General Problem Solver

General Problem Solver
Name	General Problem Solver
Developer	Newell and Simon
Released	1959–1961
Language	IPL, Lisp (later implementations)
Platform	PDP-1, PDP-6, DEC systems, simulated environments
Genre	Problem-solving system, artificial intelligence

General Problem Solver

The General Problem Solver was an early symbolic artificial intelligence program developed in the late 1950s and early 1960s by researchers at Carnegie Mellon University and RAND Corporation associated with Herbert A. Simon and Allen Newell. It sought to model human problem solving by using a formalized means-ends analysis and production rules to transform states, and influenced subsequent work at Massachusetts Institute of Technology, Stanford University, and University of California, Berkeley. The project intersected with contemporaneous efforts such as ELIZA, Logic Theorist, GPS (cognitive architecture) research, and early computer science curricula at institutions like MIT and Harvard University.

Overview

GPS aimed to provide a domain-independent solver that could handle proofs, puzzles, and routine tasks by operating on symbolic representations like those used in Logic Theorist and in the planning components of SOAR. The system used formal operators, initial states, and goal states expressed in a representation language influenced by programming languages of the era including IPL, LISP, and implementations on hardware such as the PDP-1 and PDP-6. GPS was situated within the broader movement of symbolic AI alongside projects at IBM, RAND Corporation, and Bolt, Beranek and Newman (BBN) that pursued automated reasoning, problem decomposition, and production systems.

Design and Architecture

The architecture centered on a problem-space hypothesis that modeled tasks as nodes in a space connected by operators, akin to frameworks used later in STRIPS and A* algorithm-based planners. GPS employed a control structure with recursive search, backtracking, and goal decomposition comparable to techniques used in PROLOG and in the design of Expert systems at Stanford Research Institute. Component influences included theorists and laboratories such as Herbert A. Simon, Allen Newell, John McCarthy, Marvin Minsky, and research groups at Carnegie Mellon University and MIT. The system’s representation of states and operators anticipated work in knowledge representation used by projects at SRI International and Xerox PARC.

Algorithms and Heuristics

GPS used means-ends analysis, a heuristic search method related to hill-climbing and subgoal reduction, and integrated operator selection heuristics that resembled mechanisms in production systems like those later in OPS5. The search strategy combined depth-first techniques with heuristic evaluation akin to later informed search strategies such as A* algorithm and heuristic functions developed in Edinburgh research on automated planning. Its rule-based transformations were precursors to unification and pattern-matching procedures found in PROLOG and influenced reasoning modules in architectures like SOAR and ACT-R.

Historical Development and Implementation

Development occurred during a period of intensive AI exploration at institutions including Carnegie Mellon University, RAND Corporation, MIT, and IBM Research. Early demonstrations took place on hardware at Carnegie Mellon and on systems associated with Bell Labs and DEC machines such as the PDP-1. Key publications appeared in venues connected to the Association for Computing Machinery and in symposiums like those organized by the American Psychological Association and IEEE. The team’s work drew on prior successes such as Logic Theorist and was contemporaneous with projects like ELIZA, SAM, and SHRDLU that explored language, planning, and interaction.

Applications and Impact

GPS influenced automated theorem proving in environments such as Stanford Linear Accelerator Center computing efforts and academic work at Princeton University and Yale University. Its problem-space approach seeded techniques in automated planning used at NASA for mission planning, inspired production-system implementations at Bell Labs and Xerox PARC, and informed cognitive modeling in laboratories at Harvard University and University of Michigan. Subsequent systems drawing on GPS concepts include planners and rule-based systems developed at SRI International, MITRE Corporation, Honeywell, and in academic projects at University of Edinburgh and University of Toronto.

Criticisms and Limitations

Contemporaries such as Marvin Minsky and John McCarthy critiqued GPS for its reliance on symbolic representations and brittle heuristics, limitations echoed later in discussions at DARPA and within debates over the scope of symbolic AI versus connectionist approaches advanced at institutions like Bell Labs and Cognitive Science programs at MIT. GPS struggled with combinatorial explosion in large problem spaces, faced representation bottlenecks noted in publications from Stanford University and Carnegie Mellon University, and was eventually supplanted in many applications by domain-specific planners such as STRIPS and statistical methods from research groups at University of California, Berkeley and University of Toronto.

Category:Artificial intelligence history