GPS (cognitive architecture)

GPS (cognitive architecture)
Name	General Problem Solver (GPS)
Developer	Allen Newell; Herbert A. Simon; J. C. Shaw
First release	1957
Programming language	IPL (Information Processing Language)
Platform	Early mainframes
Genre	Cognitive architecture; problem-solving system

GPS (cognitive architecture) is an early symbolic cognitive architecture designed to model human problem solving and planning. Developed in the 1950s by researchers associated with Carnegie Mellon University, RAND Corporation, and Pittsburgh, GPS aimed to demonstrate that general-purpose procedures could reproduce aspects of human reasoning across diverse domains. The system influenced research at institutions like Massachusetts Institute of Technology, Stanford University, Harvard University, University of Pennsylvania, and Bell Labs and shaped subsequent work in computational cognitive science and artificial intelligence.

Overview

GPS was conceived as a domain-independent problem solver built on the notion that human cognition can be represented by procedures operating on symbolic representations. Founders used concepts drawn from contemporary research at Carnegie Mellon University and RAND Corporation to argue that a small set of heuristics could generate behavior resembling that of human subjects tested at University of Michigan and Princeton University. Implemented in Information Processing Language on early mainframe hardware from IBM, GPS demonstrated applications ranging from puzzle solving to theorem proving in the milieu of postwar research institutions such as Bell Labs and Cornell University.

History and development

GPS emerged from collaborations led by Allen Newell and Herbert A. Simon with contributions by J. C. Shaw and colleagues during an intellectual exchange involving John von Neumann-era computing, the RAND Corporation workshops, and the cognitive psychology programs at Carnegie Mellon University and Harvard University. Early demonstrations in the late 1950s followed precedents set by systems like Logic Theorist and experiments in the Book of the Century era of Princeton University AI activity. Funding and institutional support came from organizations including Office of Naval Research, National Science Foundation, and private foundations tied to research at Stanford Research Institute. GPS's dissemination influenced curricula at Massachusetts Institute of Technology and seminars at Cornell University and led to debates at conferences such as the Dartmouth Conference (1956) and meetings of the Association for Computing Machinery.

Architecture and components

GPS combined a symbolic memory structure, production-like procedures, and heuristic control in the tradition established by earlier projects at Carnegie Mellon University and RAND Corporation. Core components included a goal stack inspired by models from Harvard University cognitive labs, a set of operators resembling production systems used later at Massachusetts Institute of Technology, and a control architecture that selected operators via heuristics influenced by work at Bell Labs and Princeton University. Implementation relied on Information Processing Language and was deployed on machines designed by IBM engineers, aligning with programming practices at University of Pennsylvania and Cornell University computing centers.

Cognitive modeling and processes

GPS modeled human-like decomposition of problems into subgoals, backtracking, and operator selection using heuristics that mirrored protocols from Harvard University and experimental paradigms at Yale University and University of Michigan. The architecture operationalized concepts from the cognitive psychology literature promoted by scholars at Carnegie Mellon University and Stanford University and engaged with theories advocated by figures associated with Princeton University and Harvard University. It sought to instantiate general problem-solving mechanisms comparable to behavioral data collected in studies at Columbia University and University of Chicago, and it informed theoretical dialogues at gatherings of the American Psychological Association and the Cognitive Science Society.

Implementations and applications

Early implementations of GPS ran on IBM 704 and similar mainframes at facilities including Carnegie Mellon University, RAND Corporation, and Bell Labs'. Applications demonstrated included solving cryptarithms, geometric proofs, and toy tasks similar to those studied at Harvard University and Stanford University. GPS influenced subsequent systems developed at Massachusetts Institute of Technology (e.g., production-system research), inspired work at SRI International (formerly Stanford Research Institute), and informed applied projects sponsored by the Office of Naval Research and Defense Advanced Research Projects Agency. Later derivatives and descendants appeared in projects at University of California, Berkeley, University of Illinois Urbana–Champaign, and Carnegie Mellon University labs focusing on automated planning and symbolic reasoning.

Evaluation and comparisons

Contemporaneous evaluations compared GPS to systems like Logic Theorist, production systems emerging from Massachusetts Institute of Technology, and planning systems developed at Stanford Research Institute. Empirical investigations at Carnegie Mellon University, Harvard University, and Yale University assessed GPS's ability to reproduce human problem-solving protocols; results fueled debates at venues including the Association for the Advancement of Artificial Intelligence and workshops organized by the National Academy of Sciences. Comparative analyses highlighted GPS's elegance and its limitations relative to domain-specific theorem provers and later architectures from Massachusetts Institute of Technology and Stanford University that used richer representations and learning mechanisms.

Criticisms and limitations

Critics affiliated with research programs at MIT, Stanford University, and Harvard University pointed to GPS's reliance on hand-coded heuristics and symbolic operators as a constraint on scalability and ecological validity. Empirical work at Columbia University, University of Michigan, and Princeton University emphasized that human problem solving often employs domain knowledge and learning mechanisms not captured by GPS, a critique echoed in discussions at RAND Corporation and policy panels convened by the National Science Foundation. Subsequent developments in connectionist models advanced at University of California, San Diego and Massachusetts Institute of Technology and probabilistic frameworks from Stanford University further highlighted areas where GPS's symbolic approach struggled to account for perceptual learning, parallel processing, and adaptive behavior.

Category:Cognitive architectures