Essay: The General Problem Solver

Overview
Herbert Simon's discussion of the General Problem Solver (GPS) presents a concrete program and a broader theoretical frame for how intelligent agents can tackle a wide range of problems. GPS was built around the idea that diverse tasks share a common structure that can be captured in terms of states, goals, and operators that transform states. By formalizing problem solving as the manipulation of symbolic representations within a search space, GPS offered a way to simulate human-like reasoning with mechanistic procedures.

Core idea: problem space and means-ends analysis
At the heart of GPS is the notion of a problem space: a set of possible states connected by operators, with a specified initial state and a goal state or condition. GPS does not rely on domain-specific algorithms; instead it treats operators as general actions that can be applied whenever their preconditions are met. Means-ends analysis is the principal control strategy. Rather than blindly exploring the space, GPS identifies differences between the current state and the goal and selects operators likely to reduce those differences. When an operator cannot be applied directly, GPS creates a subgoal to establish the operator's preconditions and pursues those subgoals recursively.

Mechanism and operation
GPS implements its ideas using a production-system style architecture: condition-action rules detect differences, choose operators, and create new subgoals. The system interleaves forward and backward reasoning by attempting to apply operators to reduce discrepancies and, if necessary, working backwards to make an operator applicable. Search is guided by heuristics embedded in the control rules, which prefer moves that reduce salient differences. When operators fail or lead to dead ends, GPS backtracks and tries alternatives. The program was tested on a variety of tasks that have clear symbolic representations and well-defined operators, demonstrating how the same control mechanisms could handle logic proofs, puzzle solving, and algebra-like manipulations.

Strengths and conceptual contribution
GPS made several important conceptual contributions. It showed that a small set of general procedures, coupled with appropriate representations, could account for a surprising range of problem-solving behavior. The explicit separation between general search control and domain-specific knowledge highlighted the importance of representations and heuristics. Means-ends analysis became a central concept in both artificial intelligence and cognitive psychology, shaping how researchers thought about strategy selection, subgoaling, and the role of heuristics in boundedly rational agents.

Limitations and critiques
GPS also revealed the limits of early symbolic approaches. The system depended on well-structured problem representations and clearly specified operators; many real-world problems are ill-defined or require perceptual grounding and rich background knowledge that GPS could not supply. Its difference-reduction strategy can get stuck in local improvements that do not lead to global solutions, and the combinatorial explosion of search remains a fundamental challenge when operators generate large branching. Subsequent work criticized GPS for its brittleness and for oversimplifying human cognition, prompting research into richer knowledge structures, learning mechanisms, and more flexible control strategies.

Legacy
Despite its limitations, GPS had a lasting impact. It crystallized the problem-space hypothesis and popularized means-ends analysis as a practical control principle. The architecture influenced later production-system models and cognitive architectures, and its emphasis on heuristics and representation continues to inform work on planning, search, and cognitive simulation. GPS stands as an early, influential demonstration that general procedures plus structured knowledge can produce behavior resembling human problem solving, while also highlighting the need for richer knowledge and adaptive control in more complex domains.