognitivism: Computational Phenomenology for Deep Learning

• Pierre Beckmann
• Guillaume Köstner
• Inês Hipólito

Minds and Machines (2023)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Scientific Reports

Foundations of human spatial problem solving

Department of Psychological and Brain Sciences, Indiana University, Bloomington, USA

Joshua W. Brown

Associated data.

The GOLSA model code for the simulations is available at https://github.com/CogControl/GolsaOrigTreasureHunt . Imaging data are available from the corresponding author on reasonable request.

Despite great strides in both machine learning and neuroscience, we do not know how the human brain solves problems in the general sense. We approach this question by drawing on the framework of engineering control theory. We demonstrate a computational neural model with only localist learning laws that is able to find solutions to arbitrary problems. The model and humans perform a multi-step task with arbitrary and changing starting and desired ending states. Using a combination of computational neural modeling, human fMRI, and representational similarity analysis, we show here that the roles of a number of brain regions can be reinterpreted as interacting mechanisms of a control theoretic system. The results suggest a new set of functional perspectives on the orbitofrontal cortex, hippocampus, basal ganglia, anterior temporal lobe, lateral prefrontal cortex, and visual cortex, as well as a new path toward artificial general intelligence.

Introduction

Great strides have been made recently toward solving hard problems with deep learning, including reinforcement learning 1 , 2 . While these are groundbreaking and show superior performance over humans in some domains, humans nevertheless exceed computers in the ability to find creative and efficient solutions to novel problems, especially with changing internal motivation values 3 . Artificial general intelligence (AGI), especially the ability to learn autonomously to solve arbitrary problems, remains elusive 4 .

Value-based decision-making and goal-directed behavior involve a number of interacting brain regions, but how these regions might work together computationally to generate goal directed actions remains unclear. This may be due in part to a lack of mechanistic theoretical frameworks 5 , 6 . The orbitofrontal cortex (OFC) may represent both a cognitive map 7 and a flexible goal value representation 8 , driving actions based on expected outcomes 9 , though how these guide action selection is still unclear. The hippocampus is important for model-based planning 10 and prospection 11 , and the striatum is important for action selection 12 . Working memory for visual cues and task sets seems to depend on the visual cortex and lateral prefrontal regions, respectively 13 , 14 .

Neuroscience continues to reveal aspects of how the brain might learn to solve problems. Studies of cognitive control highlight how the brain, especially the prefrontal cortex, can apply and update rules to guide behavior 15 , 16 , inhibit behavior 17 , and monitor performance 18 to detect and correct errors 19 . Still, there is a crucial difference between rules and goals. Rules define a mapping from a stimulus to a response 20 , but goals define a desired state of the individual and the world 21 . When cognitive control is re-conceptualized as driving the individual to achieve a desired state, or set point, then cognitive control becomes a problem amenable to control theory.

Control theory has been applied to successfully account for the neural control of movement 22 and has informed various aspects of neuroscience research, including work in C. Elegans 23 , and work on controlling states of the brain 24 and electrical stimulation placement methods 25 (as distinct from behavioral control over states of the world in the present work), and more loosely in terms of neural representations underlying how animals control an effector via a brain computer interface 26 . In Psychology, Perceptual Control Theory has long maintained that behavior is best understood as a means of controlling perceptual input in the sense of control theory 27 , 28 .

In the control theory framework, a preferred decision prospect will define a set point, to be achieved by control-theoretic negative feedback controllers 29 , 30 . Problem solving then requires 1) defining the goal state; 2) planning a sequence of state transitions to move the current state toward the goal; and 3) generating actions aimed at implementing the desired sequence of state transitions.

Algorithms already exist that can implement such strategies, including the Dijkstra and A* algorithms 31 , 32 and are commonly used in GPS navigation devices found in cars and cell phones. Many variants of reinforcement learning solve a specific case of this problem, in which the rewarded states are relatively fixed, such as winning a game of Go 33 . While deep Q networks 1 and generative adversarial networks with monte carlo tree search 33 are very powerful, what happens when the goals change, or the environmental rules change? In that case, the models may require extensive retraining. The more general problem requires the ability to dynamically recalculate the values associated with each state as circumstances, goals, and set points change, even in novel situations.

Here we explore a computational model that solves this more general problem of how the brain solves problems with changing goals 34 , and we show how a number of brain regions may implement information processing in ways that correspond to specific model components. While this may seem an audacious goal, our previous work has shown how the GOLSA model can solve problems in the general sense of causing the world to assume a desired state via a sequence of actions, as described above 34 . The model begins with a core premise: the brain constitutes a control-theoretic system, generating actions to minimize the discrepancy between actual and desired states. We developed the Goal-Oriented Learning and Selection of Action (GOLSA) computational neural model from this core premise to simulate how the brain might autonomously learn to solve problems, while maintaining fidelity to known biological mechanisms and constraints such as localist learning laws and real-time neural dynamics. The constraints of biological plausibility both narrow the scope of viable models and afford a direct comparison with neural activity.

The model treats the brain as a high-dimensional control system. It drives behavior to maintain multiple and varying control theoretic set points of the agent’s state, including low level homeostatic (e.g. hunger, thirst) and high level cognitive set points (e.g. a Tower of Hanoi configuration). The model autonomously learns the structure of state transitions, then plans actions to arbitrary goals via a novel hill-climbing algorithm inspired by Dijkstra’s algorithm 32 . The model provides a domain-general solution to the problem of solving problems and performs well in arbitrary planning tasks (such as the Tower of Hanoi) and decision-making problems involving multiple constraints 34 (“ Methods ”).

The GOLSA model works by representing each possible state of the agent and environment in a network layer, with multiple layers each representing the same sets of states (Fig.  1 A,B). The Goal Gradient layer is activated by an arbitrarily specified desired (Goal) state and spreads activation backward along possible state transitions represented as edges in the network 35 , 36 . This value spreading activation generates current state values akin to learned state values (Q values) in reinforcement learning, except that the state values can be reassigned and recalculated dynamically as goals change. This additional flexibility allows goals to be specified dynamically and arbitrarily, with all state values being updated immediately to reflect new goals, thus overcoming a limitation of current RL approaches. Essentially, the Goal Gradient is the hill to climb to minimize the discrepancy between actual and desired states in the control theoretic sense. In parallel, regarding the present state of the model system, the Adjacent States layer receives input from a node representing the current state of the agent and environment, which in turn activates representations of all states that can be achieved with one state transition. The valid adjacent states then mask the Goal Gradient layer to yield the Desired Next State representation. In this layer, the most active unit represents a state which, if achieved, will move the agent one step closer to the goal state. This desired next state is then mapped onto an action (i.e. a controller signal) that is likely to effect the desired state transition. In sum, the model is given an arbitrarily specified goal state and the actual current state of the actor. It then finds an efficient sequence of states to transit in order to reach the goal state, and it generates actions aimed at causing the current state of the world to be updated so that it approaches and reaches the goal state.

( A ) The GOLSA model determines the next desired state by hill climbing. Each layer represents the same set of states, one per neuron. The x- and y-axes of the grids represent abstracted coordinates in a space of states. Neurons are connected to each other for states that are reachable from another by one action, in this case neighbors in the x,y plane. The Goal state is activated and spreads activation through a Goal Gradient (Proximity) layer, thus dynamically specifying the value of each state given the goal, so that value is greater for states nearer the goal state. The Current State representation activates all Adjacent States, i.e. that can be achieved with one state transition. These adjacent states mask the Goal Gradient input to the Desired Next State, so that the most active unit in the Desired Next State represents a state attainable with one state transition and which will bring the state most directly toward the goal state. The black arrows indicate that the Desired Next State unit activities are the element-wise products of the corresponding Adjacent States and Goal Gradient unit activities. The font colors match the model layer to corresponding brain regions in Figs. ​ Figs.3 3 and ​ and4. 4 . ( B ) The desired state transition is determined by the conjunction of current state and desired next state. The GOLSA model learns a mapping from desired state transitions to the actions that cause those transitions. After training, the model can generate novel action sequences to achieve arbitrary goal states. Adapted from 34 .

Here we test whether and how the GOLSA model might provide an account of how various brain regions work together to drive goal-directed behavior. To do this, we ask human subjects to perform a multi-step task to achieve arbitrary goals. We then train the GOLSA model to perform the same task, and we use representational similarity analysis (RSA) to ask whether specific GOLSA model layers show similar representations to specific brain regions ( Supplementary Material ). The results will provide a tentative account of the function of specific brain regions in terms of the GOLSA model, and this account can then be tested and compared against alternative models in future work.

Study design

The details of the model implementation and the model code are available in the “ Methods ”. Behaviorally, we found that the GOLSA model is able to learn to solve arbitrary problems, such as reaching novel states in the Tower of Hanoi task (Fig.  2 A). It does this without hard-wired knowledge, simply by making initially random actions and learning from the outcomes, then synthesizing the learned information to achieve whatever state is specified as the goal state.

( A ) The GOLSA model learns to solve problems, achieving arbitrary goal states. It does this by making arbitrary actions and observing which actions cause which state transitions. Figure adapted from earlier work 34 , 37 . ( B ) Treasure Hunt task. Both the GOLSA model and the human fMRI subjects performed a simple treasure hunt task, in which subjects were placed in one of four possible starting locations, then asked to generate actions to reach any of the other possible locations. To test multi-step transitions, subjects had to first move to the location of a key needed to unlock a treasure chest, then move to the treasure chest location. Participants first saw an information screen specifying the contents of each of the four states (‘you’, ‘key’, ‘chest’, or ‘nothing’). After a jittered delay, participants selected a desired movement direction and after another delay saw an image of the outcome location. The mapping of finger buttons to game movements was random on each trial and revealed after subjects were given the task and had to plan their movements, thus avoiding motor confounds during planning. Bottom: The two state-space maps used in the experiment. One map was used in the first half of trials while the other was used in the second half, in counterbalanced order.

Having found that the model can learn autonomously to solve arbitrary problems, we then aimed to identify which brain regions might show representations and activity that matched particular GOLSA model layers. To do this, we tested the GOLSA model with a Treasure Hunt task (Fig.  2 B and “ Methods ”), which was performed by both the GOLSA model and human subjects with fMRI. All human subjects research here was approved by the IRB of Indiana University, and subjects gave full informed consent. The human subjects research was performed in accordance with relevant guidelines/regulations and in accordance with the Declaration of Helsinki. Subjects were placed in one of four starting states and had to traverse one or two states to achieve a goal, by retrieving a key and subsequently using it to unlock a treasure chest for a reward (Fig.  2 B). The Treasure Hunt task presents a challenge to standard RL approaches, because the rewarded (i.e. goal) state changes regularly. In an RL framework, the Bellman equation would regularly relearn the value of each possible state in terms of how close it is to the currently rewarded state, forgetting previous state values in the process.

Representational similarity analysis

To analyze the fMRI and model data, we used model-based fMRI with representational similarity analysis (RSA) 38 (“ Methods ”). RSA considers a set of task conditions and asks whether a model, or brain region, can discriminate between the patterns of activity associated with the two conditions, as measured by a correlation coefficient. By considering every possible pairing of conditions, the RSA method constructs a symmetric representational dissimilarity matrix (RDM), where each entry is 1-r, and r is the correlation coefficient. This RDM provides a representational fingerprint of what information is present, so that the fingerprints can be compared between a model layer and a given brain region. For our application of RSA, each representational dissimilarity matrix (RDM) represented the pairwise correlations across 96 total patterns–4 starting states by 8 trial types by 3 time points within a trial (problem description, response, and feedback). For each model layer, the pairwise correlations are calculated with the activity pattern across layer cells in one condition vs. the activity pattern in the same layer in the other condition. For each voxel in the brain, the pairwise correlations are calculated with the activity pattern in a local neighborhood of radius 10 mm (93 voxels total) around the voxel in question, for one condition vs. the other condition. The 10 mm radius was chosen to provide a tradeoff between a sufficiently high number of voxels for pattern analysis and a sufficiently small area to identify specific regions. The fMRI RSA maps are computed for each subject over all functional scans and then tested across subjects for statistical significance. The comparison between GOLSA model and fMRI RDMs consists of looking for positive correlations between elements of the upper symmetric part of a given GOLSA model layer RDM vs. the RDM around a given voxel in the fMRI RDMs. The resulting fMRI RSA maps, one per GOLSA model layer, show which brain regions have representational similarities between particular model components and particular brain regions. The fMRI RSA maps showing the similarities between a given GOLSA model layer and a given brain region are computed for each subject and then tested across subjects for statistical significance in a given brain region, with whole-brain tests for significance in all cases. Full results are in Table ​ Table2, 2 , and method details are in the “ Methods ” section. As a control, we also generated a null model layer that consisted of normally distributed noise (μ = 1, σ = 1). In the null model, no voxels exceeded the cluster defining threshold, and so no significant clusters were found, which suggests that the results below are not likely to reflect artifacts of the analysis methods.

Full list of core model layers and associated parameters.

Orbitofrontal cortex, goals, and maps

We found that the patterns of activity in a number of distinct brain regions match those expected of a control theoretic system, as instantiated in the GOLSA model (Figs. ​ (Figs.3A,B 3 A,B and ​ and4A–C); 4 A–C); Table ​ Table1). 1 ). Orbitofrontal cortex (OFC) activity patterns match model components that represent both a cognitive map 7 and a flexible goal value representation 8 , specifically matching the Goal and Goal Gradient layer activities. These layers represent the current values of the goal state and the current values of states near the goal state, respectively. The Goal Gradient layer incorporates cognitive map information in terms of which states can be reached from which other states. This suggests mechanisms by which OFC regions may calculate the values of states dynamically as part of a value-based decision process, by spreading activation of value from a currently active goal state representation backward. The GOLSA model representations of the desired next state also match overlapping regions in the orbitofrontal cortex (OFC) and ventromedial prefrontal cortex (vmPFC), consistent with a role in finding the more valuable decision option (Fig.  3 ). Reversal learning and satiety effects as supported by the OFC reduce to selecting a new goal state or deactivating a goal state respectively, which immediately updates the values of all states. Collectively this provides a mechanistic account of how value-based decision-making functions in OFC and vmPFC.

Representational Similarity Analysis (RSA) of model layers vs. human subjects performing the same Treasure Hunt task. All results shown are significant clusters across the population with a cluster defining threshold of p  < 0.001 cluster corrected to p  < 0.05 overall, and with additional smoothing of 8 mm FWHM applied prior to the population level t-test for visualization purposes. ( A ) population Z maps showing significant regions of similarity to model layers in orbitofrontal cortex. Cf. Figure  1 and Fig.  5 B. The peak regions of similarity for goal-gradient and goal show considerable overlap in right OFC. The region of peak similarity for simulated-state is more posterior. To most clearly show peaks of model-image correspondence, the maps of gradient and goal are here visualized at p  < 0.00001 while all others are visualized at p  < 0.001. ( B ) Z maps showing significant regions of similarity to model layers in right temporal cortex. The peak regions of similarity for goal-gradient and goal overlap and extend into the OFC. The peak regions of similarity for adjacent-states, next-desired-state, and -simulated-state occur in similar but not completely overlapping regions, while the cluster for queue-store is more lateral. ( C ) Fig.  1 A, copied here as a legend, where the font color of each layer name corresponds to the region colors in panels ( A) and ( B) .

Representational Similarity Analysis of model layers vs. human subjects performing the same Treasure Hunt task, with the same conditions and RSA analysis as in Fig.  3 . ( A ) Population Z maps showing significant regions of similarity to model layers in visual cortex. The peak regions of similarity for goal-gradient and goal overlap substantially, primarily in bilateral cuneus, inferior occipital gyrus, and lingual gyrus. The simulated-state layer displayed significantly similar activity to that in a smaller medial and posterior region. Statistical thresholding and significance are the same as Fig.  3 . ( B ) Z map showing significant regions of similarity to the desired-transition layer. Similarity peaks were observed for desired-transition in bilateral hippocampal gyrus as well as bilateral caudate and putamen. The desired-transition map displayed here was visualized at p  < 0.00001 for clarity. ( C ) Z maps showing significant regions of similarity to the model layers in frontal cortex. Similarity peaks were observed for queue-store in superior frontal gyrus (BA10). Action-output activity most closely resembled activity in inferior frontal gyrus (BA9), while simulated-state and goal-gradient patterns of activity were more anterior (primarily BA45). Similarity between activity in the latter two layers and activity in OFC, visual cortex, and temporal pole is also visible.

Significant similarity clusters for RSA analysis. The p and Size columns refer to cluster-corrected values. Anatomical labels are derived from the Automated Anatomical Labeling Atlas in SPM5 46 .

Lateral PFC and planning

The GOLSA model also incorporates a mechanism that allows multi-step planning, by representing a Simulated State as if the desired next state were already achieved, so that the model can plan multiple subsequent state transitions iteratively prior to committing to a particular course of action (Fig.  5 B). Those subsequent state transitions are represented in a Queue Store layer pending execution via competitive queueing, in which the most active action representation is the first to be executed, followed by the next most active representation, and so on 39 , 40 . This constitutes a mechanism of prospection 41 and planning 42 . The Simulated State layer in the GOLSA model shows strong representational similarity with regions of the OFC and anterior temporal lobe, and the Queue Store layer shows strong similarity with the anterior temporal lobe and lateral prefrontal cortex. This constitutes a mechanistic account of how the vmPFC and OFC in particular might contribute to multi-step goal-directed planning, and how plans may be stored in lateral prefrontal cortex.

( A ) Full diagram of core model. Each rectangle represents a layer and each arrow a projection. The body is a node, and two additional nodes are not shown which provide inhibition at each state-change and oscillatory control. The colored squares indicate which layers receive inhibition from these nodes. Some recurrent connections not shown. ( B ) Full diagram of extended model, with added top row representing ability to plan multiple state transition steps ahead (Simulated State, Queue Input, Queue Store, and Queue Output layers). Adapted with permission from earlier work 34 .

Visual cortex and future visual states

The visual cortex also shows representational patterns consistent with representing the goal, goal gradient, and simulated future states (Figs.  3 B and ​ and4). 4 ). This suggests a role for the visual cortex in planning, in the sense of representing anticipated future states beyond simply representing current visual input. Future states in the present task are represented largely by images of locations, such as an image of a scarecrow or a house. In that sense, an anticipated future state could be decoded as matching the representation of the image of that future state. One possibility is that this reflects an attentional effect that facilitates processing of visual cues representing anticipated future states. Another possibility is that visual cortex activity signals a kind of working memory for anticipated future visual states, similar to how working memory for past visual states has been decoded from visual cortex activity 14 . This would be distinct from predictive coding, in that the activity predicts future states, not current states 43 . In either case, the results are consistent with the notion that the visual cortex may not be only a sensory region but may play some role in planning by representing the details of anticipated future states.

Anterior temporal lobe and planning

The anterior temporal lobe likewise shows representations of the goal, goal gradient, the adjacent states, the next desired state, and simulated future and queue store states (Figs.  3 B, ​ B,4C). 4 C). In one sense this is not surprising, as the states of the task are represented by images of objects, and visual objects (especially faces) are represented in the anterior temporal lobe 44 . Still, the fact that the anterior temporal lobe shows representations consistent with planning mechanisms suggests a more active role in planning beyond feedforward sensory processing as commonly understood 45 .

Hippocampal region and prospection

Once the desired next state is specified, it must be translated to an action. The hippocampus and striatum match the representations of the Desired Transition layer in the GOLSA model. This model layer represents a conjunction of the current state and desired next state transitions, which in the GOLSA model is a necessary step toward selecting an appropriate action to achieve the desired transition. This is consistent with the role of the hippocampus in prospection 41 , and it suggests computational and neural mechanisms by which the hippocampus may play a key role in turning goals into predictions about the future, for the purpose of planning actions 10 , 11 . Finally, as would be expected, the motor output representations in the GOLSA model match motor output patterns in the motor cortex (Fig.  4 C).

The results above show how a computational neural model, the GOLSA model, provides a novel computational account of a number of brain regions. The guiding theory is that a substantial set of brain regions function together as a control-theoretic mechanism 47 , generating behaviors to minimize the discrepancy between the current state and the desired (goal) state. The OFC is understood as including neurons that represent the value of various states in the world, such as the value of acquiring certain objects. Greater activity of an OFC neuron corresponds with more value of its represented state given the current goals. Because of spreading activation, neurons will be more active if they represent states closer to the goal. This results in value representations similar to those provided by the Bellman equation of reinforcement learning 48 , with the difference being that spreading activation can instantly reconfigure the values of states as goals change, without requiring extensive iterations of the Bellman equation.

Given the current state and the goal state, the next desired state can be determined as a nearby state that can be reached and that also moves the current state of the world closer to the goal state. Table ​ Table1 1 shows these effects in the medial frontal gyrus, putamen, superior temporal gyrus, pons, and precuneus. The GOLSA model suggests this is computed as the activation of available state representations, multiplied by the OFC value for that state. Precedent for this kind of multiplicative effect has been shown in the attention literature 49 . The action to be generated is represented by neural activity in the motor cortex region. This in turn is determined on the basis of neurons that are active specifically for a conjunction of the particular current state and next desired state. Neurally, we find this conjunction represented across a large region including the striatum and hippocampus. This is consistent with the notion of the hippocampus as a generative recurrent neural network, that starts at a current state and runs forward, specifically toward the desired state 50 . The striatum is understood as part of an action gate that permits certain actions in specific contexts, although the GOLSA model does not include an explicit action gate 51 . Where multiple action steps must be planned prior to executing any of them, the lateral PFC seems to represent a queue of action plans in sequence, as sustained activity representing working memory 39 , 52 . By contrast, working memory representations in the visual cortex apparently represent the instructed future states as per the instructions for each task trial, and these are properly understood as visual sensory rather than motor working memories 14 .

Our findings overall bear a resemblance to the Free Energy principle. According to this, organisms learn to generate predictions of the most likely (i.e. rewarding) future states under a policy, then via active inference emit actions to cause the most probable outcomes to become reality, thus minimizing surprise 53 , 54 . Like active inference, the GOLSA model emits actions to minimize the discrepancy between the actual and predicted state. Of note, the GOLSA model specifies the future state as a desired state rather than a most likely state. This crucial distinction allows a state that has a high current value to be pursued, even if the probability of being in that state is very low (for example buying lottery tickets and winning). Furthermore, the model includes the mechanisms of Fig.  1 , which allow for flexible planning given arbitrary goals. The GOLSA model is a process model and simulates rate-coded neural activity as a dynamical system (“ Methods ”), which affords the more direct comparison with neural activity representations over time as in Figs.  3 and ​ and4 4 .

The GOLSA model, and especially our analysis of it, builds on recent work that developed methods to test computational neural models against empirical data. Substantial previous work has demonstrated how computational neural modeling can provide insight into the functional properties underlying empirical neural data, such as recurrent neural networks elucidating the representational structure in anterior cingulate 19 , 55 , 56 and PFC 57 ; deep neural networks accounting for object recognition in IT with representational similarity analysis 58 , and encoding/decoding of visual cortex representations 59 ; dimensionality reduction for comparing neural recordings and computational neural models 60 , and representations of multiple learned tasks in computational neural models 61 .

The GOLSA model shares some similarity with model-based reinforcement learning (MBRL), in that both include learned models of next-state probabilities as a function of current state and action pairs. Still, a significant limitation of both model-based and model free RL is that typically there is only a single ultimate goal, e.g. gaining a reward or winning a game. Q-values 62 are thus learned in order to maximize a single reward value. This implies several limitations: (1) that Q values are strongly paired with corresponding states; (2) that there is only one Q value per state at a given time, as in a Markov decision process (MDP), and (3) Q values are generally updated via substantial relearning. In contrast, real organisms will find differing reward values associated with different goals at different times and circumstances. This implies that goals will change over time, and re-learning Q-values with each goal change would be inefficient. Instead, a more flexible mechanism will dynamically assign values to various goals and then plan accordingly. The GOLSA model exemplifies this approach, essentially replacing the learned Q values of MBRL and MDPs with an activation-based representation of state value, which can be dynamically reconfigured as goals change. This overcomes the three limitations above.

Our work has several limitations. First, regarding the GOLSA model itself, the main limitation is its present implementation of one-hot state representations. This makes a scale-up to larger and continuous state spaces challenging. Future work may overcome this limitation by replacing the one-hot representations with vector-valued state representations and the learned connections with deep network function approximators. This would require corresponding changes in the search mechanisms of Fig.  1 A, from parallel, spreading activation to a serial, monte carlo tree search mechanism. This would be consistent with evidence of serial search during planning 63 , 64 and would afford a new approach to artificial general intelligence that is both powerful and similar to human brain function. Another limitation is that the Treasure Hunt task is essentially a spatial problem solving task. We anticipate that the GOLSA model could be applied to solve more general, non-spatial problems, but this remains to be demonstrated.

The fMRI analysis here has several limitations as well. First, a correspondence of representations does not imply a correspondence of computations, nor does it prove the model correct in an absolute sense 65 . There are other computational models that use diffusion gradients to solve goal-directed planning 66 , and more recent work with deep networks to navigate from arbitrary starting to arbitrary ending states 50 . The combined model and fMRI results here constitute a proposed functional account of the various brain regions, but our results do not prove that the regions compute exactly what the corresponding model regions do, nor can we definitively rule out competing models. Nevertheless the ability of the model to account for fMRI data selectively in specific brain regions suggests that it merits further investigation and direct tests against competing models, as a direction for future research. Future work might compare other models besides GOLSA against the fMRI data using RSA, to ascertain whether other model components might provide a better fit to, and account of, specific brain regions. While variations of model-based and model-free reinforcement learning models would seem likely candidates, we know of only one model by Banino et al. 50 endowed with the ability to flexibily switch goals and thus perform the treasure hunt task as does the GOLSA model 34 . It would be instructive to compare the overall abilities of GOLSA and the model of Banino et al. to account the RDMs of specific brain regions in the Treasure Hunt task, although it is unclear how to do a direct comparison given that the two models consist of very different mechanisms.

The GOLSA model may in principle be extended hierarchically. The frontal cortex has a hierarchical representational structure, in which higher levels of a task may be represented as more anterior 67 . Such hierarchical structure has been construed to represent higher, more abstract task rules 13 , 15 , 68 . The GOLSA model suggests another perspective, that higher level representations consist of higher level goals instead of higher level rules. In the coffee-making task for example 69 , the higher level task of making coffee may require a lower level task of boiling water. If the GOLSA model framework were extended hierarchically, the high level goal of having coffee prepared would activate a lower level goal of having the water heated to a specified temperature. The goal specification framework here is intrinsically more robust than a rule or schema based framework–rules may fail to produce a desired outcome, but if an error occurs in the GOLSA task performance, replanning simply calculates the optimal sequence of events from whatever the current state is, and the error will be automatically addressed.

This incidentally points to a key difference between rules and goals, in that task rules define a mapping from stimuli to responses 15 , in a way that is not necessarily teleological. Goals, in contrast, are by definition teleological. This distinction roughly parallels that between model-free and model-based reinforcement learning 70 The rule concept, as a stimulus–response mapping, implies that an error is a failure to generate the action specified by the stimulus, regardless of the final state of a system. In contrast, the goal concept implies that an error is precisely a failure to generate the desired final state of a system. Well-learned actions may acquire a degree of automaticity over time 71 , but arguably the degree of automaticity is independent of whether an action is rule oriented vs. goal-directed. If a goal-directed action becomes automatized, this does not negate the teleological nature, namely that errors in the desired final state of the world can be detected and lead to corrective action to achieve the desired final state. Rule-based action, whether deliberate or automatized, does not necessarily entail corrective action to achieve a desired state. Where actions are generated, and possibly corrected, to achieve a desired state of the world, this may properly be referred to as goal-directed behavior.

We have investigated the GOLSA model here to examine whether and how it might account for the function of specific brain regions. With RSA analysis, we found that specific layers of the GOLSA model show strong representational similarities with corresponding brain regions. Goals and goal value gradients matched especially the orbitofrontal cortex, and also some aspects of the visual and anterior temporal cortices. The desired transition layer matched representations in the hippocampus and striatum, and simulated future states matched representations in the middle frontal gyrus and superior temporal pole. Not surprisingly, the model motor layer representations matched the motor cortex. Collectively, these results constitute a proposal that the GOLSA model can provide an organizing account of how multiple brain regions interact to form essentially a negative feedback controller, with time varying behavioral set points derived from motivational states. Future work may investigate this proposal in more depth and compare against alternative models.

Model components

The GOLSA model is constructed from a small set of basic components, and the model code is freely available as supplementary material . The main component class is a layer of units, where each unit represents a neuron (or, more abstractly, a small subpopulation of neurons) corresponding to either a state, a state transition, or an action. The activity of units in a layer represents the neural firing rate and is instantiated as a vector updated according to a first order differential equation (c.f. Grossberg 72 ). The activation function varies between layers, but all units in a particular layer are governed by the same equation. The most typical activation function for a single unit is,

where a represents activation, i.e. the firing rate, of a model neuron. The four terms of this equation represent, in order: passive decay - λ a ( t ) , shunting excitation 1 - a t E , linear inhibition - I , and random noise ε N ( t ) dt . “Shunting” refers to the fact that excitation (E) scales inversely as current activity increases, with a natural upper bound of 1. The passive decay works in a similar fashion, providing a natural lower bound activity of 0. The inhibition term linearly suppresses unit activity, while the final term adds normally distributed noise N (μ = 0, σ = 1), with strength ε . Because the differential equations are approximated using the Euler method, the noise term is multiplied by dt to standardize the magnitude across different choices of dt 73 , 74 . The speed of activity change is determined by a time constant τ. The parameters τ, λ, ε vary by layer in order to implement different processes. E and I are the total excitation and inhibition, respectively, impinging on a particular unit for every presynaptic unit j in every projection p onto the target unit,

where a p j is the activation of a presynaptic model neuron that provides exciation, and w p j is the synaptic weight that determines how much excitation per unit of presynaptic activity will be provided to the postsynaptic model neuron.

A second activation function used in several places throughout the model is,

This function is identical to Eq. ( 1 ) except that the inhibition is also shunting, such that it exhibits a strong effect on highly active units and a smaller effect as unit activity approaches 0. While more typical in other models, shunting inhibition has a number of drawbacks in the current model. Two common uses for inhibition in the GOLSA model are winner-take-all dynamics and regulatory inhibition which resets layer activity. Shunting inhibition impedes both of these processes because inhibition fails to fully suppress the appropriate units, since it becomes less effective as unit activity decreases.

Projections

Layers connect to each other via projections, representing the synapses connecting one neural population to another. The primary component of projections is a weight matrix specifying the strength of connections between each pair of units. Learning is instantiated by updating the weights according to a learning function. These functions vary between the projections responsible for the model learning and are fully described in the section below dealing with each learning type. Some projections also maintain a matrix of traces updated by a projection-specific function of presynaptic or postsynaptic activity. The traces serve as a kind of short-term memory for which pre or postsynaptic units were recently activated, which serve a very similar role to eligibility traces as in Barto et al. 75 , though with a different mathematical form.

Nodes are model components that are not represented neurally via an activation function. They represent important control and timing signals to the model and are either set externally or update autonomously according to a function of time. For instance, sinusoidal oscillations are used to gate activity between various layers. While in principle rate-coded model neurons could implement a sinusoidal wave, the function is simply hard coded into the update function of the node for simplicity. In some cases, it is necessary for an entire layer to be strongly inhibited when particular conditions hold true, such as when an oscillatory node is in a particular phase. Layers therefore also have a list of inhibitor nodes that prevent unit activity within the layer when the node value meets certain conditions. In a similar fashion, some projections are gated by nodes such that they allow activity to pass through and/or allow the weights to be updated only when the relevant node activity satisfies a particular condition. Another important node provides strong inhibition to many layers when the agent changes states.

Environment

The agent operates in an environment consisting of discrete states, with a set of allowable state transitions. Allowable state transitions are not necessarily bidirectional, but for the present simulations, they are deterministic (unlike the typical MDP formulation used in RL). In some simulations, the environment also contains different types of reward located in various states, which can be used to drive goal selection. In other simulations, the goal is specified externally via a node value.

Complete network

Each component and subnetwork of the model is described in detail below or in the main text, but for reference and completeness a full diagram of the core network is shown in Fig.  5 A, and the network augmented for multi-step planning is shown in Fig.  5 B. Some of the basic layer properties are summarized in Table ​ Table2. 2 . Layers and nodes are referred to using italics, such that the layer representing the current state is referred to simply as current-state.

Representational structure

In Fig.  5 B, the layers Goal, Goal Gradient, Next State, Adjacent States, Previous States, Simulated State, and Current State all have the same number of nodes and the same representational structure, i.e. one state per node.

The layers Desired Transition, Observed Transition, Transition Output, Queue Input, Queue Output, and Queue Store likewise have the same representational structure, which is the number of possible states squared. This allows a node in these layers to represent a transition from one specific state to another specific state.

The layers Action Input, Action Output, and Previous Action all have the same representational structure, which is one possible action per node.

Task description

The Treasure Hunt task (Fig.  2 ) was created and presented in OpenSesame, a Python-based toolbox for psychological task design 76 . In the task, participants control an agent which can move within a small environment comprised of four distinct states. The nominal setting is a farm, and the states are a field with a scarecrow, the lawn in front of the farm house, a stump with an axe, and a pasture with cows. Each is associated with a picture of the scene obtained from the internet. These states were chosen to exemplify categories previously shown to elicit a univariate response in different brain regions, namely faces, houses, tools, and animals 77 – 79 .

Over the course of the experiment, participants were told the locations of treasure chests and the keys needed to open them. By arriving at a chest with the key, participants earned points which were converted to a monetary bonus at the end of the experiment. The states were arranged in a square, where each state was accessible from the two adjacent states but not the state in the opposite corner (diagonal movement was not allowed).

Each trial began with the presentation of a text screen displaying the relevant information for the next trial, namely the locations of the participant, the key, and the chest (Fig.  2 ). Because the neural patterns elicited during the presentation were the primary target of the decoding analysis, it was important that visual information be as similar as possible across different goal configurations, to avoid potential confounds. To hold luminance as constant as possible across conditions, each line always had the same number of characters. Since, for instance, “Farm House: key” has fewer characters than “Farm House: Nothing”, filler characters were added to the shorter lines, namely Xs and Os. On some trials Xs were the filler characters on the top row and Os were the filler characters on the bottom rows. This manipulation allowed us to attempt to decode the relative position of the Xs and Os to test whether decoding could be achieved due only to character-level differences in the display. We found no evidence that our results reflect low level visual confounds such as the properties of the filler characters.

Participants were under no time constraint on the information screen and pressed a button when they were ready to continue. A delay screen then appeared consisting of four empty boxes. After a jittered interval (1-6 s, distributed exponentially), arrows appeared in the boxes. The arrows represented movement directions and the boxes corresponded to four buttons under the participants left middle finger, left index finger, right index finger, and right middle finger, from left to right. Participants pressed the button corresponding to the box with the arrow pointing in the desired direction to initiate a movement. A fixation cross then appeared for another jittered delay of 0–4 s, followed by a 2 s display of the newly reached location if their choice was correct or an error screen if it was incorrect.

If the participant did not yet have the key required to open the chest, the correct movement was always to the key. Sometimes the key and chest were in the same location in which case the participant would earn points immediately. If they were in different locations, then on the next trial the participant had to move to the chest. This structure facilitated a mix of goal distances (one and two states away) while controlling the route required to navigate to the goal.

If the chosen direction was incorrect, participants saw an error screen displaying text and a map of the environment. Participants advanced from this screen with a button press and then restarted the failed trial. If the failed trial was the second step in a two-step sequence (i.e., if they had already gotten the key and then moved to the wrong state to get to the chest), they had to repeat the previous two trials.

Repeating the failed trial ensured that there were balanced numbers of each class of event for decoding, since an incorrect response indicated that some information was not properly maintained or utilized. For example, if a participant failed the second step of a two-trial sequence, then they may not have properly encoded the final goal when first presented with the information screen on the previous trial, which specified the location of the key and the chest.

Halfway through the experiment, the map was reconfigured such that states were swapped across the diagonal axes of the map. This was necessary because otherwise, each state could be reached by exactly two movement directions and exactly two movement directions could be made from it. For instance, if the farm house was the state in the lower left, the farmhouse could only be reached by moving left or down from adjacent states, and participants starting at the farm house could only move up or to the right. If this were true across the entire experiment, above-chance classification of target state, for instance, could appear in regions that in fact only contain information about the intended movement direction.

Each state was the starting state for one quarter of the trials and the target destination for a different quarter of the trials. All trials were one of three types. One category consisted of single-trial (single) sequences in which the chest and key were in the same location. The sequences in which the chest and key were in separate locations required two trials to complete, one to move from the initial starting location to the key and another to move from the key location to the chest location. These two steps formed the other two classes of trials, the first-of-two (first) and second-of-two (second) trials. Recall that on second trials, no information other than the participant’s current location is presented on the starting screen to ensure that the participant maintained the location of the chest in memory across the entire two-trial sequence (if it was presented on the second trial, there would be no need to maintain that information through the first trial). The trials were evenly divided into single, first, and second classes with 64 trials in each class. Therefore, every trial had a starting state and an immediate goal, while one third of trials also had a more distant final goal.

Immediately prior to participating in the fMRI version of the task, participants completed a short 16-trial practice outside the scanner to refresh their memory. Before beginning the first run inside the scanner, participants saw a map of the farm states and indicated when they had memorized it before moving on. Within each run, participants completed as many trials as they could within eight minutes. As described above, exactly halfway through the trials, the state space was rearranged with each state moving to the opposite corner. Therefore, when participants completed the first half of the experiment, the current run was terminated and participants were given time to learn the new state space before scanning resumed. At the end of the experiment, participants filled out a short survey about their strategy.

Participants

In total, 49 participants (28 female) completed the behavioral-only portion of the experiment, including during task piloting (early versions of the behavioral task were slightly different than described below). Participants provided written informed consent in accordance with the Institutional Review Board at Indiana University, and were compensated $10/hour for their time plus a performance bonus based on accuracy up to an additional $10. The behavioral task first served as a pilot during task design and then as a pre-screen for the fMRI portion, in that only participants with at least 90% accuracy were invited to participate. Additional criteria for scanning were that the subjects be right handed, free of metal implants, free of claustrophobia, weigh less than 440 pounds, and not be currently taking psychoactive medication. In total, 25 participants participated in the fMRI task but one subject withdrew shortly after beginning, leaving 24 subjects who completed the imaging task (14 female). Across the 24 subjects, the average error rate of responses during the fMRI task was 2.4%, and error trials were modeled separately in the fMRI analysis. These were not analyzed further as there were too few error trials for a meaningful analysis.

fMRI acquisition and data preprocessing

Imaging data were collected on a Siemens Magnetom Trio 3.0-Tesla MRI scanner and a 32 channel head coil. Foam padding was inserted around the sides of the head to increase participant comfort and reduce head motion. Functional T2* weighted images were acquired using a multiband EPI sequence 80 with 42 contiguous slices and 3.44 × 3.44 × 3.4 mm 3 voxels (echo time = 28 ms; flip angle = 60; field of view = 220, multiband acceleration factor = 3). For the first subject, the TR was 813 ms, but during data collection for the second subject the TR changed to 816 ms for unknown reasons. The scanner was upgraded after collecting data from an additional five subjects, at which point the TR remained constant at 832 ms. All other parameters remained unchanged. High-resolution T 1 –weighted MPRAGE images were collected for spatial normalization (256 × 256 × 160 matrix of 1 × 1 × 1mm 3 voxels, TR = 1800, echo time = 2.56 ms; flip angle = 9).

Functional data were spike-corrected using AFNI’s 3dDespike ( http://afni.nimh.nih.gov/afni ). Functional images were corrected for difference in slice timing using sinc-interpolation and head movement using a least-squares approach with a 6-parameter rigid body spatial transformation. For subjects who moved more than 3 mm total or 0.5 mm between TRs, 24 motion regressors were added to subsequent GLM analyses 81 .

Because MVPA and representation similarity analysis (RSA) rely on precise voxelwise patterns, these analyses were performed before spatial normalization. For the univariate analyses, structural data were coregistered to the functional data and segmented into gray and white matter probability maps 82 . These segmented images were used to calculate spatial normalization parameters to the MNI template, which were subsequently applied to the functional data. As part of spatial normalization, the data were resampled to 2 × 2 × 2mm 3 , and this upsampling allowed maximum preservation of information. All analyses included a temporal high-pass filter (128 s) and correction for temporal autocorrelation using an autoregressive AR(1) model.

Univariate GLM

For initial univariate analyses, we measured the neural response associated with each outcome state at the outcome screen (when an image of the state was displayed), as well as the signal at the start of the trial associated with each immediate goal location. Five timepoints were modeled in the GLM used in this analysis, namely the start of the trial, the button press to advance, the appearance of the arrows and subsequent response, the start of the feedback, and the end of the feedback. The regressors marking the start of the trial and the start of the feedback screen were further individuated by the immediate goal on the trial. A separate error regressor was used when the response was incorrect, meaning they did not properly pursue the immediate goal and received error feedback. All correct trials in which participants moved to, for instance, the cow field, used the same trial start and feedback start regressors.

The GLM was fit to the normalized functional images. The resulting beta maps were combined at the second level with a voxel-wise threshold of p  < 0.001 and cluster corrected ( p  < 0.05) to control for multiple comparisons. We assessed the univariate response associated with each outcome location, by contrasting each particular outcome location with all other outcome locations. The response to the error feedback screen was assessed in a separate contrast against all correct outcomes. To test for any univariate responses related to the immediate goal, we performed an analogous analysis using the trial start regressors which were individuated based on the immediate goal. For example, the regressor ‘trialStartHouseNext’ was associated with the beginning of every trial where the farmhouse was the immediate goal location. To assess the univariate signal associated with the farmhouse immediate goal, we performed a contrast between this regressor and all other trial start regressors.

Representational similarity analysis (RSA)

As before, a GLM was fit to the realigned functional images. The following events were modeled with impulse regressors: trial onset (information screen), key press to advance to the decision screen, the prompt and immediately subsequent action (modeled as a single regressor), the onset of the outcome screen, and the termination of the outcome screen. The RSA analysis used beta maps derived from the regressors marking trial onset, prompt/response, and outcome screen onset.

Each of these regressors (except those used in error trials) were further individuated by the (start state, next state, final goal) triple constituting the goal path. There were 8 distinct trial types starting in each state. Each state could serve as the starting point of two single-step sequences (in which the key and treasure chest are in the same location) and four two-step sequences (in which the key and treasure chest are in different locations). Each state could also be the midpoint of a two-step sequence with the treasure chest located in one of two adjacent states. With three regressors used for each trial, there were 4 starting states * 8 trial types * 3 time points = 96 total patterns used to create the Representational Dissimilarity Matrix (RDM) in each searchlight region, where cell x ij in the RDM is defined as one minus the Pearson correlation between the ith and jth patterns. Values close to 2 therefore represent negative correlation (high representational distance) while values close to 0 indicate a positive correlation (low representational distance).

To derive the model-based RDMs, the GOLSA model was run on an analogue of the goal pursuit task, using a four state state-space with four actions corresponding to movement in each cardinal direction. The model layer timecourses of activity are shown in Figs.  6 and ​ and7 7 for one- and two-step trials, respectively. The base GOLSA model is not capable of maintaining a plan across an arbitrary delay, but instead acts immediately to make the necessary state transitions. The competitive queue 83 module allows state transition sequences to be maintained and executed after a delay, and was therefore necessary to model the task in the most accurate manner possible. However, the goal-learning module was not necessary since goals were externally imposed. Because participants had to demonstrate high performance on the task before entering the scanner, little if any learning took place during the experiment. As a result, the model was trained extensively on the state space before performing any trials used in data collection. To further simulate likely patterns of activity in the absence of significant learning, the input from state to goal-gradient (used in the learning phase of an oscillatory cycle) was removed and the goal-gradient received steady input from the goal layer, interrupted only by the state-change inhibition signal. In other words, the goal-gradient layer continuously represented the actual goal gradient rather than shifting into learning mode half of the time.

Model activity during a simulated one-step sequence of the Treasure Hunt task. The competitive queueing module first loads a plan and then executes it sequentially. State activity shows that the agent remains in state 1 for the first half of the simulation, while simulated-state (StateSim) shows the state transition the agent simulates as it forms its plan. Adjacent-states (Adjacent) receives input from stateSim which, along with goal-gradient (Gradient) activity determines the desired next state and therefore the appropriate transition to make. The plan is kept in queue-store (Store) which receives a burst of input from queue-input (QueueIn) and finally executes the plan by sending output to queue-output (QueueOut) which drives the motor system. The vertical dashed lines indicating the different phases of the simulation used in the creation of the model RDMs. For each layer, activity within each period was averaged across time to form a single vector representing the average pattern for that time period in the trial type being simulated. The bounds of each phase were determined qualitatively. The planning period is longer than the acting and outcome periods because the model takes longer to form a plan than execute it or observe the outcome.

Model activity during a simulated two-step sequence of the Treasure Hunt task. The competitive queueing module first loads a plan and then executes it sequentially. State activity shows that the agent remains in state 1 for the first half of the simulation, while simulated-state shows the state transitions the agent simulates as it forms its plan. Adjacent-states receives input from simulated-state which, along with goal-gradient activity determines the desired next state and therefore the appropriate transitions to make. The plan is kept in queue-store which receives bursts of input from queue-input and finally executes the plan by sequentially sending output to queue-output which drives the motor system. To force the agent to go to the appropriate intermediate state, goal activity first reflects the key location and then the chest location. The vertical dashed lines indicate time periods used when creating the RDMs for the two-step sequence simulations. The first three time periods correspond to the first trial in the sequence while the latter three correspond to the second trial in the sequence. Again, the first planning period is much longer due to the nature of the model dynamics. During the second “planning” period (P2), the plan was already formed as must have been the case in the actual experiment since on the second trials in a two-step sequence, no information was presented at the start of the trial and had to be remembered from the previous trial.

In the task, participants first saw an information screen from which they could determine the immediate goal state and the appropriate next action. This plan was maintained over a delay before being implemented. At the beginning of each trial simulation, the queuing module was set to “load” while the model interactions determined the best method of getting from the current state to the goal state. This period is analogous to the period in which subjects look at the starting information screen and plan their next move. Then, the queuing module was set to “execute,” modeling the period in which participants are prompted to make their selection. Finally, the chosen action implements a state transition and the environment provides new state information to the state layer, modeling the outcome phase of the experiment.

Some pairs of trials in the task comprised a two-step sequence in which the final goal was initially two states away from the starting state. On the second trial of such sequences, participants were not provided any information on the information screen at the start of the trial, ensuring that they had encoded and maintained all goal-related information from the information screen presented at the start at the first trial in the sequence. These pairs of trials were modeled within a single GOLSA simulation. The model seeks the quickest path to the goal, identifying immediately available subgoals as needed. However, in the task, the location of the key necessitated a specific path to reach the final goal of the treasure chest. To provide these instructions to the model at the start of a two-step simulation, the goal representation from the subgoal (the key) was provided to the model first until the appropriate action was loaded and then the goal representation shifted to the final goal (the chest). Once the full two-step state transition sequence was loaded in the queue, the actions were read out sequentially, as shown in Fig.  7 .

A separate RDM was generated for each model layer. Patterns were extracted from three time intervals per action (six total for the two-step sequence simulations). Due of the time required to load the queue, the first planning period was longer than all other intervals. For each simulation and time point, the patterns of activity across the units were averaged over time, yielding one vector. Each trial type was repeated 10 times and the patterns generated in the previous step were averaged across simulation repetitions. The activity of each layer was thus summarized with at most 96 patterns of activity which were converted into an RDM by taking one minus the Pearson correlation between each pattern. Patterns in which all units were 0 were ignored since the correlation is undefined for constant vectors.

We looked for neural regions corresponding to the layers that played a critical role in the model during the acting phase in the typical learning oscillation since in these simulations the learning phase of the oscillation was disabled. We created RDMs from the following layers: current-state, adjacent-states, goal, goal-gradient, next-desired-state, desired-transition, action-out, simulated-state, and queue-store. As a control, we also added a layer component which generated normally distributed noise (μ = 1, σ = 1).

RSA searchlight

The searchlight analysis was conducted using Representational Similarity Analysis Toolbox, developed at the University of Cambridge ( http://www.mrc-cbu.cam.ac.uk/methods-and-resources/toolboxes/license/ ). For each of these layer RDM, a searchlight of radius of 10 mm was moved through the entire brain. At each voxel, an RDM was created by from the patterns in the spherical region centered on that voxel.

An r value was obtained for each voxel by computing the Spearman correlation between the searchlight RDM and the model layer RDM, ignoring trial time periods in which all model units showed no activity. A full pass of the searchlight over the brain produced a whole-brain r map for each subject for each layer. Voxels in regions that perform a similar function to the model component will produce similar RDMs to the model component RDM and thus will be assigned relatively high values. The r maps were then Fisher-transformed into z maps ( z = 1 2 ln 1 + r 1 - r ). The z maps were normalized into the MNI template but were not smoothed, as the searchlight method already introduces substantial smoothing. Second level effects were assessed with a t test on the normalized z maps, with a cluster defining threshold of p  < 0.001, cluster corrected to p  < 0.05 overall. The cluster significance was determined by SPM5 and verified for clusters >  = 24 voxels in size with a version of 3DClustSim (compile date Jan. 11, 2017) that corrects for the alpha inflation found in pervious 3DClustSim versions 84 . The complete results are shown in Table ​ Table1 1 .

Supplementary Information

Acknowledgements.

We thank A. Ramamoorthy for helpful discussions and J. Fine and W. Alexander for helpful comments on the manuscript. Supported by the Indiana University Imaging Research Facility. JWB was supported by NIH R21 DA040773.

Author contributions

J.W.B. and N.Z. designed the model and experiment. N.Z. implemented and simulated the model, implemented and ran the fMRI experiment, and analyzed the data. J.W.B. and N.Z. wrote the paper.

Data availability

Competing interests.

The authors declare no competing interests.

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

The online version contains supplementary material available at 10.1038/s41598-023-28834-3.

• Health, Fitness & Dieting
• Psychology & Counseling

Enjoy fast, free delivery, exclusive deals, and award-winning movies & TV shows with Prime Try Prime and start saving today with fast, free delivery

Amazon Prime includes:

Fast, FREE Delivery is available to Prime members. To join, select "Try Amazon Prime and start saving today with Fast, FREE Delivery" below the Add to Cart button.

• Cardmembers earn 5% Back at Amazon.com with a Prime Credit Card.
• Unlimited Free Two-Day Delivery
• Streaming of thousands of movies and TV shows with limited ads on Prime Video.
• A Kindle book to borrow for free each month - with no due dates
• Listen to over 2 million songs and hundreds of playlists
• Unlimited photo storage with anywhere access

Important:  Your credit card will NOT be charged when you start your free trial or if you cancel during the trial period. If you're happy with Amazon Prime, do nothing. At the end of the free trial, your membership will automatically upgrade to a monthly membership.

Buy new: $59.95 $59.95 FREE delivery: May 1 - 3 Ships from: Amazon.com Sold by: Amazon.com

Return this item for free.

Free returns are available for the shipping address you chose. You can return the item for any reason in new and unused condition: no shipping charges

• Go to your orders and start the return
• Select the return method

Buy used: $54.74

Other sellers on amazon.

Download the free Kindle app and start reading Kindle books instantly on your smartphone, tablet, or computer - no Kindle device required .

Read instantly on your browser with Kindle for Web.

Using your mobile phone camera - scan the code below and download the Kindle app.

Image Unavailable

• To view this video download Flash Player

Follow the authors

Human Problem Solving Hardcover – February 5, 2019

Purchase options and add-ons.

First published in 1972, this monumental work develops and defends the authors' information processing theory of human reasoning. Human reasoners, they argue, can be modeled as symbolic "information processing systems" (IPSs), abstracted entirely from physiological bases. Modeling subjects with IPSs yields predictive theories of their problem-solving behavior and performance, and psychological insight into their heuristics and methods. Newell and Simon's previous epoch-making collaborations included the General Problem Solver, the Logic Theorist, and the Information Processing Language. This book is a careful application of those ideas from artificial intelligence - the ideas of AI's first golden age - to cognitive psychology. The authors first develop the formal theory of information processing systems. They then report studies of three symbolic reasoning tasks, and analyze that data using the information processing paradigm. In the final section, they state their comprehensive theory of human problem-solving. The success of the models of cognition given in this work was a major piece of evidence for the physical symbol system hypothesis, which Newell and Simon would first state a few years later. Newell went on to co-develop the Soar cognitive architecture, and Simon to receive the Nobel Prize in Economics. The two jointly received the Turing Award in 1975 for the research program of which Human Problem Solving was the culmination.

• Print length 938 pages
• Language English
• Publisher Echo Point Books & Media
• Publication date February 5, 2019
• Dimensions 6.69 x 1.94 x 9.61 inches
• ISBN-10 1635617928
• ISBN-13 978-1635617924
• See all details

Frequently bought together

Customers who viewed this item also viewed

Editorial Reviews

"...perhaps the most important book on the scientific study of human thinking in the 20th century." - E.A. Feigenbaum, A. M. Turing Award Laureate

Product details

• Publisher ‏ : ‎ Echo Point Books & Media; Reprint ed. edition (February 5, 2019)
• Language ‏ : ‎ English
• Hardcover ‏ : ‎ 938 pages
• ISBN-10 ‏ : ‎ 1635617928
• ISBN-13 ‏ : ‎ 978-1635617924
• Item Weight ‏ : ‎ 3.69 pounds
• Dimensions ‏ : ‎ 6.69 x 1.94 x 9.61 inches
• #2,780 in Medical Cognitive Psychology
• #4,077 in Popular Social Psychology & Interactions
• #4,280 in Cognitive Psychology (Books)

About the authors

Allen newell.

Discover more of the author’s books, see similar authors, read author blogs and more

Herbert A. Simon

Customer reviews.

Customer Reviews, including Product Star Ratings help customers to learn more about the product and decide whether it is the right product for them.

To calculate the overall star rating and percentage breakdown by star, we don’t use a simple average. Instead, our system considers things like how recent a review is and if the reviewer bought the item on Amazon. It also analyzed reviews to verify trustworthiness.

• Sort reviews by Top reviews Most recent Top reviews

Top reviews from the United States

There was a problem filtering reviews right now. please try again later..

Top reviews from other countries

• Amazon Newsletter
• About Amazon
• Accessibility
• Sustainability
• Press Center
• Investor Relations
• Amazon Devices
• Amazon Science
• Sell on Amazon
• Sell apps on Amazon
• Supply to Amazon
• Protect & Build Your Brand
• Become an Affiliate
• Become a Delivery Driver
• Start a Package Delivery Business
• Advertise Your Products
• Self-Publish with Us
• Become an Amazon Hub Partner
• › See More Ways to Make Money
• Amazon Visa
• Amazon Store Card
• Amazon Secured Card
• Amazon Business Card
• Shop with Points
• Credit Card Marketplace
• Reload Your Balance
• Amazon Currency Converter
• Your Account
• Your Orders
• Shipping Rates & Policies
• Amazon Prime
• Returns & Replacements
• Manage Your Content and Devices
• Recalls and Product Safety Alerts
• Conditions of Use
• Privacy Notice
• Consumer Health Data Privacy Disclosure
• Your Ads Privacy Choices

Herbert A. Simon and the Science of Decision Making

Herbert A. Simon (1916-2001)

On June 15 , 1916 , American political scientist , economist , sociologist , psychologist , and computer scientist Herbert Alexander Simon was born. Simon was among the founding fathers of several of today’s important scientific domains, including artificial intelligence , information processing, decision-making, problem-solving, organization theory , complex systems , and computer simulation of scientific discovery . With almost a thousand highly cited publications , he was one of the most influential social scientists of the 20th century.

“(If) there were no limits to human rationality administrative theory would be barren. It would consist of the single precept: Always select that alternative, among those available, which will lead to the most complete achievement of your goals”, – Herbert A. Simon, Administrative Behavior, 1947.

Herbert A. Simon – Early Years

Herbert Alexander Simon was born in Milwaukee, Wisconsin to Arthur Simon, an electrical engineer who had come to the United States from Germany. His mother, Edna Marguerite Merkel, was an accomplished pianist. Simon was educated as a child in the public school system in Milwaukee where he developed an interest in science. Through his uncle’s books on economics and psychology, Simon discovered the social sciences. In 1933, Simon entered the University of Chicago, and studied the social sciences and mathematics. Originally, Simon was interested in biology, but chose not to study it because of his “color-blindness and awkwardness in the laboratory”. Simon received both his B.A. (1936) and his Ph.D. (1943) in political science, from the University of Chicago, where he studied under Harold Lasswell , Nicholas Rashevsky , Rudolf Carnap ,[ 7 ]  Henry Schultz , and Charles Edward Merriam .

Academic Career

After enrolling in a course on “Measuring Municipal Governments,” Simon was invited to be a research assistant for Clarence Ridley , with whom he coauthored the book, Measuring Municipal Activities , in 1938. After graduating with his undergraduate degree, Simon obtained a research assistantship in municipal administration which turned into a directorship at the University of California, Berkeley. From 1942 to 1949, Simon was a professor of political science and also served as department chairman at Illinois Institute of Technology. In 1949, he became a professor of administration and psychology at the Carnegie Institute of Technology (now Carnegie Mellon University), later becoming the Richard King Mellon University Professor of Computer Science and Psychology there. He began a more in-depth study of economics in the area of institutionalism there.

Corporate Decision Making

“The criterion of efficiency dictates that choice of alternatives which produces the largest result for the given application of resources.” – H. Simon (1945), as quoted in [17] 

A Pioneer of Artificial Intelligence

“In an information-rich world, the wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes. What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.” Simon, H. A. (1971) “Designing Organizations for an Information-Rich World” 

Herbert Simon has made a great number of profound and in depth contributions to both economic analysis and applications. Simon also was a pioneer in the field of artificial intelligence, creating with Allen Newell the Logic Theory Machine (1956) and the General Problem Solver (GPS) (1957) programs. Both programs were developed using the Information Processing Language (IPL) (1956) developed by Newell, Cliff Shaw, and Simon. In 1957, Simon predicted that computer chess would surpass human chess abilities within “ten years” when, in reality, that transition took about forty years.

Simulating Human Problem Solving

In the early 1960s psychologist Ulric Neisser asserted that while machines are capable of replicating ‘cold cognition’ behaviors such as reasoning, planning, perceiving, and deciding, they would never be able to replicate ‘hot cognition’ behaviors such as pain, pleasure, desire, and other emotions. Simon responded to Neisser’s views in 1963 by writing a paper on emotional cognition, which was largely ignored by the artificial intelligence research community, but subsequent work on emotions by Sloman and Picard helped refocus attention on Simon’s paper and eventually, made it highly influential on the topic. With Allen Newell, Simon developed a theory for the simulation of human problem solving behavior using production rules The study of human problem solving required new kinds of human measurements and, with Anders Ericsson , Simon developed the experimental technique of verbal protocol analysis. Simon was interested in the role of knowledge in expertise. He said that to become an expert on a topic required about ten years of experience and he and colleagues estimated that expertise was the result of learning roughly 50,000 chunks of information. A chess expert was said to have learned about 50,000 chunks or chess position patterns. In 1975 Herbert A. Simon was awarded the ACM A.M. Turing Award along with Allen Newell.

Simon’s three stages in Rational Decision Making: Intelligence, Design, Choice (IDC), MrunaltPatel, CC BY 3.0 <https://creativecommons.org/licenses/by/3.0>, via Wikimedia Commons

Organizational Decision Making and Nobel Prize

Simon was interested in how humans learn and, with Edward Feigenbaum , he developed the EPAM (Elementary Perceiver and Memorizer) theory, one of the first theories of learning to be implemented as a computer program. Simon also has been credited for revolutionary changes in microeconomics, where he introduced the concept of organizational decision-making as it is known today. He was the first to discuss this concept in terms of uncertainty, in the sense that it is impossible to have perfect and complete information at any given time to make a decision. It was in this contribution that he was awarded the Nobel Prize in 1978.

New Institutionalist Economics

In January 2001, he underwent surgery at UPMC Presbyterian to remove a cancerous tumor in his abdomen. Although the surgery was successful, Simon later succumbed to the complications that followed on February 9, 2001. 

References and Further Reading:

• [1] Herbert A. Simon , American Social Scientist, at Britannica Online.
• [2] A tribute to Herbert A. Simon , at CMU
• [3] D. Klahr, K. Kotovsky: A Life of the Mind: Remembering Herb Simon , American Psychological Society, 2001.
• [4] Herbert A. Simon , at The New World Encyclopedia
• [5] Herbert A. Simon, “ Literary Criticism: A Cognitive Approach ” from Stanford Humanities Review, 1995, with peer reviews and critique.
• [6] Herbert A. Simon at Wikidata
• [7]  Rudolf Carnap and the Logical Structure of the World , SciHi Blog
• [8]  Herbert A. Simon,  A Theory of Emotional Behavior . Carnegie Mellon University Complex Information Processing (CIP) Working Paper #55, June 1, 1963.
• [9]  Herbert Alexander Simon   at the   Mathematics Genealogy Project
• [10]  Herbert Alexander Simon   at the AI Genealogy Project.
• [11]  “Herbert A. Simon – Biographical” .  nobelprize.org .  
• [12]  Herbert A. Simon,   A Theory of Emotional Behavior . Carnegie Mellon University Complex Information Processing (CIP) Working Paper #55, June 1, 1963.
• [13]  Herbert A. Simon,  “Motivational and Emotional Controls of Cognition” .   Psychological Review , 1967, Vol. 74, No. 1, 29-39.
• [14] Simon, Herbert A.   ‘Organizations and markets’ ,   Journal of Economic Perspectives , vol. 5, no. 2 (1991), pp. 25–44.
• [15] Frantz, R., and Marsh, L. (Eds.) (2016).   Minds, Models and Milieux: Commemorating the Centennial of the Birth of Herbert Simon . Palgrave Macmillan.
• [16]  Herbert Simon : September 9, 1979, Current Research ,  at Carnegie Mellon University,  cmurobotics  @ youtube
• [17] Harry M. Johnson (1966) Sociology: A Systematic Introduction
• [18] Timeline of Nobel Laureates in Economics , via Wikidata

Harald Sack

Related posts, norman borlaug and the green revolution, sidney fox and his research for the origins of life, frederick william twort and the bacteriophages, edward c. kendall and the adrenal cortex hormones, leave a reply cancel reply.

Your email address will not be published. Required fields are marked *

Further Projects

• February (28)
• January (30)
• December (30)
• November (29)
• October (31)
• September (30)
• August (30)
• January (31)
• December (31)
• November (30)
• August (31)
• February (29)
• February (19)
• January (18)
• October (29)
• September (29)
• February (5)
• January (5)
• December (14)
• November (9)
• October (13)
• September (6)
• August (13)
• December (3)
• November (5)
• October (1)
• September (3)
• November (2)
• September (2)
• Entries RSS
• Comments RSS
• WordPress.org

Legal Notice

• Privacy Statement

• SUGGESTED TOPICS
• The Magazine
• Newsletters
• Managing Yourself
• Managing Teams
• Work-life Balance
• The Big Idea
• Data & Visuals
• Reading Lists
• Case Selections
• HBR Learning
• Topic Feeds
• Account Settings
• Email Preferences

Share Podcast

Do You Understand the Problem You’re Trying to Solve?

To solve tough problems at work, first ask these questions.

• Apple Podcasts
• Google Podcasts

Problem solving skills are invaluable in any job. But all too often, we jump to find solutions to a problem without taking time to really understand the dilemma we face, according to Thomas Wedell-Wedellsborg , an expert in innovation and the author of the book, What’s Your Problem?: To Solve Your Toughest Problems, Change the Problems You Solve .

In this episode, you’ll learn how to reframe tough problems by asking questions that reveal all the factors and assumptions that contribute to the situation. You’ll also learn why searching for just one root cause can be misleading.

Key episode topics include: leadership, decision making and problem solving, power and influence, business management.

HBR On Leadership curates the best case studies and conversations with the world’s top business and management experts, to help you unlock the best in those around you. New episodes every week.

• Listen to the original HBR IdeaCast episode: The Secret to Better Problem Solving (2016)
• Find more episodes of HBR IdeaCast
• Discover 100 years of Harvard Business Review articles, case studies, podcasts, and more at HBR.org .

HANNAH BATES: Welcome to HBR on Leadership , case studies and conversations with the world’s top business and management experts, hand-selected to help you unlock the best in those around you.

Problem solving skills are invaluable in any job. But even the most experienced among us can fall into the trap of solving the wrong problem.

Thomas Wedell-Wedellsborg says that all too often, we jump to find solutions to a problem – without taking time to really understand what we’re facing.

He’s an expert in innovation, and he’s the author of the book, What’s Your Problem?: To Solve Your Toughest Problems, Change the Problems You Solve .

  In this episode, you’ll learn how to reframe tough problems, by asking questions that reveal all the factors and assumptions that contribute to the situation. You’ll also learn why searching for one root cause can be misleading. And you’ll learn how to use experimentation and rapid prototyping as problem-solving tools.

This episode originally aired on HBR IdeaCast in December 2016. Here it is.

SARAH GREEN CARMICHAEL: Welcome to the HBR IdeaCast from Harvard Business Review. I’m Sarah Green Carmichael.

Problem solving is popular. People put it on their resumes. Managers believe they excel at it. Companies count it as a key proficiency. We solve customers’ problems.

The problem is we often solve the wrong problems. Albert Einstein and Peter Drucker alike have discussed the difficulty of effective diagnosis. There are great frameworks for getting teams to attack true problems, but they’re often hard to do daily and on the fly. That’s where our guest comes in.

Thomas Wedell-Wedellsborg is a consultant who helps companies and managers reframe their problems so they can come up with an effective solution faster. He asks the question “Are You Solving The Right Problems?” in the January-February 2017 issue of Harvard Business Review. Thomas, thank you so much for coming on the HBR IdeaCast .

THOMAS WEDELL-WEDELLSBORG: Thanks for inviting me.

SARAH GREEN CARMICHAEL: So, I thought maybe we could start by talking about the problem of talking about problem reframing. What is that exactly?

THOMAS WEDELL-WEDELLSBORG: Basically, when people face a problem, they tend to jump into solution mode to rapidly, and very often that means that they don’t really understand, necessarily, the problem they’re trying to solve. And so, reframing is really a– at heart, it’s a method that helps you avoid that by taking a second to go in and ask two questions, basically saying, first of all, wait. What is the problem we’re trying to solve? And then crucially asking, is there a different way to think about what the problem actually is?

SARAH GREEN CARMICHAEL: So, I feel like so often when this comes up in meetings, you know, someone says that, and maybe they throw out the Einstein quote about you spend an hour of problem solving, you spend 55 minutes to find the problem. And then everyone else in the room kind of gets irritated. So, maybe just give us an example of maybe how this would work in practice in a way that would not, sort of, set people’s teeth on edge, like oh, here Sarah goes again, reframing the whole problem instead of just solving it.

THOMAS WEDELL-WEDELLSBORG: I mean, you’re bringing up something that’s, I think is crucial, which is to create legitimacy for the method. So, one of the reasons why I put out the article is to give people a tool to say actually, this thing is still important, and we need to do it. But I think the really critical thing in order to make this work in a meeting is actually to learn how to do it fast, because if you have the idea that you need to spend 30 minutes in a meeting delving deeply into the problem, I mean, that’s going to be uphill for most problems. So, the critical thing here is really to try to make it a practice you can implement very, very rapidly.

There’s an example that I would suggest memorizing. This is the example that I use to explain very rapidly what it is. And it’s basically, I call it the slow elevator problem. You imagine that you are the owner of an office building, and that your tenants are complaining that the elevator’s slow.

Now, if you take that problem framing for granted, you’re going to start thinking creatively around how do we make the elevator faster. Do we install a new motor? Do we have to buy a new lift somewhere?

The thing is, though, if you ask people who actually work with facilities management, well, they’re going to have a different solution for you, which is put up a mirror next to the elevator. That’s what happens is, of course, that people go oh, I’m busy. I’m busy. I’m– oh, a mirror. Oh, that’s beautiful.

And then they forget time. What’s interesting about that example is that the idea with a mirror is actually a solution to a different problem than the one you first proposed. And so, the whole idea here is once you get good at using reframing, you can quickly identify other aspects of the problem that might be much better to try to solve than the original one you found. It’s not necessarily that the first one is wrong. It’s just that there might be better problems out there to attack that we can, means we can do things much faster, cheaper, or better.

SARAH GREEN CARMICHAEL: So, in that example, I can understand how A, it’s probably expensive to make the elevator faster, so it’s much cheaper just to put up a mirror. And B, maybe the real problem people are actually feeling, even though they’re not articulating it right, is like, I hate waiting for the elevator. But if you let them sort of fix their hair or check their teeth, they’re suddenly distracted and don’t notice.

But if you have, this is sort of a pedestrian example, but say you have a roommate or a spouse who doesn’t clean up the kitchen. Facing that problem and not having your elegant solution already there to highlight the contrast between the perceived problem and the real problem, how would you take a problem like that and attack it using this method so that you can see what some of the other options might be?

THOMAS WEDELL-WEDELLSBORG: Right. So, I mean, let’s say it’s you who have that problem. I would go in and say, first of all, what would you say the problem is? Like, if you were to describe your view of the problem, what would that be?

SARAH GREEN CARMICHAEL: I hate cleaning the kitchen, and I want someone else to clean it up.

THOMAS WEDELL-WEDELLSBORG: OK. So, my first observation, you know, that somebody else might not necessarily be your spouse. So, already there, there’s an inbuilt assumption in your question around oh, it has to be my husband who does the cleaning. So, it might actually be worth, already there to say, is that really the only problem you have? That you hate cleaning the kitchen, and you want to avoid it? Or might there be something around, as well, getting a better relationship in terms of how you solve problems in general or establishing a better way to handle small problems when dealing with your spouse?

SARAH GREEN CARMICHAEL: Or maybe, now that I’m thinking that, maybe the problem is that you just can’t find the stuff in the kitchen when you need to find it.

THOMAS WEDELL-WEDELLSBORG: Right, and so that’s an example of a reframing, that actually why is it a problem that the kitchen is not clean? Is it only because you hate the act of cleaning, or does it actually mean that it just takes you a lot longer and gets a lot messier to actually use the kitchen, which is a different problem. The way you describe this problem now, is there anything that’s missing from that description?

SARAH GREEN CARMICHAEL: That is a really good question.

THOMAS WEDELL-WEDELLSBORG: Other, basically asking other factors that we are not talking about right now, and I say those because people tend to, when given a problem, they tend to delve deeper into the detail. What often is missing is actually an element outside of the initial description of the problem that might be really relevant to what’s going on. Like, why does the kitchen get messy in the first place? Is it something about the way you use it or your cooking habits? Is it because the neighbor’s kids, kind of, use it all the time?

There might, very often, there might be issues that you’re not really thinking about when you first describe the problem that actually has a big effect on it.

SARAH GREEN CARMICHAEL: I think at this point it would be helpful to maybe get another business example, and I’m wondering if you could tell us the story of the dog adoption problem.

THOMAS WEDELL-WEDELLSBORG: Yeah. This is a big problem in the US. If you work in the shelter industry, basically because dogs are so popular, more than 3 million dogs every year enter a shelter, and currently only about half of those actually find a new home and get adopted. And so, this is a problem that has persisted. It’s been, like, a structural problem for decades in this space. In the last three years, where people found new ways to address it.

So a woman called Lori Weise who runs a rescue organization in South LA, and she actually went in and challenged the very idea of what we were trying to do. She said, no, no. The problem we’re trying to solve is not about how to get more people to adopt dogs. It is about keeping the dogs with their first family so they never enter the shelter system in the first place.

In 2013, she started what’s called a Shelter Intervention Program that basically works like this. If a family comes and wants to hand over their dog, these are called owner surrenders. It’s about 30% of all dogs that come into a shelter. All they would do is go up and ask, if you could, would you like to keep your animal? And if they said yes, they would try to fix whatever helped them fix the problem, but that made them turn over this.

And sometimes that might be that they moved into a new building. The landlord required a deposit, and they simply didn’t have the money to put down a deposit. Or the dog might need a $10 rabies shot, but they didn’t know how to get access to a vet.

And so, by instigating that program, just in the first year, she took her, basically the amount of dollars they spent per animal they helped went from something like $85 down to around $60. Just an immediate impact, and her program now is being rolled out, is being supported by the ASPCA, which is one of the big animal welfare stations, and it’s being rolled out to various other places.

And I think what really struck me with that example was this was not dependent on having the internet. This was not, oh, we needed to have everybody mobile before we could come up with this. This, conceivably, we could have done 20 years ago. Only, it only happened when somebody, like in this case Lori, went in and actually rethought what the problem they were trying to solve was in the first place.

SARAH GREEN CARMICHAEL: So, what I also think is so interesting about that example is that when you talk about it, it doesn’t sound like the kind of thing that would have been thought of through other kinds of problem solving methods. There wasn’t necessarily an After Action Review or a 5 Whys exercise or a Six Sigma type intervention. I don’t want to throw those other methods under the bus, but how can you get such powerful results with such a very simple way of thinking about something?

THOMAS WEDELL-WEDELLSBORG: That was something that struck me as well. This, in a way, reframing and the idea of the problem diagnosis is important is something we’ve known for a long, long time. And we’ve actually have built some tools to help out. If you worked with us professionally, you are familiar with, like, Six Sigma, TRIZ, and so on. You mentioned 5 Whys. A root cause analysis is another one that a lot of people are familiar with.

Those are our good tools, and they’re definitely better than nothing. But what I notice when I work with the companies applying those was those tools tend to make you dig deeper into the first understanding of the problem we have. If it’s the elevator example, people start asking, well, is that the cable strength, or is the capacity of the elevator? That they kind of get caught by the details.

That, in a way, is a bad way to work on problems because it really assumes that there’s like a, you can almost hear it, a root cause. That you have to dig down and find the one true problem, and everything else was just symptoms. That’s a bad way to think about problems because problems tend to be multicausal.

There tend to be lots of causes or levers you can potentially press to address a problem. And if you think there’s only one, if that’s the right problem, that’s actually a dangerous way. And so I think that’s why, that this is a method I’ve worked with over the last five years, trying to basically refine how to make people better at this, and the key tends to be this thing about shifting out and saying, is there a totally different way of thinking about the problem versus getting too caught up in the mechanistic details of what happens.

SARAH GREEN CARMICHAEL: What about experimentation? Because that’s another method that’s become really popular with the rise of Lean Startup and lots of other innovation methodologies. Why wouldn’t it have worked to, say, experiment with many different types of fixing the dog adoption problem, and then just pick the one that works the best?

THOMAS WEDELL-WEDELLSBORG: You could say in the dog space, that’s what’s been going on. I mean, there is, in this industry and a lot of, it’s largely volunteer driven. People have experimented, and they found different ways of trying to cope. And that has definitely made the problem better. So, I wouldn’t say that experimentation is bad, quite the contrary. Rapid prototyping, quickly putting something out into the world and learning from it, that’s a fantastic way to learn more and to move forward.

My point is, though, that I feel we’ve come to rely too much on that. There’s like, if you look at the start up space, the wisdom is now just to put something quickly into the market, and then if it doesn’t work, pivot and just do more stuff. What reframing really is, I think of it as the cognitive counterpoint to prototyping. So, this is really a way of seeing very quickly, like not just working on the solution, but also working on our understanding of the problem and trying to see is there a different way to think about that.

If you only stick with experimentation, again, you tend to sometimes stay too much in the same space trying minute variations of something instead of taking a step back and saying, wait a minute. What is this telling us about what the real issue is?

SARAH GREEN CARMICHAEL: So, to go back to something that we touched on earlier, when we were talking about the completely hypothetical example of a spouse who does not clean the kitchen–

THOMAS WEDELL-WEDELLSBORG: Completely, completely hypothetical.

SARAH GREEN CARMICHAEL: Yes. For the record, my husband is a great kitchen cleaner.

You started asking me some questions that I could see immediately were helping me rethink that problem. Is that kind of the key, just having a checklist of questions to ask yourself? How do you really start to put this into practice?

THOMAS WEDELL-WEDELLSBORG: I think there are two steps in that. The first one is just to make yourself better at the method. Yes, you should kind of work with a checklist. In the article, I kind of outlined seven practices that you can use to do this.

But importantly, I would say you have to consider that as, basically, a set of training wheels. I think there’s a big, big danger in getting caught in a checklist. This is something I work with.

My co-author Paddy Miller, it’s one of his insights. That if you start giving people a checklist for things like this, they start following it. And that’s actually a problem, because what you really want them to do is start challenging their thinking.

So the way to handle this is to get some practice using it. Do use the checklist initially, but then try to step away from it and try to see if you can organically make– it’s almost a habit of mind. When you run into a colleague in the hallway and she has a problem and you have five minutes, like, delving in and just starting asking some of those questions and using your intuition to say, wait, how is she talking about this problem? And is there a question or two I can ask her about the problem that can help her rethink it?

SARAH GREEN CARMICHAEL: Well, that is also just a very different approach, because I think in that situation, most of us can’t go 30 seconds without jumping in and offering solutions.

THOMAS WEDELL-WEDELLSBORG: Very true. The drive toward solutions is very strong. And to be clear, I mean, there’s nothing wrong with that if the solutions work. So, many problems are just solved by oh, you know, oh, here’s the way to do that. Great.

But this is really a powerful method for those problems where either it’s something we’ve been banging our heads against tons of times without making progress, or when you need to come up with a really creative solution. When you’re facing a competitor with a much bigger budget, and you know, if you solve the same problem later, you’re not going to win. So, that basic idea of taking that approach to problems can often help you move forward in a different way than just like, oh, I have a solution.

I would say there’s also, there’s some interesting psychological stuff going on, right? Where you may have tried this, but if somebody tries to serve up a solution to a problem I have, I’m often resistant towards them. Kind if like, no, no, no, no, no, no. That solution is not going to work in my world. Whereas if you get them to discuss and analyze what the problem really is, you might actually dig something up.

Let’s go back to the kitchen example. One powerful question is just to say, what’s your own part in creating this problem? It’s very often, like, people, they describe problems as if it’s something that’s inflicted upon them from the external world, and they are innocent bystanders in that.

SARAH GREEN CARMICHAEL: Right, or crazy customers with unreasonable demands.

THOMAS WEDELL-WEDELLSBORG: Exactly, right. I don’t think I’ve ever met an agency or consultancy that didn’t, like, gossip about their customers. Oh, my god, they’re horrible. That, you know, classic thing, why don’t they want to take more risk? Well, risk is bad.

It’s their business that’s on the line, not the consultancy’s, right? So, absolutely, that’s one of the things when you step into a different mindset and kind of, wait. Oh yeah, maybe I actually am part of creating this problem in a sense, as well. That tends to open some new doors for you to move forward, in a way, with stuff that you may have been struggling with for years.

SARAH GREEN CARMICHAEL: So, we’ve surfaced a couple of questions that are useful. I’m curious to know, what are some of the other questions that you find yourself asking in these situations, given that you have made this sort of mental habit that you do? What are the questions that people seem to find really useful?

THOMAS WEDELL-WEDELLSBORG: One easy one is just to ask if there are any positive exceptions to the problem. So, was there day where your kitchen was actually spotlessly clean? And then asking, what was different about that day? Like, what happened there that didn’t happen the other days? That can very often point people towards a factor that they hadn’t considered previously.

SARAH GREEN CARMICHAEL: We got take-out.

THOMAS WEDELL-WEDELLSBORG: S,o that is your solution. Take-out from [INAUDIBLE]. That might have other problems.

Another good question, and this is a little bit more high level. It’s actually more making an observation about labeling how that person thinks about the problem. And what I mean with that is, we have problem categories in our head. So, if I say, let’s say that you describe a problem to me and say, well, we have a really great product and are, it’s much better than our previous product, but people aren’t buying it. I think we need to put more marketing dollars into this.

Now you can go in and say, that’s interesting. This sounds like you’re thinking of this as a communications problem. Is there a different way of thinking about that? Because you can almost tell how, when the second you say communications, there are some ideas about how do you solve a communications problem. Typically with more communication.

And what you might do is go in and suggest, well, have you considered that it might be, say, an incentive problem? Are there incentives on behalf of the purchasing manager at your clients that are obstructing you? Might there be incentive issues with your own sales force that makes them want to sell the old product instead of the new one?

So literally, just identifying what type of problem does this person think about, and is there different potential way of thinking about it? Might it be an emotional problem, a timing problem, an expectations management problem? Thinking about what label of what type of problem that person is kind of thinking as it of.

SARAH GREEN CARMICHAEL: That’s really interesting, too, because I think so many of us get requests for advice that we’re really not qualified to give. So, maybe the next time that happens, instead of muddying my way through, I will just ask some of those questions that we talked about instead.

THOMAS WEDELL-WEDELLSBORG: That sounds like a good idea.

SARAH GREEN CARMICHAEL: So, Thomas, this has really helped me reframe the way I think about a couple of problems in my own life, and I’m just wondering. I know you do this professionally, but is there a problem in your life that thinking this way has helped you solve?

THOMAS WEDELL-WEDELLSBORG: I’ve, of course, I’ve been swallowing my own medicine on this, too, and I think I have, well, maybe two different examples, and in one case somebody else did the reframing for me. But in one case, when I was younger, I often kind of struggled a little bit. I mean, this is my teenage years, kind of hanging out with my parents. I thought they were pretty annoying people. That’s not really fair, because they’re quite wonderful, but that’s what life is when you’re a teenager.

And one of the things that struck me, suddenly, and this was kind of the positive exception was, there was actually an evening where we really had a good time, and there wasn’t a conflict. And the core thing was, I wasn’t just seeing them in their old house where I grew up. It was, actually, we were at a restaurant. And it suddenly struck me that so much of the sometimes, kind of, a little bit, you love them but they’re annoying kind of dynamic, is tied to the place, is tied to the setting you are in.

And of course, if– you know, I live abroad now, if I visit my parents and I stay in my old bedroom, you know, my mother comes in and wants to wake me up in the morning. Stuff like that, right? And it just struck me so, so clearly that it’s– when I change this setting, if I go out and have dinner with them at a different place, that the dynamic, just that dynamic disappears.

SARAH GREEN CARMICHAEL: Well, Thomas, this has been really, really helpful. Thank you for talking with me today.

THOMAS WEDELL-WEDELLSBORG: Thank you, Sarah.  

HANNAH BATES: That was Thomas Wedell-Wedellsborg in conversation with Sarah Green Carmichael on the HBR IdeaCast. He’s an expert in problem solving and innovation, and he’s the author of the book, What’s Your Problem?: To Solve Your Toughest Problems, Change the Problems You Solve .

We’ll be back next Wednesday with another hand-picked conversation about leadership from the Harvard Business Review. If you found this episode helpful, share it with your friends and colleagues, and follow our show on Apple Podcasts, Spotify, or wherever you get your podcasts. While you’re there, be sure to leave us a review.

We’re a production of Harvard Business Review. If you want more podcasts, articles, case studies, books, and videos like this, find it all at HBR dot org.

This episode was produced by Anne Saini, and me, Hannah Bates. Ian Fox is our editor. Music by Coma Media. Special thanks to Maureen Hoch, Adi Ignatius, Karen Player, Ramsey Khabbaz, Nicole Smith, Anne Bartholomew, and you – our listener.

See you next week.

• Subscribe On:

Latest in this series

This article is about leadership.

• Decision making and problem solving
• Power and influence
• Business management

Partner Center

Collaborative Robotics is prioritizing ‘human problem solving’ over humanoid forms

Humanoids have sucked a lot of the air out of the room. It is, after all, a lot easier to generate press for robots that look and move like humans. Ultimately, however, both the efficacy and scalability of such designs have yet to be proven out. For a while now, Collaborative Robotics founder Brad Porter has eschewed robots that look like people. Machines that can potentially reason like people, however, is another thing entirely.

As the two-year-old startup’s name implies, Collaborative Robotics (Cobot for short) is interested in the ways in which humans and robots will collaborate, moving forward. The company has yet to unveil its system, though last year, Porter told me that the “novel cobot” system is neither humanoid nor a mobile manipulator mounted to the back of an autonomous mobile robot (AMR).

The system has, however, begun to be deployed in select sites.

“Getting our first robots in the field earlier this year, coupled with today’s investment, are major milestones as we bring cobots with human-level capability into the industries of today,” Porter says. “We see a virtuous cycle where more robots in the field lead to improved AI and a more cost-effective supply chain.”

Further deployment will be helped along by a fresh $100 million Series B, led by General Catalyst and featuring Bison Ventures, Industry Ventures and Lux Capital. That brings the Bay Area firm’s total funding up to $140 million. General Catalyst’s Teresa Carlson is also joining the company in an advisory role.

Cobot has the pedigree, as well, with staff that includes former Apple, Meta, Google, Microsoft, NASA and Waymo employees. Porter himself spent more than 13 years at Amazon. When his run with the company ended in summer 2020, he was leading the retail giant’s industrial robotics team.

Amazon became one of the world’s top drivers and consumer of industrial robotics during that time, and the company’s now ubiquitous AMRs stand as a testament to the efficiency of pairing human and robot workers together.

AI will, naturally, be foundational to the company’s promise of “human problem solving,” while the move away from the humanoid form factor is a bid, in part, to reduce the cost of entry for deploying these systems.

Study: Monkeys are much smarter than we thought they were

I n a groundbreaking study published today in the journal Nature Neuroscience , researchers have discovered that monkeys, much like humans , are capable of complex deliberation and careful decision-making. 

This new finding challenges the long-held belief that humans alone possess the ability to think deeply about a problem and consider multiple factors such as costs, consequences, and constraints in order to arrive at optimal outcomes.

"Humans are not the only animals capable of slow and thoughtful deliberation," said study senior author Dr. William Stauffer from the University of Pittsburgh School of Medicine. "Our work shows that monkeys have a rich mental state that renders them capable of intelligent thinking. It's a new paradigm for studying the neurophysiological basis for deliberative thought."

The study raises important questions about the nature of thought processes and decision-making in animals, and whether other species are also capable of engaging in the same level of complexity as humans. It also helps to shed light on the cognitive processes at work when we, as humans, make decisions about various aspects of our lives, such as who to spend time with or what to study in school.

Several decades ago, Dr. Daniel Kahneman, a Nobel Prize laureate, revolutionized the field of behavioral economics with his Prospect Theory. In his seminal book, "Thinking Fast and Slow," Dr. Kahneman posited that humans employ two distinct systems of thinking: one nearly instantaneous and automatic, and the other much slower and reliant on conscious logical reasoning that requires greater mental effort.

How the study was done

Dr. Kahneman referred to the first type of thinking as "slow" and the second as "fast." Slow, effortful thinking enables us to engage in complex activities such as writing music, developing scientific hypotheses, and balancing our checkbooks. Until now, it was believed that slow thinking was a uniquely human trait.

However, this latest research turns that notion on its head. By presenting monkeys with combinatorial optimization problems, which the researchers dubbed the "knapsack task," and rewarding the animals based on the value of their solutions, the study demonstrated that monkeys employed sophisticated mathematical reasoning and used efficient computational algorithms to tackle complex problems.

The scientists found that the animals' performance and speed of deliberation were dependent on the task's complexity, and that their solutions closely mirrored those generated by efficient computer algorithms specifically designed to solve the optimization problem.

"Results from this work will contribute neurophysiological evidence to enlighten centuries of discussions about dual process theories of the mind, the structure of thoughts, and the neurobiological basis of intuition and reasoning," wrote Stauffer in an accompanying research briefing.

Tao Hong of Carnegie Mellon University is the lead author of the paper. The study's findings not only provide valuable insights into the cognitive abilities of monkeys but also pave the way for a new paradigm in studying the neurophysiological basis for deliberative thought, with potential implications for better understanding the complex nature of decision-making across various species.

More about monkeys

Monkeys are a diverse group of primates that belong to the infraorder Simiiformes. They are divided into two major groups: New World monkeys, native to Central and South America, and Old World monkeys, native to Africa and Asia. Monkeys are known for their intelligence, social behavior, and adaptability to different environments.

Physical characteristics

Monkeys vary greatly in size and appearance, ranging from the tiny pygmy marmoset, which measures just 4.6-6.2 inches (12-16 cm) in length, to the large mandrill, which can reach up to 37 inches (94 cm) in length. 

Monkeys typically have forward-facing eyes, flat faces, and dexterous hands with opposable thumbs. Some species also have prehensile tails, which they use to grasp and manipulate objects or to hang from branches.

Most monkeys are omnivores, eating a diverse diet that includes fruits, leaves, seeds, insects, and small animals. Some species, like the howler monkey, primarily consume leaves, while others, like the capuchin monkey, have a more varied diet.

Social behavior

Monkeys are highly social animals that usually live in groups called troops. These troops can range in size from just a few individuals to hundreds of members. Social hierarchies are common in monkey troops, with dominant individuals enjoying benefits like better access to food and mating opportunities. Monkeys communicate through vocalizations, body language, and facial expressions, and they often engage in grooming behaviors to maintain social bonds.

Intelligence and tool use

Monkeys are known for their cognitive abilities, problem-solving skills, and in some cases, their use of tools. Capuchin monkeys, for example, have been observed using rocks to crack open nuts, while some macaques have been seen using sticks to extract insects from tree bark. 

Research has also shown that monkeys are capable of understanding basic arithmetic and recognizing themselves in mirrors, which is considered a sign of self-awareness.

Conservation

Many monkey species are threatened by habitat loss, hunting, and the illegal pet trade. Conservation efforts are underway to protect these primates and their habitats, including the establishment of protected areas, reintroduction programs, and education campaigns to raise awareness about the importance of monkey conservation.

In conclusion, monkeys are fascinating and intelligent creatures with complex social structures and diverse behaviors. As we continue to study these primates, we gain a greater understanding of their cognitive abilities and the evolutionary links between humans and other primates.

Other animals that demonstrate problem-solving ability

Yes, numerous animals demonstrate problem-solving abilities, indicating the presence of intelligence and cognitive skills across various species. Some examples of animals with notable problem-solving capabilities include:

Crows and other corvids

These birds are known for their exceptional problem-solving skills and have been observed using tools to access food. For instance, they can use sticks to extract insects from tree bark or crevices and even bend wires to create hooks for retrieving food from hard-to-reach places.

Elephants are highly intelligent animals capable of complex problem-solving. They have been observed using sticks and branches to swat flies or scratch hard-to-reach areas and can also recognize themselves in mirrors, suggesting self-awareness. Elephants have displayed the ability to cooperate and work together to solve problems, such as pulling a rope simultaneously to access food.

Dolphins are known for their intelligence and problem-solving abilities. They have been observed using tools like sponges to protect their snouts while foraging on the ocean floor. Dolphins can also learn and understand complex commands and have been shown to recognize themselves in mirrors, indicating self-awareness.

These highly intelligent invertebrates have demonstrated remarkable problem-solving skills. Octopuses have been observed opening jars, navigating mazes, and escaping from enclosures by manipulating objects and their environment. Their impressive learning and memory capabilities make them formidable problem solvers.

Domesticated dogs have evolved alongside humans and have developed a range of problem-solving skills. They can learn commands, understand gestures, and follow human cues to solve problems, such as locating hidden objects or navigating obstacles. Some breeds, like border collies and poodles, are especially known for their intelligence and problem-solving abilities.

Chimpanzees

As our closest living relatives, chimpanzees share many cognitive traits with humans. They have been observed using tools, such as sticks to extract termites from their mounds, and leaves as sponges to collect water. Chimpanzees also display complex social behaviors, such as cooperation and deception, which require problem-solving skills.

Rats are intelligent rodents that have shown the ability to solve problems and learn from their experiences. They can navigate complex mazes, recognize patterns, and demonstrate a rudimentary understanding of cause and effect. Rats have also been observed using tools and adapting their behavior based on previous experiences.

These examples illustrate that problem-solving abilities are not exclusive to humans and can be found across various animal species. Studying these animals and their cognitive skills can provide valuable insights into the evolution of intelligence and the diversity of problem-solving strategies in the animal kingdom.

Check us out on EarthSnap , a free app brought to you by Eric Ralls and Earth.com .

Crossing the principle–practice gap in AI ethics with ethical problem-solving

• Original Research
• Open access
• Published: 15 April 2024

Cite this article

You have full access to this open access article

• Nicholas Kluge Corrêa   ORCID: orcid.org/0000-0002-5633-6094 1 , 4 ,
• James William Santos   ORCID: orcid.org/0000-0002-9806-1172 4 , 5 ,
• Camila Galvão   ORCID: orcid.org/0000-0002-4814-4164 2 , 4 ,
• Marcelo Pasetti   ORCID: orcid.org/0000-0003-1993-1422 2 , 4 ,
• Dieine Schiavon   ORCID: orcid.org/0000-0001-8090-2386 2 , 4 ,
• Faizah Naqvi   ORCID: orcid.org/0009-0003-1225-0824 3 ,
• Robayet Hossain   ORCID: orcid.org/0000-0003-0511-3285 3 &
• Nythamar De Oliveira   ORCID: orcid.org/0000-0001-9241-1031 2 , 4  

41 Accesses

Explore all metrics

The past years have presented a surge in (AI) development, fueled by breakthroughs in deep learning, increased computational power, and substantial investments in the field. Given the generative capabilities of more recent AI systems, the era of large-scale AI models has transformed various domains that intersect our daily lives. However, this progress raises concerns about the balance between technological advancement, ethical considerations, safety measures, and financial interests. Moreover, using such systems in sensitive areas amplifies our general ethical awareness, prompting a re-emergence of debates on governance, regulation, and human values. However, amidst this landscape, how to bridge the principle–practice gap separating ethical discourse from the technical side of AI development remains an open problem. In response to this challenge, the present work proposes a framework to help shorten this gap: ethical problem-solving (EPS). EPS is a methodology promoting responsible, human-centric, and value-oriented AI development. The framework’s core resides in translating principles into practical implementations using impact assessment surveys and a differential recommendation methodology. We utilize EPS as a blueprint to propose the implementation of an Ethics as a Service Platform , currently available as a simple demonstration. We released all framework components openly and with a permissive license, hoping the community would adopt and extend our efforts into other contexts. Available in the following URL https://nkluge-correa.github.io/ethical-problem-solving/ .

Avoid common mistakes on your manuscript.

1 Introduction

The late 2010s, especially after the beginning of the deep learning revolution [ 1 , 2 ], marked a rising interest in AI research, underlined by an exponential increase in the academic work related to the field [ 3 , 4 , 5 ]. The confluence of technical advancements (i.e., breakthroughs in deep learning and increased computational power), data availability (i.e., the proliferation of data through social media, smartphones, and IoT devices), and massive investments (i.e., governments and private companies began investing heavily in AI R &D) played a crucial part in the expansion of the field [ 6 , 7 , 8 ]. In the past few years (2022–2023), AI has entered an era where organizations release large-scale foundation models every few months [ 4 ], with systems like GPT-4 [ 9 ], Llama 2 [ 10 ], Gemini [ 11 ], Whisper [ 12 ], CLIP [ 13 ], among many others, becoming the basis of many modern AI applications. The capabilities of such models, ranging from human-like text generation and analysis to image synthesis and unprecedented speech recognition, have revolutionized the public consciousness of AI and transformed how we interact with technology. Footnote 1

This accelerated progress comes with its own set of challenges. Notably, academia released the majority of state-of-the-art machine learning models until 2014. Since then, the market-driven industry has taken over [ 3 , 4 ]. Big tech companies are the current major players in the research and development of AI applications. This shift leads us to question the balance between ethical considerations, safety measures, technological progress, and revenue shares. In other words, prioritization of revenue and “progress in the name of progress” may undercut ethical and safety concerns. Meanwhile, the use of AI systems in sensitive areas, such as healthcare [ 15 , 16 ], banking services [ 17 ], public safety [ 18 , 19 ], among others [ 20 , 21 , 22 ] prompted a re-emergence of the debate surrounding the ethical issues related to the use, development, and governance of these systems and technologies in general [ 23 , 24 , 25 , 26 ].

The AI safety field of research emerges as one possible solution to this unprecedented expansion. AI safety focuses on approaches for developing AI beneficial to humanity and alerts us to the unforeseen consequences of AI systems untied to human values [ 27 , 28 , 29 , 30 ]. At the same time, the growth of the research field in terms of regulation and governance demonstrates an apparent broad consensus on the human values (principles) relevant to AI development and the necessity of enforceable rules and guidelines to apply these values [ 3 , 31 , 32 , 33 , 34 ]. Under this scenario, events like the “Pause Giant AI Experiments open letter” [ 35 ], among others [ 36 , 37 ], are a symptom that even the industry recognizes that the unrestrained AI development impacts will not be positive or manageable [ 30 ]. With that realization, the outcry for regulatory input in the AI development industry grew stronger [ 38 , 39 , 40 ]. Ultimately, the debate on whether consensus on ethical principles exists, how to translate principles to practices, and the interests underlying the push for creating normative guidelines presents critical unanswered questions around AI research [ 41 ].

This work seeks to tackle a critical point within these aforementioned unanswered questions. The principle–practice gap (i.e., translating principles to practice [ 42 ]) presents the sociotechnical challenge of managing the expectations of those who seek a magical normative formula and working with developers to foster design ingrained with ethical principles. With this in mind, we present ethical problem-solving (EPS), a method to aid developers and other professionals responsible for creating autonomous systems (Fig.  1 ).

EPS seeks to bridge the principle-practice gap between principles and practical implementations, giving developers tools to use in the development cycle of an AI system. The general workflow of this method consists of an evaluation (Impact Assessment) and a Recommendation stage, structured in a WHY, SHOULD, and HOW format

Our framework’s core resides in translating abstract principles into practical implementations through impact assessment and a differential recommendation methodology. The EPS builds and improves upon similar works, where proposals are usually limited to paper or worksheet frameworks, while also presenting novel contributions: an implementation of EPS tailored to Brazil’s context. In the following sections, we will explore the challenges faced in ethical AI development, the path toward the EPS, its process, and resources. We released all components of our current implementation openly and with a permissive license, hoping the community can adopt and extend our efforts into other contexts. Footnote 2

2 Related works

This section provides an overview of practical frameworks for AI ethics, specifically emphasizing approaches and methods that help translate ethical principles into developmental insight.

According to Taddeo and Floridi [ 42 ], creating methods to anticipate ethical risks and opportunities and prevent unwanted consequences is the crucial aspect of what the authors call translational ethics. One significant issue for translational ethics is the definition of the appropriate set of fundamental ethical principles to guide the creation, regulation, and application of AI to benefit and respect individuals and societies. Another is the formulation of foresight methodologies to indicate ethical risks and opportunities and prevent unwanted consequences (e.g., impact assessments, stakeholder engagements, and governance frameworks). In the context of AI ethics, translational ethics aims towards concrete and applicable practices that realize broadly accepted ethical principles. In other words, the translation of theoretical conclusions into adequate practice [ 43 ]. However, the authors do not present a particular foresight methodology in the translational framework proposed by Taddeo and Floridi. Nevertheless, they have an optimistic view towards a global convergence of ethical principles that could bring regulatory advancements, as shown in other fields of applied ethics.

Another approach that aims to assess the ethical aspects of AI systems is the VCIO-based (Values, Criteria, Indicators, and Observables) description of systems for AI trustworthiness characterization [ 44 ]. First introduced by the AI Ethics Impact Group in the “From Principles to Practice—An Interdisciplinary Framework to Operationalize AI Ethics” [ 45 ], the VCIO approach identifies core values, defines specific criteria, establishes measurable indicators, and utilizes observable evidence to measure the alignment of AI systems according to the selected values (i.e., transparency, accountability, privacy, justice, reliability, and environmental sustainability), which steam from a meta-analysis of relevant publications. The VCIO approach culminates in an AI Ethics Label. The label offers a rating for each value assessment proposed by the VCIO approach and later communicates an AI system’s ethical score. However, no prerequisites exist to build a label based on the VCIO score. Companies, users, or government bodies can set the requirements for a minimum level of risk within this framework. Moreover, the following publication tied to the VCIO approach [ 44 ] clearly states that the VCIO’s description/evaluation is independent of the risk posed by the technology under consideration and does not define any minimum requirements. It merely describes compliance with the specified values.

Even though VCIO is a valid contribution to the field, its limitations are also apparent in its description. First, the source of the values stated in the papers and the parameters of the meta-analysis that originated them are unclear. Despite being well-known principles in the AI ethics literature [ 31 , 32 , 33 , 34 ], we argue that not knowing how they are defined or uncovered represents a weak spot in the underlying methodology of the VCIO framework. Second, the self-imposing character of the risk evaluation requirements can be an issue and potentially undermine the effect of the approach. If there is no objective measure of the potential risks and impacts on ethical principles, the whole project of applying principles to practice may become moot.

Another approach that instrumentalizes ethical principles into palpable development tools is the Google People AI Research Guidebook [ 46 ], developed by PAIR’s (People + AI Research) multidisciplinary team. The guidebook aims to empower researchers and developers to create AI systems that are fair, inclusive, and unbiased. The guidebook provides six chapters and their respective worksheets that acknowledge the potential risks associated with AI development and emphasize the importance of addressing bias, fairness, interpretability, privacy, and security. The guidebook encourages researchers to adopt a multidisciplinary approach incorporating insights from diverse fields. The worksheets provide strategies for understanding and mitigating bias, creating interpretable AI models, implementing privacy-preserving techniques, and promoting responsible data-handling practices.

PAIR’s guidebook represented a commendable step towards building AI systems that are ethically sound, unbiased, and beneficial to all. Nevertheless, the guidebook and respective worksheets do not present a standard to evaluate whether there is progress in the development. Users of the methodology are left free to gauge their successes and failures, making the whole approach tied to the expectations of the user, who is also the evaluator. Also, the worksheets themselves do not present an approachable user interface.

Other examples of methods created to deal with the same issues previously mentioned works sought to tackle include:

The Digital Catapult AI Ethics Framework [ 47 ] is a set of seven principles and corresponding questions to ensure AI technologies’ responsible and ethical development. The framework advises the developer to consider the principles and their questions. The Digital Catapult underlines that the objective of the questions is to highlight the various scenarios in which ethical principles may apply to the project. Nevertheless, just like in the VCIO methodology [ 44 , 45 ], their method needs to clarify the constitution of the set of relevant values. Meanwhile, the voluntary aspects of the approach rely on the awareness developers must have regarding AI’s potential risks and ethical blank spots.

Microsoft’s AI Ethical Framework [ 48 ] is a set of principles and guidelines designed to ensure AI’s responsible and ethical use. The framework encompasses six principles that should guide AI development and use: fairness, privacy and security, transparency, trust and security, inclusion, and accountability. Each (group of) principle(s) sets its goals and practical measures, bringing back the idea of pragmatizing ethical principles. By adhering to these principles, Microsoft aims to contribute to positive societal impact while mitigating potential risks associated with AI technologies. While the source of the ethical principles remains undisclosed in the paper (a recurring theme in the literature related to this work), the tools meant to achieve said principles are underrepresented, leaving much of the heavy lifting to the developers themselves.

Morley et al. [ 49 ] investigation also revolves around the gap between AI ethics principles, practical implementations, and evaluation of the existing translational tools and methods. The authors propose an assistance method for AI development akin to a Platform as a Service (PaaS). PaaS represents a set-up where core infrastructure, such as operating systems, computational hardware, and storage, is provided to enable developers to create custom software or applications expeditiously. Meanwhile, Ethics as a Service (EaaS) seeks to contextually build an ethical approach with shared responsibilities, where the EaaS provides the infrastructure needed for moral development. In their work, Morley et al. mentions several components that should be a part of this kind of platform, for example, an independent multi-disciplinary ethics board, a collaboratively developed ethical code, and the AI practitioners themselves, among others. EaaS as an idea has the advantage of being relatable to modern tech companies, where organizations subsidize much of their work to specialized third parties. Also, having an ethical evaluation performed by a third party has merits. However, the actual operation of the EaaS framework and the content of the service provided are still ongoing research for the authors.

Baker-Brunnbauer [ 50 ] proposes the Trustworthy Artificial Intelligence Implementation (TAII) Framework, a management guideline for implementing trusted AI systems in enterprises. The framework contains several steps, starting with an overview of the system to address the company’s values and business model. Then, the focus shifts to the definition and documentation of the stakeholders and the exact regulations and standards in play. The assessment of risks and compliance with the common good follows. Ultimately, the framework should generate ethical principles suitable for translation into practical implementations, while executing and certifying the results should be the last step. However, two vulnerabilities of the TAII approach are its self-imposing nature and the many steps involved in the framework. Given these flaws, the iteration of the many measures proposed could leave evaluators blind to the weaknesses of their creation.

As our last mention, we cite the framework created by Ciobanu and Mesnita [ 51 ] for implementing ethical AI in the industry. The proposed framework comprises AI Embedded Ethics by Design (AI EED) and AI Desired State Configuration (AI DSC). The AI EED stage is where the developer can train its model to address the specific ethical challenges of a particular AI application. Meanwhile, the developer or the consumer can define the relevant AI principles using the VCIO approach as the normative source. The AI DSC stage focuses on actively managing the AI system post-implementation through constant user feedback. However, it is unclear how the framework operationalizes its stages, except for the VCIO approach, which also has shortcomings. Also, the framework relies heavily on the interest and acute awareness of the general public to provide feedback on an end-to-end process where the AI system is adapted to the contextual reality where it is implemented.

As shown above, we can see many efforts to incorporate ethics in AI development and deployment. Many of these attempts are still extra-empirical and subject to improvement, modification, and actual deployment. Also, these attempts further justify our introduction’s main point: the imperative necessity to anticipate and mitigate ethical risks in AI system development.

To build upon the work mentioned above, in the following sections, we propose a framework (EPS) that, we argue, is ethically and theoretically grounded, simple, practical, and not self-imposing.

3 Methodology

From the gaps in previous proposals and grounded in our critical analyses of the field of AI ethics, the EPS presents:

A set of assessment surveys aimed at helping evaluators estimate the possible impacts of a given AI application.

An evaluation matrix to classify an AI application’s impact level according to AI principles grounded in empirical research.

A recommendation approach customized to each impact level. These recommendations provide practical suggestions for enhancing the ethical standards of the evaluated system or application.

Before diving into the implementational aspects of our method, let us first revise the philosophical roots of ethical problem-solving.

3.1 Theoretical grounding

The EPS is grounded on Critical Theory’s social diagnosis methods and emancipatory goals, particularly Rahel Jaeggi’s “Critique of Forms of Life”. [ 52 ]. That said, this work’s approach to normativity is not focused on the more traditional aspects of normative theory (i.e., how to judge a particular action as moral and which parameters to use) but is aligned with the notion that normativity is created and reproduced through social practices as an ongoing process. Hence, Jaeggi’s approach fits with normativity as an integral part of sociality, which cannot be dissociated from it to judge what is right or good. This argument is in line with her approach to immanent critique, and it resonates with a few authors who methodologically sustain immanent criticism as the path to avoid universalistic or relativistic tendencies [ 53 , 54 , 55 , 56 ]. However, what differentiates Jaeggi’s approach from others is her conceptualization of forms of life, her normative recuperation of Hegelian theory, her problem-solving approach, and the open practical questions it leaves us with.

Jaeggi draws on the concept of “forms of Life” introduced by Ludwig Wittgenstein [ 57 ], which broadly refers to how individuals and communities organize their activities and ways of understanding the world. Jaeggi reformulates this concept to analyze the social and cultural structures that shape human existence, examining how they might limit human flourishing, autonomy, and self-realization. Jaeggi understands forms of life as clusters of social practices addressing historically contextualized and normatively defined problems. Ultimately, what shapes human existence is how exactly these problems are solved within a form of life.

Broadly speaking, the practices that constitute a form of life are connected in practical-functional ways. Some have a tangible sense of how functionally interdependent they are, such as agricultural practices required for urban consumption. In contrast, others do not, such as playing with children as an essential part of parenthood. In this sense, practices bring their interpretation as something (descriptive) and the functional assignment as being good for something (evaluative) in correlation with each other. Jaeggi thus understands forms of life as being geared to solve problems because their description already carries a meshed functional and ethical perspective. In other words, forms of life always entail an inherent evaluation, excluding pure functionality in human activities. The critical theory of technology, also known as Science and Technology Studies (STS), sustains a similar position that technology is value-laden like other social realities that frame our everyday existence [ 58 , 59 ]. However, the argument of technology being permeated by values or biases in its inception only partially resonates with the proposition of this work. It is Jaeggi’s proposal of the value-laden argument setting up a normative foundation and a problem-solving approach that fits the present endeavor.

Jaeggi asserts that the normative dimensions of forms of life are not static but rather rooted in a triangular relationship involving the current empirical state (is), normative claims (ought), and changing objective conditions. The normative claims reflect the expectation of particular manifestations of social practices as they developed historically, and the current empirical state reflects the actual state of social practices against the expectations and facing objective conditions. This continuous process of dynamic normativity can be illustrated by the development of the rural-feudal extended family into the bourgeois nuclear family due to changed socioeconomic (objective) conditions demanding changes in normative expectations. Further developments, ranging from patchwork families to polyamorous relationships, could also be understood as reactions to the new objective conditions now posed in turn, but not solved, by the bourgeois family [ 52 ].

To address the unsolved issues of social practices and ever-changing objective conditions, Jaeggi proposes problem-solving in the form of a hermeneutic anticipation (i.e., a recommendation) of an assumed solution (i.e., of a desirable goal) and the validity of such a recommendation can only be determined after addressing the identified problems. As we can see, Jaeggi’s approach foreshadows a dynamic normativity that renews itself through iteration without necessarily implying progress from the outset, considering that the success or failure of the recommendation determines the evaluation of the addressed problem. Footnote 3 Jaeggi’s proposal also raised criticisms regarding her conceptualization of forms of life, which was deemed exceedingly vague [ 70 ]. Also, despite Jaeggi’s push to include some notion of social reality in her theoretical proposition, it seems that Jaeggi needed to go further as the approach lacks a clear connection with sociality [ 71 , 72 ]. Although Jaeggi’s work does open practical avenues, much of the work toward changing practices is left out of her proposal.

At this point comes the inspiration for the EPS; at first glance, it might seem striking that there is a slight difference in scope between Jaeggi’s theoretically proposed criticism for societies and the EPS, which seeks to bridge values toward actionable practices of ethical AI development. It is worth underlining that the EPS is not attempting an overarching criticism of technology. However, it can still take advantage of a problem-solving approach that tackles a complex conceptualization and responsively grasps normativity. The EPS puts into practice the problem-solving process theoretically proposed on a different scale but with a complex subject matter and ethical stakes nonetheless. There is no doubt that AI systems involve ensembles of directives to solve diverse issues and that much of the inner workings can be elusive to our current understanding of the subject, not unlike the elusive character of our social practices and the historical background that supports them.

Nevertheless, despite the elusive character of the subject, there is an apparent demand for normalization or, at least, to understand how suitable norms should arise. To this point, the EPS utilizes the dynamic normative approach to test Jaeggi’s theoretical proposal further. The EPS gauges the empirical state of systems (through survey assessment) to trace if their normative assignments are aligned with the ever-changing landscape of AI ethics (considering the state-of-the-art in the field). If there are discrepancies, then problem-solving takes over to align the system. The EPS shines on problem-solving because it enacts something only theorized by Jaeggi; it enables a clear connection between dysfunctionality and its normative assignment to reframe the current state of an AI system with the normative expectations of the field of AI ethics through its practical recommendations. Ultimately, the EPS adaptation of Jaeggi’s dynamic normativity and problem-solving approach transforms the principle–practice gap into an ongoing task open to correction and responsive to its surrounding ethical field.

3.2 Finding values: a descriptive analysis of the field

Much like the previous works mentioned [ 45 , 46 , 47 , 48 ], we embarked on the essential task of surveying the landscape of AI ethics to identify its relevant values. However, one factor that distinguishes our framework from the prevailing body of literature rests in the work of descriptive ethics that preceded the development of EPS, where we rooted the principiological foundations of our framework through a descriptive analysis of how the field defines, instrumentalizes, and proposes AI principles worldwide. This descriptive work is entitled Worldwide AI Ethics (WAIE) [ 3 ].

For starters, WAIE draws inspiration from earlier meta-analytical works [ 31 , 32 , 33 , 73 , 74 ] and meticulously surveys a wide range of ethical guidelines related to AI development and governance through a massive effort of descriptive ethics implemented as a data science venture. In it, 200 documents are thoroughly analyzed, including private company governance policies, academic statements, governmental and non-governmental recommendations, and other ethical guidelines published by various stakeholders in over 30 countries spanning six continents. As a result, WAIE identified 17 resonating principles prevalent in the policies and guidelines of its dataset, which now provide the principles we utilize as the basis for the succeeding stages of the EPS, from the assessments to the recommendations.

Besides the fact that EPS stems from our own meta-analysis of the field, we argue that our principiological foundation differs in the following ways from other works:

WAIE uses a worldwide sample of 200 documents, a more diverse representation of global ethical discourse around AI than previous studies.

WAIE delivers its information in a data visualization way that is interactive and searchable and allows the study of correlations.

WAIE is granular, presenting a series of typologies that increase the insight users can gain.

WAIE is open source, allowing users to replicate and extend our results.

By employing the WAIE review to sustain the EPS methodology, we offer a more nuanced and empirically substantiated perspective on the ethical underpinnings of artificial intelligence, thereby enhancing the depth and rigor of our axiological basis. Footnote 4

However, while the utilization of WAIE has undoubtedly provided a valuable foundation for descriptive, and now normative, ethics in AI, we must emphasize the significance of augmenting our value analyses with a critical evaluation of the risks associated with recently released AI models [ 9 , 10 , 10 , 12 , 13 , 75 , 76 ]. As already stated by Bengio et al. [ 30 ], the field’s dynamic and rapidly evolving landscape necessitates constant vigilance in assessing emerging technologies’ potential pitfalls and challenges (e.g., disinformation, algorithmic discrimination, environmental impacts, social engineering, technological unemployment, intellectual fraud, etc.). Hence, we augmented the development of EPS with a risk analysis of large-scale models released in the last 5 years [ 77 ]. Again, we released all materials tied to this analysis as an open and reproducible project. Footnote 5

By incorporating a critical evaluation of values and known risks, we not only provide a more holistic perspective on AI ethics but also equip stakeholders with a timely understanding of the complex ethical considerations surrounding the deployment of AI systems. This integrated approach ensures that our work remains forward-looking and responsive to the ever-changing landscape in the AI field. These are all fundamental aspects for implementing our method, which, in the first instance, is supported by extensive descriptive work. An empirical-descriptive grounding is paramount to any attempt to pragmatize ethics.

However, this descriptive work only serves as a starting point. As already pointed out by Whittlestone et al. [ 78 ] in their critique of the proliferation of AI principles, we must be ready to realize that, by themselves, principles and risks are insufficient to guide ethics in practice. Hence, now that we have made clear the philosophical, axiological, and descriptive roots of our work, we showcase how we translated these into a practical framework for AI development in the following sections.

3.3 Defining risk and impact with algorithmic impact assessment

The first step in the EPS methodology is to gauge the state of a particular system via an impact assessment survey. Our decision to utilize an impact assessment approach comes from the bourgeoning landscape of legislative efforts worldwide currently focused on this topic (i.e., European Union [ 79 ], Brazil [ 80 ], the United States of America [ 81 , 82 ] [ 83 ], several African states [ 84 , 85 , 86 ], Australia [ 87 ], Argentina [ 88 ], Egypt [ 89 ], Japan [ 90 ], Israel [ 91 ], Estonia [ 92 ], Peru [ 93 ], China [ 94 ], Russia [ 95 ], United Kingdom [ 96 ], Canada [ 97 ], among many others). Whether still in production or already enacted, the realization that governments should legally regulate AI is trending toward unanimity. For example, Brazil (the context where EPS came to be) does not have a bill regulating AI systems specifically. However, several bills to govern such technologies are currently the subject of debate in the National Congress. Nevertheless, from our analysis of these aforementioned regulatory efforts, we argue that two main trends are evident:

Determining the fundamental ethical principles to be protected (which we have achieved through the WAIE review).

A risk-based governance approach toward autonomous systems.

For disambiguation purposes, a risk-based governance approach implies a differential treatment concerning AI systems, i.e., different types of systems demand differential treatment pertaining to the risks they pose. For example, a spam filter would not require the same level of auditing and ethical concern as an autonomous vehicle.

In the EPS framework, we argue that the concepts of risk and impact concerning AI ethics differ in the ex-ante and ex-post relationships. Risk refers to the likelihood of negative consequences arising from the deployment and use of artificial intelligence systems, assessing the potential harm that could result from a particular AI application (ex-ante). On the other hand, impact refers to the magnitude and significance of the actual damage or benefit that occurs when these risks materialize or are mitigated, considering the real-world effects of AI systems on individuals, society, and the environment (ex-post). Regardless of their differences, both concepts are related to the impact these technologies can have on the wild.

Even though both of these terms are used interchangeably in many situations, we choose to use the term impact, aiming to encompass the potential problems (risks) and the actual consequences of harmful AI technologies, including their ethical, social, legal, and economic implications. At the same time, the term “impact” assessment is already accepted and used by the literature [ 98 , 99 , 100 ]. We argue that choosing the term “risk” assessment could entail only a preemptive approach toward assessing AI systems’ negative impacts while introducing less-used terminology. Finally, we also point out that many of the known negative impacts of AI systems are currently documented in the form of “impacts” in many publications that present ethical assessments [ 101 , 102 , 103 , 104 , 105 , 106 ].

Another important aspect related to the development of impact assessment methods is that the impact of AI technologies can vary significantly depending on the cultural, social, political, and economic contexts. For example, concerns for the indigenous population must be considered sensitive topics in contexts such as countries that were former subjects of colonial rule. These topics are not quite paramount in the global north. An AI system that is thoughtfully designed concerning local culture, laws, socioeconomic conditions, and ethics is more likely to succeed and less likely to cause harm in our diverse global landscape [ 107 , 108 ]. Hence, accounting for context is a critical step in the EPS framework to ensure a context-sensitive evaluation. As our theoretical grounding session stated, context is paramount to tracing and addressing the normative assignments in our daily practices. In our results section, we will showcase how we executed this in our implementation.

Finally, in our conception, the evaluation stage of any framework entailed in auditing AI systems should not be fully automated. The sole purpose of an evaluation should be to aid and inform a team of human evaluators, e.g., an ethics board, that can use such information to make a normative decision (e.g., a medical ethics board). In other words, we do not agree that ethical evaluations concerning human issues should be a matter of non-human deliberation. The development of this kind of evaluation board is beyond the scope of this study. Our only input is that such a group should be formed in the most interdisciplinary way possible, having representatives of all areas concerning AI ethics, e.g., Computer Science, Philosophy, Law, etc, as already suggested by other works [ 44 , 45 , 46 , 109 , 110 ]. Also, it is essential to note that in our conception, such evaluation should be administered by third parties or sources outside of the developmental cycle of the technologies under consideration, making the EPS framework not a self-evaluation method but a process that requires the collective engagement of AI developers and auditing parties.

3.4 WHY–SHOULD–HOW: a three-step approach to ethical problem-solving

The EPS framework acts as the bridge between the recognition of AI system dysfunctionality and its normative assignment. As previously mentioned, through EPS, it becomes possible to identify, assess, and understand the ethical implications of an AI system. Moreover, EPS provides practical recommendations to reframe an AI system’s actual condition in alignment with the expectations of the AI ethics community. Therefore, problem-solving becomes the act of providing recommendations informed by the current state of the AI system and the normative assignments demanded from within the system and by the field of AI ethics. The WHY–SHOULD–HOW methodology appeared as a palpable and direct form to underline the relevancy of ethical principles in AI development while stating the normative standards and offering comprehensive recommendations to address the issues. This three-step process attempts to grasp the essential features of shortening the principle–practice gap in AI development.

First, the WHY component serves as the foundation, demonstrating the relevance of the AI principles to the specific issue at hand. It encourages practitioners and organizations to acknowledge why they should uphold particular values and the consequences of the opposite. This step is crucial as it sets the stage for a deeper understanding of AI applications’ implications and broader societal and ethical impacts. Also, the WHY step is aligned with the epistemological claim that informed developers are better than uninformed ones, which revolves around the fundamental idea that ethical knowledge and understanding are essential to technological development. In this context, “informed” developers would better understand the principles, best practices, and technologies relevant to their field, while “uninformed” developers may lack this knowledge. In other words, explaining why something is suitable is the first step in any approach that seeks to promote moral reasoning convincingly. Otherwise, starting with an imperative claim may seem authoritative.

While the WHY step represents the foundation and contextualization of principles, SHOULD and HOW are its pragmatization. The last two stages are associated with different levels of impact, inferred at the EPS assessment stage (low, intermediate, and high). Footnote 6 The SHOULD aspect outlines the necessary steps to tackle ethical problems. By necessary, we mean that the measures indicated in this step represent the axiological content of each principle in the framework and are integral to developing an ethically attuned AI system. We can also define the SHOULD stage as an implementation of normative ethics in an applied format, where besides defining explicit “oughts,” we stipulate criteria to help users and evaluators assess the compliance of a given system. This step traces a causal relationship between principles and observables, taking the VCIO approach as inspiration [ 44 , 45 ].

However, the EPS aims to go beyond the normative guidance, presenting the how-to-do, i.e., the practical step. Hence, the HOW component offers valuable tools and strategies to implement the ethical recommendations in the SHOULD stage. In short, it equips developers, researchers, and organizations with the means to put ethical principles into practice. The scarcity of practical tools to address ethical matters within AI is a significant and concerning gap in AI ethics [ 3 , 32 , 111 , 112 ]. Most of the literature that brings ethics to the development of applications does so through its descriptions of principles and extensive flowcharts of how the AI development process should be, failing to provide the practical support to address ethical problems or achieve the principles it underlines. In the meantime, developers often face unique and context-specific dilemmas, and without practical guidance, they may resort to ad-hoc solutions or bypass critical considerations altogether. Without available tools to guide developers or the know-how of how to use them, it is no surprise that there is a lack of standardized ethical practices in AI development [ 113 , 114 ], resulting in blank spots and inconsistencies across the life cycle of AI projects.

Thus, given the deficits mentioned above, the normative step alone is insufficient to cross the principle–practice gap, making the absence of the HOW to step a clear blank spot in other works that our framework seeks to surpass. Ultimately, the WHY–SHOULD–HOW approach culminates in an educational step, acknowledging that responsible AI development is far from standard practice in the curriculum of many STEM-related fields that sprout most AI developers. In our envisioned form, this whole process aims to integrate professionals from humane sciences to STEM-field areas, and vice-versa, bringing developmental focus to ethics and developmental mindset to ethical considerations. Hence, if we suppose, taking a Virtue Ethics stance, that moral behavior can only stem from practice, our proposed framework allows practitioners to develop their virtues through training, which is why we implemented the HOW step as an “educational” step, besides a practical one.

In the following section, we delve into the heart of our work, presenting an implementation of the ethical problem-solving framework. This implementation offers a blueprint for constituting an EaaS platform to apply our envisioned framework.

In this section, we will show an implementation of the EPS framework. The idea of Ethics as a Service guided the creation of this implementation. Following the attempt of Morley et al. [ 49 ], this EaaS implementation would provide the infrastructure for ethical AI development akin to what a Platform as a Service offers, i.e., a platform where developers can submit their systems to ethical auditing.

As mentioned before, the EaaS idea has the advantage of being relatable to modern companies, where third parties constantly subsidize services and infrastructure that is too costly to maintain. While this may not be the case for large organizations (i.e., organizations that, besides having their own technological infrastructure, also have their own ethics boards), an EaaS may as well be a valid tool for companies that cannot sustain or afford their own AI ethics and safety auditing. Also, we again stress the value of a neutral, third-party auditing platform, regardless of the organization’s size.

The components of our envisioned implementation are:

The EPS framework (questionnaires, evaluation metrics, recommendations, and educational aid).

A platform to apply this methodology.

An ethics board to perform the evaluations.

The subsections below present a step-by-step implementation of the EPS framework as an EaaS platform tailored to the Brazilian context.

4.1 Evaluation stage: algorithmic impact assessment and ethical troubleshoot

The flow of the ethical problem-solving framework begins with a pre-algorithmic impact assessment. The pre-assessment gauges preemptively the realm of impact of a particular system, leading to the actual tools of impact assessment. In other words, this preliminary assessment informs the user what algorithmic impact assessment surveys (AIAs) are required to fulfill the evaluation stage. For example, the user must perform the privacy and data protection AIA if the intended application utilizes personally identifiable information. Hence, after this brief assessment, the user is directed to the next step: the EPS’ algorithmic impact assessment surveys (Fig.  2 )

This flowchart illustrates the evaluation structure of the EPS framework. The dotted lines trace the pre-assessment leading through the algorithmic impact assessment surveys. The straight line represents the indispensable survey throughout the framework: the Ethical Troubleshoot. All survey assessments lead to a human-centered evaluation process

The algorithmic impact assessment surveys consist of questionnaires with pre-defined questions and answers that can be single-choice or multiple-choice (Fig.  3 ). Our current implementation of these covers the following themes: data protection and privacy, protection of children and adolescents, antidiscrimination, and consumer rights. Choosing these areas was a strategic move to gather resources since they are rich in legislative content in the Brazilian context. Even though Brazil still needs specific AI regulations, other legal sources can still be used to determine what is and is not allowed regarding the use and development of AI technologies. This design choice also highlights another important aspect of the EPS framework: the importance of contextual information while developing the evaluation stage of an implementation of our framework. This implementation aims to show stakeholders and researchers how these assessments can be created even if AI is not explicitly regulated.

We developed the questions from the algorithmic impact assessment surveys to infer the level of impact a particular system may have on different ethical principles. Each question’s response could either raise the impact, remain unchanged, or lower it if mitigating measures have been found

The questionnaires for the algorithmic impact assessment surveys entail that each of the questions identifies the AI’s compliance with at least three ethical principles identified by the WAIE review. Hence, each of these generates impact scores relative to these assessed principles. As a result, distinct principles may serve as the basis for each question in each AIA. At the same time, each question’s response could either raise the impact, remain unchanged, or lower it if mitigating measures have been found. In other words, we use objective questions tied to contextually relevant legally binding norms intended to guarantee a good life to infer the impact caused by a technology under consideration. Our current implementation of these impact assessment surveys presents for each question:

Their possible answers.

The scores related to each answer.

The principles impacted by the answers to each question.

Ultimately, these assessments can generate a standardized impact level on each ethical principle evaluated by each AIA. At the same time, the overall cumulative impact of all assessed principles represents the general impact of a system against a specific AIA. For example, in our privacy and data protection AIA, the following principles could be impacted, depending on the answers given by the user: privacy, transparency, and accountability. Hence, the final result of the privacy and data protection AIA presents an individual impact score for each principle and an overall score on the AIA itself (the standardized summation of each evaluated principle):

The algorithmic impact assessment surveys use legally binding standards to deduce the implications of AI systems through an objective lens. However, these questionnaires provide an impact score that cannot address all of the ubiquities attached to the ethical issues that AI systems present. Hence, in our current implementation of the EPS framework, we developed a more qualitative survey to accompany them, entitled Ethical Troubleshoot, aimed at going beyond an objective evaluation.

In short, the Ethical Troubleshoot survey seeks to allow the respondent to divulge how the development of a given AI system or application has been done in a human-centric way, e.g., how the needs of the intended users have been considered. It utilizes a combination of multiple-choice, single-choice, and open-ended questions to gauge the system’s scope, its intended and unintended uses, and its target audience. We argue that a mixture of objective evaluation modes and more qualitative assessment forms can only augment a human-centric ethical evaluation (Fig.  4 ). In other words, this qualitative survey is meant to capture the aspects that the more objective and rigid AIAs could not assess. Our implementation of the Ethical Troubleshoot survey was mainly achieved by reverse engineering the VCIO method [ 44 , 45 ] and the Google-PAIR worksheets [ 46 ].

The EPS framework’s evaluations are designed to help human evaluators assess the level of impact of an AI system. Each evaluated principle has three distinct levels of impact. After being informed by the outputs of the evaluation stage, human evaluators prescribe the particular impact level of an AI system regarding the ethical principles being considered

As already mentioned, the sole purpose of these evaluation surveys is to help inform a team of human evaluators, e.g., an ethics board, that can use such information to make a human-centered evaluation. Footnote 7 The output of this decision is the recommendation stage.

4.2 Recommendation stage: WHY–SHOULD–HOW

After the evaluation stage, the EPS framework requires that human evaluators classify the system under consideration in an impact matrix. The matrix comprises three recommendation levels tailored to each impact level—high, intermediate, and low—and six ethical principles gathered from the WAIE review, i.e., fairness, privacy, transparency, reliability, truthfulness, and sustainability. Hence, each principle has three distinct possible recommendations tailored to specific impact levels, e.g., Sustainability-low, Sustainability-intermediate, and Sustainability-high (Fig.  5 ).

After being moderated and revised by an ethics board (human-centered evaluation), the assessment output is an ethical framing, where the system under consideration is classified with an impact level (high, intermediate, and low) for each of the evaluated principles

The WHY–SHOULD–HOW method is the format in which our approach presents the evaluation’s outcome. First, the WHY step is structured to demonstrate the relevancy of each principle, providing the conceptualization and highlighting paradigmatic cases of deficit implementation in a structure that answers the questions “ What is said principle? ” and “ Why should you care about it? ”. Second, the SHOULD step provides the metric utilized to gauge the level of recommendation regarding the corresponding principle, the level of recommendations indicated for the specific case, and the set of recommendations in a summarized form. Third and finally, the HOW component offers the practical tools and strategies required to implement the recommendations made in the SHOULD stage, i.e., it pragmatizes the normative recommendations of the previous step while also providing how-to instructions on using them (Fig.  6 ).

Each level of recommendation regarding the principles utilized is structured around the WHY–SHOULD–HOW method. This allows the evaluators to make differential recommendations based on each principle’s inferred level of impact. Subsequently, each level of impact presents differential recommendations with tailored tools and practices for that specific impact level

We developed the EPS framework to address the principle–practice gap explicitly. Given that we wished to go beyond simply “pointing to what tools developers can use,” we developed an open-source repository containing a collection of demonstrations of how to use the developmental tools we recommend as part of the EPS [ 116 ]. This repository has many examples of tools and techniques developed to deal with the potential issues of an AI application (e.g., algorithmic discrimination, model opacity, brittleness, etc.), using as test cases some of the most common contemporary AI applications (e.g., computer vision, natural language processing, forecasting, etc.) (Fig.  7 ).

We structured the HOW step around the question, “ How can a developer increase the robustness of a system in regards to a specific ethical principle? ”, laying down step-by-step instructions, toolkits, educational resources, testing, and training procedures, among other resources that can help apply the principles used in the EPS

The effectiveness of a tool or framework often hinges on its ease of use and implementation. For instance, PyTorch’s popularity in deep learning stems from, besides the inherent value of its automatic differentiation engine, its comprehensive documentation, which facilitates widespread adoption by lowering the entry barrier for newcomers to the field, simplifying complex procedures, like neural network engineering and training, to simple how-to-do examples. In the case of the EPS, we try to accomplish the same for the practices related to AI ethics and safety. To better exemplify this, let us describe a hypothetical use of our framework as an EaaS platform:

Hypothetical use case: An organization in Brazil is in the process of developing an AI-empowered product. Before deploying it to its first users, the organization applies the EPS method via an EaaS platform to access ethical and legal compliance. During the evaluation stage, the organization answers the surveys in the best way possible, disclosing all information required, protected by a non-disclosure agreement between both parties. After the evaluation stage, the ethics board working behind the platform receives the results of the product evaluations. This information is also disclosed to the organization since it gives valuable information about the legal compliance of the product under several legislative works. Imbued with the results of the evaluation stage, the ethics board frames the product into the pre-established impact levels, giving rise to a particular set of recommendations (WHY–SHOULD–HOW). The EaaS platform delivers this documentation back to the organization, together with their tailored review. This deliverable presents criteria and tools to improve a product according to the principles under consideration. To further help, these tools are offered in a pedagogical form, i.e., via documented examples of use cases (e.g., how to perform adversarial exploration, evaluate fairness metrics, interpret language models, etc.), to improve their adoption and use. These are presented as step-by-step procedures to improve the organization’s product further.

The workflow of our implementation combines aspects related to ethical principles, legal compliance, and technical implementations, articulating all of them together akin to the “Stronger Together” proposal of Pistilli et al. [ 117 ]. This, we argue, ultimately leads to tightening the principle–practice gap, i.e., from AI to beneficial AI. Meanwhile, the point in which our framework goes beyond past works in shortening the principle–practice gap lies heavily in our pedagogical aspect. Past frameworks almost always give you the normative (the ought) and, more rarely, the practical (the how). Besides giving normative recommendations, our framework seeks to teach developers how to use tools to tackle ethical and safety problems. At the same time, the materials that support our framework are all openly available, making this study one of the first efforts to tailor an AI governance approach to the Brazilian context in an open-source fashion while also allowing for spin-offs tailored to different contexts. Readers can access the source code and all other materials tied to the EPS framework at the following URL https://nkluge-correa.github.io/ethical-problem-solving/ .

5 Limitations and future works

EPS represents a novel attempt to bridge the principle–practice gap between ethical principles and the practical implementation of AI development. However, much work remains.

First, while the framework provides a structured approach to addressing ethical concerns, handling AI’s ever-evolving landscape is a tiring feature we must come to terms with. Many foundations of our approach rest on work that is bound to be outdated. The axiological roots of the EPS framework (the WAIE review), the legislative sources we used to create our impact assessment surveys, our risk assessment catalog, and the practices we recommend and exemplify are all bound to become irrelevant as advances in these fronts occur. Hence, the fast-moving nature of the field requires implementations of our framework to undergo constant recycling; otherwise, we risk falling into uselessness. This fact leads us to the question, “Can humane sciences accompany the accelerated pace of the technological industry?” which, in our opinion, has till now been answered negatively and with pessimism. As mentioned, bridging the principle–practice gap is a continuous problem-solving process, and as Jaeggi points out [ 52 ], problem-solving is a never-ending work. Hence, an obvious future avenue of work involves updating and extending EPS. For example, there is undoubtedly space to expand the legislative contribution in subsequent implementations of the EPS, even more so if this expansion encompasses legislation specifically focused on AI systems (e.g., Generative AI). Also, it remains open the possibility to integrate more general frameworks, like international human rights [ 118 , 119 , 120 ], which are already a part of some impact assessment tools tailored to the assessment of human rights [ 121 ].

Another sensitive point of our framework regarding its evaluation method is its scoring process. In our current implementation, we constrained our impact score scale to a standard range, where answers to questions could maximally impact a given principle with a score of 1 or decrease its impact with a score of − 1. At the same time, we chose to have more ways in which impact scores could be increased rather than decreased, i.e., \(\approx \frac{1}{5}\) of the general score from each AIA produces a reduction in impact, given that the main objective of our evaluations is to assess the lack of ethical compliance. Hence, one issue we face is the feasibility of translating regulatory standards into a cardinal evaluation scale. The problem of intertheoretic comparisons and the gauging of how much “utility” a developmental choice should represent is not trivial, being an area of open research in Moral Philosophy and Metanormativity [ 122 , 123 , 124 , 125 ]. Given that we developed these evaluations to inform a body of experts imbued with making a normative framing, finding ways to present evaluations understandably and unambiguously is crucial. While our approach translates standardized cardinal values to unambiguous impact classes (low, intermediate, and high), other methods might be better suited. Searching for improved ways to perform this translation is an area of study worth pursuing.

Meanwhile, the idea of a human-centric evaluation presents its own problem. This human-centric aspect, which, in the end, comes down to the biased and subjective view of a group of individuals, is one of its weak spots. Like many other forms of evaluation and certification that rely on human oversight, the EPS may also fall short of its promise if its human element is unaligned with the core purpose of the framework. While the idea of an EaaS platform that should always be managed (or at least audited) by an external party, such as a government body or a neutral auditing organization, may help avoid specific perils without a proper normative engine (a.k.a. an evaluation board or an ethics committee), the whole idea of EaaS could deteriorate into Ethics Washing as a Service [ 126 ]. We remain committed to the concept of not automatizing ethics. However, we argue that the success of this type of framework also rests in the question of how to train and educate good ethical boards to perform this crucial role [ 109 , 110 ], which is another avenue for future studies.

This work also explored the limitations of shortening the principle–practice gap with a toolbox mentality. In other words, we encountered several cases where we may need more than mere tools and the knowledge of how to use them to fulfill our higher goals. For example, one can use statistical metrics and other tools to assess the fairness of an AI system and further correct them. However, these do not attack the root cause of inequality that becomes imprinted in our systems and applications [ 127 ]. One can use several methods to protect and obfuscate personally identifiable information during an AI development cycle. Yet, robust privacy can only be achieved through collaboration among governments, private companies, and other stakeholders [ 128 , 129 ]. One can use carbon tracking to offset their footprint and promote sustainable practices. But unfortunately, sustainability in AI ethics cannot be reduced to such mere calculations, given the many other side effects of our technological progress do not have an easy-to-measure metric, e.g., the depletion of our natural resources [ 130 ], the humanitarian costs related to their extraction [ 5 , 131 , 132 ], production of e-waste [ 133 ], etc. All these cases stress that AI ethics has challenges beyond the “lack of implementational techniques” or “knowledge gaps,” which should incentivize works and agendas that use a different approach than the one we utilized in the EPS.

Other limitations that fall outside the scope of our framework but can prevent (or improve) its success are these:

Regulatory supports and incentives: frameworks like the EPS and the development of EaaS platforms can become a future necessity if regulatory bodies make this evaluation a mandatory procedure for AI products above a certain impact level. At the same time, regulatory bodies could adopt frameworks like this and serve it as a certification system. On the other hand, it could also be the case that frameworks like these would only have adoption with regulatory support and, without it, would find no adoption in the industry. Just as in the case of organizations that provide cybersecurity and GDPR compliance services, their adoption is tied to regulations that make compliance necessary for them.

Lack of attention to ethical issues in entrepreneurial environments: as already mentioned by previous studies [ 114 ], entrepreneurial settings, where many modern technologies become commoditized into products, are not environments that usually take ethical concerns too seriously, where these are generally seen as a nuisance or barrier to further progress. Currently, the field necessitates minimizing knowledge asymmetry between all sectors, from humane sciences to STEM fields. To research and develop business and applications. However, there are still many obstacles to this type of collaboration and how to overcome the challenges of interdisciplinary research.

Lack of virtues in AI and software developers: studies have already shown that we might have a Humane Sciences gap in STEM areas [ 113 , 134 ], while questions related to AI ethics and safety are still far from the mainstream in terms of Academic curricula. However, we argue that practices like the ones promoted in our framework could help patch this in the educational development of STEM professionals acting as AI engineers and developers. Regardless, improving the “Humane aspect” of the formation of these professionals could undoubtedly improve their sensibility to the issues dealt with by frameworks like the EPS.

These are research areas that can, directly and indirectly, improve not just the success of this study’s objectives, i.e., shortening the principle–practice gap, but AI Ethics itself. In this landscape, our framework is a blueprint for other researchers to build upon and expand. The entirety of the proposed process gives more than enough space for the many areas related to AI ethics to contribute, like, for example, improving evaluation methods, coming up with recommendations for ethical design choices, creating tools, or teaching developers how to use them. Our objective for the foreseeable future is to fully implement and test the EPS framework as an EaaS platform in Brazil while supporting and updating our open-source repositories. We hope this work and service may provide novel pathways and opportunities to AI ethicists and better general guidance and assistance to the field.

6 Conclusion

In this work, we presented ethical problem-solving, a broad framework pursuing the betterment of AI systems. Our framework employs impact assessment tools and a differential recommendation methodology. The EPS framework differentiates itself from other works by its theoretical grounding, axiological foundation, and operationalization, serving as the blueprints for an EaaS platform that mediates the normative expectations of the field and the reality of AI system development. However, crossing the principle–practice gap in AI development is an ongoing process. Even though many problems remain without immediate technical solutions, efforts like this can help institute a culture of responsible AI practice in a way that can keep pace with the advances in the field. Finally, by opening our work and results, we hope other researchers can easily extend and surpass our work.

Data availability

The authors confirm that all data related to this study is available at the following URL: https://github.com/Nkluge-correa/ethical-problem-solving .

In this study, the term Artificial Intelligence (AI), and more specifically, AI systems and applications, are defined as the products generated by the interdisciplinary field of computer science, cognitive sciences, and engineering that focus on creating machines capable of performing tasks that typically require human intelligence (e.g., natural language processing, problem-solving, pattern recognition, decision-making, forecasting, etc.) [ 14 ].

All materials tied to this study are available in https://nkluge-correa.github.io/ethical-problem-solving/ .

Jaeggi’s contribution to Critical Theory with the reclamation of Hegelian normative touchstones, drawing on Dewey [ 60 , 61 ], MacIntyre [ 62 , 63 , 64 ], and Pinkard [ 65 ], sets itself apart from the deconstructive [ 66 , 67 ] and the overarching constructivist [ 68 , 69 ] branches of the tradition.

All materials tied to the WAIE review are available in https://nkluge-correa.github.io/worldwide_AI-ethics/ .

Available in https://github.com/Nkluge-correa/Model-Library .

Increasing levels of impact demand additional recommendations and more severe implementations.

We define “human-centeredness,” also known as human-centric design or user-centered design [ 115 ], as a methodology employed across various disciplines to prioritize human needs, behaviors, and preferences in creating and optimizing products, services, and systems. Hence, by a “human-centered evaluation,” we here refer to a process of ethical deliberation and reasoning that requires human involvement and not automatized processes.

Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst. 25 (2012)

Sejnowski, T.J.: The Deep Learning Revolution. MIT Press, Cambridge (2018)

Book   Google Scholar  

Corrêa, N.K., Galvão, C., Santos, J.W., Del Pino, C., Pinto, E.P., Barbosa, C., Massmann, D., Mambrini, R., Galvão, L., Terem, E., Oliveira, N.: Worldwide AI ethics: a review of 200 guidelines and recommendations for AI governance. Patterns 4 (10), 100857 (2023)

Article   Google Scholar  

Maslej, N., Fattorini, L., Brynjolfsson, E., Etchemendy, J., Ligett, K., Lyons, T., Manyika, J., Ngo, H., Niebles, J.C., Parli, V., Shoham, Y., Wald, R., Clark, J., Perrault, R.: The AI index 2023 annual report. Technical report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April (2023)

Falk, S., van Wynsberghe, A.: Challenging AI for sustainability: what ought it mean? AI Ethics (2023). https://doi.org/10.1007/s43681-023-00323-3

Schmidhuber, J.: Deep learning in neural networks: an overview. Neural Netw. 61 , 85–117 (2015)

LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521 (7553), 436–444 (2015)

Chollet, F.: Deep Learning with Python. Manning, Shelter Island (2017)

Google Scholar  

OpenAI: Gpt-4 technical report. arXiv, pp. 2303–08774 (2023)

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752 (2021)

Gemini Team, Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356 (2022)

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, pp. 8748–8763. PMLR (2021)

Russell, S.J., Norvig, P.: Artificial Intelligence a Modern Approach. Prentice-Hall, Upper Saddle River (2010)

Noorbakhsh-Sabet, N., Zand, R., Zhang, Y., Abedi, V.: Artificial intelligence transforms the future of health care. Am. J. Med. 132 (7), 795–801 (2019)

Schwalbe, N., Wahl, B.: Artificial intelligence and the future of global health. Lancet 395 (10236), 1579–1586 (2020)

Umamaheswari, S., Valarmathi, A., et al.: Role of artificial intelligence in the banking sector. J. Surv. Fish. Sci. 10 (4S), 2841–2849 (2023)

Joh, E.E.: Artificial intelligence and policing: first questions. Seattle UL Rev. 41 , 1139 (2017)

Ezzeddine, Y., Bayerl, P.S., Gibson, H.: Safety, privacy, or both: evaluating citizens’ perspectives around artificial intelligence use by police forces. Policing Soc. 33 , 861–876 (2023)

Lin, L., Hou, Z.: Combat COVID-19 with artificial intelligence and big data. J. Travel Med. 27 (5), taaa080 (2020)

Cugurullo, F., Acheampong, R.A.: Fear of AI: an inquiry into the adoption of autonomous cars in spite of fear, and a theoretical framework for the study of artificial intelligence technology acceptance. AI Soc (2023). https://doi.org/10.1007/s00146-022-01598-6

World Economic Forum: Artificial intelligence in wildfire prediction and prevention (2023). Accessed 8 Sept 2023

Governança e ética da inteligência artificial: perspectivas globais (2019)

Inteligência artificial e governança eficaz: uma agenda de revisão, crítica e pesquisa. 2 , 100004 (2020)

Governança de dados: organizando dados para inteligência artificial confiável. 37 (3), 101493 (2020)

Taeihagh, A.: Governance of artificial intelligence. Policy Soc. 40 (2), 137–157 (2021)

Russell, S.: Human Compatible: Artificial Intelligence and the Problem of Control. Penguin, London (2019)

Juric, M., Sandic, A., Brcic, M.: AI safety: state of the field through quantitative lens. I: 2020 43rd International Convention on Information, Communication and Electronic Technology (MIPRO), pp. 1254–1259. IEEE (2020)

Hendrycks, D., Carlini, N., Schulman, J., Steinhardt, J.: Unsolved problems in ml safety (2022)

Bengio, Y., et al.: Managing AI risks in an era of rapid progress (2023)

Jobin, A., Ienca, M., Vayena, E.: The global landscape of AI ethics guidelines. Nat. Mach. Intell. 1 (9), 389–399 (2019)

Hagendorff, T.: The ethics of AI ethics—an evaluation of guidelines. Minds Mach. 30 , 99–120 (2019)

Fjeld, J., Achten, N., Hilligoss, H., Nagy, A., Srikumar, M.: Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-Based Approaches to Principles for AI. Berkman Klein Center for Internet and Society, Cambridge (2020)

Schiff, D.S., Laas, K., Biddle, J.B., Borenstein, J.: Global AI ethics documents: what they reveal about motivations, practices, and policies. In: Codes of Ethics and Ethical Guidelines, pp. 121–143. Springer International Publishing, Cham (2022)

Chapter   Google Scholar  

Future of Life Institute: Pause giant AI experiments: an open letter (2023)

Bergengruen, V.: AI regulation takes baby steps on capitol hill (2023)

Kang, K.: Openai’s sam altman urges A.I. regulation in senate hearing (2023)

Gerke, S., Babic, B., Evgeniou, T., Glenn Cohen, I.: The need for a system view to regulate artificial intelligence/machine learning-based software as medical device. NPJ Digit. Med. 3 (1), 53 (2020)

Hine, E., Floridi, L.: The blueprint for an AI bill of rights: in search of enaction, at risk of inaction. Minds Mach. 33 , 285–292 (2023)

Finocchiaro, G.: The regulation of artificial intelligence. AI Soc. (2023). https://doi.org/10.1007/s00146-023-01650-z

Schiff, D.S., Laas, K., Biddle, J.B., Borenstein, J.: Global AI Ethics Documents: What They Reveal About Motivations, Practices, and Policies, pp. 121–143. Springer International Publishing, Cham (2022)

Taddeo, M., Floridi, L.: How AI can be a force for good. Science 361 (6404), 751–752 (2018)

Article   MathSciNet   Google Scholar  

Bærøe, K.: Translational ethics: an analytical framework of translational movements between theory and practice and a sketch of a comprehensive approach. BMC Med. Ethics 15 (1), 1–7 (2014)

Peylo, C., Slama, D., Hallensleben, S., Hauschke, A., Hildebrandt, S.: VCIO based description of systems for AI trustworthiness characterisation (2022). Accessed 14 Sept 2023

AI Ethics Impact Group: From principles to practice—an interdisciplinary framework to operationalize AI ethics (2020). Accessed 14 Sept 2023

Google: People + AI guidebook (2019). Accessed 14 Sept 2023

Digital Catapult: Ethics framework (2021). Accessed 14 Sept 2023

Microsoft: Microsoft responsible AI standard, v2 (2022). Accessed 14 Sept 2023

Morley, J., Elhalal, A., Garcia, F., Kinsey, L., Mökander, J., Floridi, L.: Ethics as a service: a pragmatic operationalisation of AI ethics. Minds Mach. 31 , 239–256 (2021)

Baker-Brunnbauer, J.: TAII framework for trustworthy AI systems. ROBONOMICS J. Autom. Econ. 2 , 17 (2021)

Constantin Ciobanu, A., Mesnită, G.: AI ethics for industry 5.0—from principles to practice (2022)

Jaeggi, R., Cronin, C.: Critique of Forms of Life. Belknap Press, Cambridge (2018)

Iser, M.: Recommendation and Progress: Foundations of a Critical Theory of Society, vol. 64. Campus Verlag, Franfurt Am Main (2008)

Liu, X., Xie, L., Wang, Y., Zou, J., Xiong, J., Ying, Z., Vasilakos, A.V.: Privacy and security issues in deep learning: a survey. IEEE Access 9 , 4566–4593 (2021)

Stahl, T.: Habermas and the project of immanent critique (2013)

Iser, M.: Rational reconstruction. In: The Habermas Handbook, pp. 614–618. Columbia University Press, New York City (2017)

Wittgenstein, L.: Philosophical investigations (1994)

Feenberg, A., et al.: Critical Theory of Technology, vol. 5. Oxford University Press, New York (1991)

Feenberg, A.: Questioning Technology. Routledge, London (2012)

Dewey, J.: Logic: the theory of inquiry, 1938 (1999)

Dewey, J., Rogers, M.L.: The Public and Its Problems: An Essay in Political Inquiry. Penn State Press, University Park (2012)

MacIntyre, A.: Three Rival Versions of Moral Enquiry: Encyclopaedia, Genealogy, and Tradition. University of Notre Dame Press, Notre Dame (1994)

MacIntyre, A.: After Virtue. A &C Black, London (2013)

MacIntyre, A.: Whose justice? Which rationality? In: The New Social Theory Reader, pp. 130–137. Routledge, London (2020)

Pinkard, T.: Hegel’s Phenomenology: The Sociality of Reason. Cambridge University Press, Cambridge (1994)

Mendieta, E.: Global Fragments: Globalizations, Latin Americanisms, and Critical Theory. State University of New York Press, Albany (2012)

Allen, A.: Critique on the Couch: Why Critical Theory Needs Psychoanalysis. Columbia University Press, New York City (2021)

Forst, R.: Justice, Democracy and the Right to Justification: Rainer Forst in Dialogue. Bloomsbury Academic, London (2014)

Benhabib, S.: Another Cosmopolitanism. Oxford University Press, Oxford (2008)

Niederberger, A., Weihrauch, T.: Rahel jaeggi: Kritik von lebensformen. Notre Dame Philos. Rev. 2015 , 25 (2015)

Pinkard, T.: Kritik von lebensformen, by Rahel Jaeggi. Suhrkamp verlag, Frankfurt (2017)

Santos, J.W., Sobottka, E.A.: Genealogy, immanent critique and forms of life: a path for decolonial studies. Hum. Aff. 33 (1), 101–114 (2023)

Zeng, Y., Lu, E., Huangfu, C.: Linking artificial intelligence principles. arXiv preprint arXiv:1812.04814 (2018)

Schiff, D., Borenstein, J., Biddle, J., Laas, K.: Ai ethics in the public, private, and NGO sectors: a review of a global document collection. IEEE Trans. Technol. Soc. 2 (1), 31–42 (2021)

Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., et al.: Constitutional AI: harmlessness from AI feedback. arXiv preprint arXiv:2212.08073 (2022)

Anil, R., Dai A.M., Firat, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)

Corrêa, N.K., Naqvi, F., Rossain, R.: Model library. https://github.com/Nkluge-correa/Model-Library (2024)

Whittlestone, J., Nyrup, R., Alexandrova, A., Cave, S.: The role and limits of principles in AI ethics: towards a focus on tensions. In: Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pp. 195–200 (2019)

The European Parliament: Amendments adopted by the European Parliament on 14 June 2023 on the proposal for a Regulation of the European Parliament and of the Council on laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts (2023). 14.06.2023

Senado Federal do Brasil: Projeto de lei n \(^\circ\) 2338, de 2023. dispõe sobre o uso da inteligência artificial. Relator Senador Eduardo Gomes (2023)

Zhu, K.: The state of state AI laws: 2023. 8 (2023)

NCSL: National Conference of State Legislatures: Artificial intelligence 2023 legislation 9 (2023)

White House Office of Science and Technology Policy: Blueprint for an AI bill of rights: making automated systems work for the American people. The White House, October (2022)

Federal Ministry of Communications and Digital Economy: National digital economy policy and strategy (2020–2030) for a digital Nigeria (2020)

Floyd, R., Goriola, O.: How Africa can benefit from artificial intelligence: the AI regulation landscape 8 (2023)

Dugbazah, J.: The African Union artificial intelligence continental strategy for Africa (2023)

Energy Department of Industry, Science and Resources: Supporting responsible AI: discussion paper 7 (2023)

de Tecnología de la Información, S.: Recomendaciones para una inteligencia artificial fiable (2023)

Ministry of Communications and Information Technology: The national AI strategy (2019). 2023-10-11

Trade Ministry of Economy and Industry: Governance guidelines for implementation of AI principles (2022)

Science Minister of Innovation and Technology: Minister of innovation, science and technology policy—regulation and ethics for the field of artificial intelligence for public comments (2023)

Ministry of Economic Affairs and Communications: Report of estonia’s ai taskforce (2019)

Congress of the Republic of Peru: Law n \(^\circ\) . 31814 (law promoting the use of artificial intelligence for and social development of the country) (2023)

Cyberspace Administration of China. Measures for the management of generative artificial intelligence services (draft for comment) (2023)

Russian Federation: Federal law of the Russian Federation of 02.07.2021 no. 331-f. on experimental legal regimes in the field of digital innovations in the Russian Federation (2021)

Innovation Parliament by the Secretary of State for Science and Technology by Command of His Majesty: A pro-innovation approach to AI regulation (2023). 2023-10-11

Government of Canada. The artificial intelligence and data act (AIDA)—companion document (2023). Date modified: 2023-03-13

Government of Canada. Algorithmic impact assessment tool (2023)

The Danish Institute for Human Rights: Human Rights Impact Assessments (2023)

United Nations Educational, Scientific and Cultural Organization: Ethical impact assessment: a tool of the recommendation on the ethics of artificial intelligence. United Nations Educational, Scientific and Cultural Organization, 7, place de Fontenoy, 75352 Paris 07 SP, France (2023)

LaRosa, E., Danks, D.: Impacts on trust of healthcare AI. In: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pp. 210–215 (2018)

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33 , 1877–1901 (2020)

Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Kaplan, J, Edwards, H., Burda, Y., Joseph, N., Brockman, G., et al.: Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 (2021)

Ligozat, A.-L., Lefèvre, J., Bugeau, A., Combaz, J.: Unraveling the hidden environmental impacts of AI solutions for environment. arXiv preprint arXiv:2110.11822 (2021)

Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: WebGPT: browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021)

Saisubramanian, S., Roberts, S.C., Zilberstein, S.: Understanding user attitudes towards negative side effects of AI systems. In: Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pp. 1–6 (2021)

McLennan, S., Fiske, A., Celi, L.A., Müller, R., Harder, J., Ritt, K., Haddadin, S., Buyx, A.: An embedded ethics approach for AI development. Nat. Mach. Intell. 2 (9), 488–490 (2020)

Hermann, E., Hermann, G., Tremblay, J.-C.: Ethical artificial intelligence in chemical research and development: a dual advantage for sustainability. Sci. Eng. Ethics 27 , 1–16 (2021)

Coelho, A.Z., et al.: Framework for AI ethics committees: governance of artificial intelligence in organizations (2023)

Schuett, J., Reuel, A., Carlier, A.: How to design an AI ethics board. arXiv preprint arXiv:2304.07249 (2023)

Morley, J., Kinsey, L., Elhalal, A., Garcia, F., Ziosi, M., Floridi, L.: Operationalising AI ethics: barriers, enablers and next steps. AI Soc. 38 , 411–423 (2021)

Georgieva, I., Lazo, C., Timan, T., van Veenstra, A.F.: From AI ethics principles to data science practice: a reflection and a gap analysis based on recent frameworks and practical experience. AI Ethics 2 (4), 697–711 (2022)

McNamara, A., Smith, J., Murphy-Hill, E.: Does ACM’s code of ethics change ethical decision making in software development? In: Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pp. 729–733 (2018)

Vakkuri, V., Kemell, K.-K., Jantunen, M., Abrahamsson, P.: “This is just a prototype’’: how ethics are ignored in software startup-like environments. In: International Conference on Agile Software Development, pp. 195–210. Springer, Cham (2020)

Boy, G.A.: The Handbook of Human–Machine Interaction: A Human-Centered Design Approach. CRC Press, Boca Raton (2017)

Corrêa, N.K.: Teeny-tiny castle (2024)

Pistilli, G., Muñoz Ferrandis, C., Jernite, Y., Mitchell, M.: Stronger together: on the articulation of ethical charters, legal tools, and technical documentation in ML. In: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp. 343–354 (2023)

UNH Rights: Universal declaration of human rights (1961)

Kilkelly, U.: The Child and the European Convention on Human Rights. Routledge, London (2017)

UN General Assembly: Convention on the rights of the child. U. N. Treaty Ser. 1577 (3), 1–23 (1989)

Kemp, D., Vanclay, F.: Human rights and impact assessment: clarifying the connections in practice. Impact Assess. Project Apprais. 31 (2), 86–96 (2013)

Gracely, E.J.: On the noncomparability of judgments made by different ethical theories. Metaphilosophy 27 (3), 327–332 (1996)

MacAskill, W.: Normative uncertainty. PhD thesis, University of Oxford (2014)

MacAskill, M., Bykvist, K., Ord, T.: Moral Uncertainty. Oxford University Press, Oxford (2020)

Corrêa, N.K., de Oliveira, N.F.: Metanormativity: solving questions about moral and empirical uncertainty (2020)

van Maanen, G.: Ai ethics, ethics washing, and the need to politicize data ethics. Digit. Soc. 1 (2), 9 (2022)

Speicher, T., Heidari, H., Grgic-Hlaca, N., Gummadi, K.P., Singla, A., Weller, A., Zafar, M.B.: A unified approach to quantifying algorithmic unfairness: measuring individual & group unfairness via inequality indices. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 2239–2248 (2018)

Humerick, M.: Taking AI personally: how the EU must learn to balance the interests of personal data privacy & artificial intelligence. Santa Clara High Tech. LJ 34 , 393 (2017)

Yanisky-Ravid, S., Hallisey, S.: ‘Equality and privacy by design’: ensuring artificial intelligence (AI) is properly trained & fed: a new model of AI data transparency & certification as safe harbor procedures. Available at SSRN 3278490 (2018)

Filho, W.L., Hunt, J., Lingos, A., Platje, J., Vieira, L.W., Will, M., Gavriletea, M.D.: The unsustainable use of sand: reporting on a global problem. Sustainability 13 (6), 3356 (2021)

Mustapha, A.O., Mbuzukongira, P., Mangala, M.J.: Occupational radiation exposures of artisans mining columbite-tantalite in the eastern democratic Republic of Congo. J. Radiol. Prot. 27 (2), 187 (2007)

Sutherland, E.: Coltan, the Congo and your cell phone. Available at SSRN 1752822 (2011)

Islam, M.T., Iyer-Raniga, U.: Life cycle assessment of e-waste management system in Australia: case of waste printed circuit board (PCB). J. Clean. Prod. 418 , 138082 (2023)

Gogoll, J., Zuber, N., Kacianka, S., Greger, T., Pretschner, A., Nida-Rümelin, J.: Ethics in the software development process: from codes of conduct to ethical deliberation. Philos. Technol. 34 , 1085–1108 (2021)

Download references

Acknowledgements

This research was funded by RAIES (Rede de Inteligência Artificial Ética e Segura). RAIES is a project supported by FAPERGS (Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul) and CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico).

Open Access funding enabled and organized by Projekt DEAL.

Author information

Authors and affiliations.

University of Bonn, Bonn, NRW, Germany

Nicholas Kluge Corrêa

PUCRS, Porto Alegre, Brazil

Camila Galvão, Marcelo Pasetti, Dieine Schiavon & Nythamar De Oliveira

Brown University, Providence, USA

Faizah Naqvi & Robayet Hossain

RAIES (Rede de Inteligência Artificial Ética e Segura) network, Porto Alegre, RS, Brazil

Nicholas Kluge Corrêa, James William Santos, Camila Galvão, Marcelo Pasetti, Dieine Schiavon & Nythamar De Oliveira

Universität Erfurt, Erfurt, Germany

James William Santos

You can also search for this author in PubMed   Google Scholar

Contributions

This study involved a collaborative effort from a multidisciplinary team. N.K.C. contributed to the development of the methodology, the creation of the demo and related code repositories, and the writing of the article. J.W.S. contributed to the development of the methodology, the documentation of the demo, and the writing of the article. C.G. contributed to the development of the AIAs, to writing the corresponding sections related to the AIAs, and aided in the development of the overall methodology. M.P. contributed to the development of the AIAs, to writing the corresponding sections related to the AIAs, and aided in the development of the overall methodology. D.S. contributed to the development of code repositories, to the writing of the article, and aided in the development of the overall methodology. F.N. contributed to the development of the recommendation approach and aided in the development of the overall methodology. R.H. contributed to the development of the recommendation approach and aided in the development of the overall methodology. N.O. is the project coordinator.

Corresponding author

Correspondence to Nicholas Kluge Corrêa .

Ethics declarations

Conflict of interest.

The authors declare no conflict of interest.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Corrêa, N.K., Santos, J.W., Galvão, C. et al. Crossing the principle–practice gap in AI ethics with ethical problem-solving. AI Ethics (2024). https://doi.org/10.1007/s43681-024-00469-8

Download citation

Received : 17 December 2023

Accepted : 07 March 2024

Published : 15 April 2024

DOI : https://doi.org/10.1007/s43681-024-00469-8

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Artificial intelligence
• Principle–practice gap
• Ethics as a service
• Find a journal
• Publish with us
• Track your research