Learning Search Strategies through Discrimination

SAGE is an adaptive production system model of strategy learning. The system begins a task with weak, overly general operators and uses these to find a solution to some problem by trial and error. The program then attempts to resolve the problem, using its knowledge of the solution path to determine blame when an error occurs. Once the faulty operator has been found, the system employs a process of discrimination to generate more conservative versions of the rule containing additional conditions. Such variants are strengthened each time they are relearned, until they come to override their precursors. The program continues to learn until it can solve the problem without errors. SAGE has learned useful heuristics in the domains of the slide-jump puzzle, solving simple algebra equations and seriating blocks of different lengths.