Herbert Simon (1916-2001) is best known for his work in economics, but he was a true polymath whose research contributions spanned many other fields such as public and business administration, organizational theory, information and decision theory, cognitive psychology, computer science, philosophy of science, and artificial intelligence (AI). The common thread in Simon’s body of research was his self-described “monomania” with decision-making, particularly within the context of social organizations.
Simon began his academic career studying political science and economics at the University of Chicago. His early research, starting with his doctoral dissertation, focused on developing a theory of administration for structuring and operating organizations in order to perform correctly and efficiently. He addressed topics including marshaling and coordinating workers, and incentivizing them to align their interests to their organization when making decisions. He also defined a supporting process model for decision-making, consisting of data gathering, formulating decision options, and choosing a course of action. Two themes from this initial work remained central to his later research:
Hierarchy: a principle for structuring organizations (or decisions) top-down and coordinating workers (or problem-solving)
Limits of rationality: how workers’ skills, values and knowledge influence and constrain their ability to perform, including making correct decisions.
Following his graduate work at Chicago and positions at the University of California and Illinois Institute of Technology, Simon joined the faculty at Carnegie Mellon University (CMU) in 1949, where he remained for the rest of his academic career, working in departments including industrial administration, computer science, and psychology.
Simon is best known for his theories of bounded rationality and satisficing, which attacked rational choice theory. This latter theory, a cornerstone of mainstream economics, asserts that people act as perfectly rational actors when making decisions: they are assumed to possess full information about their decision options and consequences, have the ability to define complete and consistent preferences about outcomes (and quantify them as utilities), and always chose the option that maximizes utility. Simon disputed all of these points, arguing that human rationality is bounded because we are finite and fallible beings operating in complex social environments, such as markets. Real-world constraints include limited cognitive capacities, bounded time intervals and budgets, incomplete information about decisions, uncertainty about future events, and imperfect knowledge for accurately predicting consequences. As a result, we are incapable of choosing an optimal (i.e., utility maximizing) decision. Instead, we are confined to satisfice or select a decision option that appears to be “good enough” given moderate effort.
At CMU, Simon studied mathematical economics, exploring formal methods for making optimal decisions, both under certainty and uncertainty. He also began a long collaboration with Allen Newell and J. C. Shaw to study human problem-solving and decision-making behaviors. They soon hit upon the strategy of using digital computers to model and simulate human cognitive processes. Much of this work was coupled with psychological experiments consisting of interviews with human test subjects as they solved puzzles and performed other tasks. The trio developed the first programming language tailored for symbolic reasoning—the cornerstone of early AI. They used this list processing language to create what is widely considered to be the first AI program, which automatically proved theorems of mathematical logic by applying rules of deductive inference.
Simon, Newell, and Shaw went on to develop another important software program called General Problem Solver (GPS). This was the first AI system that explicitly separated generalized problem-solving strategies from knowledge specific to particular domains, captured in the form of logical if-then rules called productions. GPS was limited to solving relatively simple problems, such as cryptarithmetic, logic, and block stacking puzzles. Solving such puzzles is driven largely by deductive reasoning, possibly supplemented by some simple strategy productions to guide reasoning. Understanding how people and computer programs solve such problems provided a baseline for Simon’s cognitive model of deliberate reasoning.
Simon went on to study more semantically rich problems, for which solutions require detailed knowledge about specific domains. Examples include chess, route planning, and resource allocation. GPS provided a generalized architecture for building AI “engines” to solve these complex problems: Simon’s if-then production rules could be used to model and apply requisite knowledge and specialized expertise. Programs called expert systems automated reasoning that could augment or replace humans in performing complex tasks such as medical diagnosis, equipment troubleshooting and repair, architectural design, and scheduling and planning. For example, medical knowledge could be captured by rules of the form IF Observation1 = X AND Test2 = Y AND … THEN conclude DiagnosisZ).
Exploring problem-solving was a natural extension of Simon’s research on decision-making. In particular, creating alternative plans is a complex design problem that represents a crucial stage in any deliberate decision-making process. Simon’s interviews with human experts and his AI simulations of problem-solving enabled him to better understand constraints on human and artificial reasoning to design decision options and evaluate (or search among) large numbers of possible solutions. This research also helped validate his concept of satisficing.
Simon also made important contributions to the philosophy of science, in the areas of causality, complexity, and the logic of scientific discovery. Notably, he collaborated on developing BACON, an AI system that automatically uncovered regularities in numeric experimental data. BACON re-derived physical laws including Kepler’s third law of planetary motion (astronomy), Ohm’s law (electricity), and Snell’s law (optics). Simon viewed scientific discovery as an “ill-structured” problem because it only has a weakly defined goal—finding “interesting” or “significant” patterns in data. By contrast, critical economic decisions generally have more well-defined goals such as achieving 5% annual growth. BACON supported Simon’s belief that his models of human and artificial cognition could account for both kinds of problem-solving.
A prolific author, Simon published over 800 papers and 27 books. He won numerous awards including the Nobel Prize in Economics in 1978 for his work on bounded rationality and decision-making processes within economic organizations, the prestigious Turing Award (along with colleague Alan Newell) in 1975 for pioneering contributions to AI, cognitive psychology, and list processing, and won the National Medal of Science in the Behavioral Sciences in 1986.
For More Information
See Herbert Simon’s Sciences of the Artificial for an overview of his research on bounded rationality, cognitive psychology, and artificial intelligence. See also Simon’s Nobel Prize lecture, his autobiography, and CMU links to his research. Simon is one of the central figures in my book on critical decision-making, Bending the Law of Unintended Consequences.