The General Problem Solver’s Impact on Early AI Research

This article explores how the General Problem Solver (GPS), developed in 1957 by Herbert A. Simon, J.C. Shaw, and Allen Newell, shaped the foundational era of artificial intelligence. It covers the core mechanism behind the program—means-ends analysis—its attempt to mimic human cognitive processes, its structural separation of knowledge from problem-solving strategies, and how its eventual limitations guided the shift toward modern, specialized AI systems.

At the core of the General Problem Solver’s influence was a technique called means-ends analysis. This method allowed the program to solve problems by defining a current state and a desired goal state, and then identifying the differences between the two. GPS would search for actions, or "operators," that could reduce these differences one step at a time. When an operator could not be applied directly, GPS created a sub-goal to change the current state so the operator could be used. This step-by-step planning approach became a standard technique in computer science and remains the basis for many modern automated planning and routing algorithms.

Beyond pure mechanics, GPS profoundly influenced cognitive science by attempting to replicate how humans actually think. Simon, Shaw, and Newell did not just want a machine that could solve puzzles; they wanted a machine that solved them the way a person would. By recording people talking through logic problems aloud and comparing those transcripts to the steps taken by GPS, the researchers established the field of cognitive modeling. This proved that computer programs could be used as scientific tools to study and test theories about human psychology.

Another major contribution of GPS to early AI was the separation of problem-solving knowledge from the strategy used to apply it. Before GPS, programs were tightly coded to do one specific thing, such as play checkers or prove mathematical theorems. GPS was designed with a universal problem-solving engine that could theoretically tackle any formalized problem, provided the user inputted the specific rules and environment of that problem. This structural separation laid the direct groundwork for the "expert systems" of the 1970s and 1980s, which separated the "inference engine" (the logic) from the "knowledge base" (the facts).

Despite its initial success with simple logic puzzles like the Tower of Hanoi, the failures of GPS were just as influential as its triumphs. As researchers fed more complex, real-world problems into the program, it suffered from a "combinatorial explosion." The number of possible steps and sub-goals grew so large that the computer ran out of memory and processing power, making it unable to find a solution.

This limitation taught early AI researchers a critical lesson: general-purpose search strategies are not enough to achieve high-level intelligence. The inability of GPS to handle complex problems caused a major pivot in the AI community during the late 1960s and 1970s. Researchers moved away from searching for a single, universal problem-solving algorithm and began focusing on domain-specific programs packed with specialized knowledge, forever changing the trajectory of artificial intelligence development.