Susan Epstein, Jack Gelfand, Esther Lock
If a system is to both satisfice and compute with a high degree of reliability, a balance must be struck. This paper describes an architecture that effectively integrates correct reasoning with satisficing heuristics. As a program inductively learns new heuristics, the architecture underlying it robustly incorporates them. Many of these heuristics represent significant information previously inexpressible in the program’s representation and, in some cases, not readily deducible from a description of the problem class. After training, the program exhibits both an emphasis on its new heuristics and the ability to respond correctly to novel situations with them. This significantly improves the program’s performance. The domain of investigation is board games, but the methods outlined here are applicable to an autonomous learning program in any spatial domain.