Identifying the Right Reasons: Learning to Filter Decision Makers

Given a domain of related problem classes and a set of general decision-making procedures applicable to them, this paper describes AWL, an algorithm to filter out those procedures that prove irrelevant, selfcontradictory, or untrustworthy for a particular class. With an external model of expertise as its performance criterion, the algorithm uses a perceptron-like model to learn problem-class-specific weights for its decision-making procedures. Learning improves both the efficiency of the decision-making process and the performance of the system. 1. The Reasoning Framework The problem-solving and learning architecture called FORR (FOrr the Right Reasons) relies upon a set Advisors, general purpose, heuristic rationales that make decisions across a set of related problem classes (Epstein 1994). This paper describes an algorithm that learns the relevance of each Advisor to any particular problem class. In this context, relevance is the usefulness of a generalpropose heuristic for decision making in a specific problena class. The thesis behind FORR is that a "general expert" in a domain can become a "specific expert" for some problem class in that domain by learning problem-class-specific data (useful knowledge) that is potentially applicable and probably correct As it combines powerful heuristics, FORR simulates a synergy among good reasons as a mental model for decision making. A background theory for a general domain is defined in FORR by aproblem frame that delineates the nature of a problem class, a behavioral script that represents how an expert proceeds in the domain, a useful knowledge frame that identifies what can be learned about a problem class, and a set of "right reasons," heuristic procedures (Advisors) that represent reasonable arguments for decision making throughout the domain. One or more learning methods is attached to each useful knowledge slot and triggered by the behavioral script. The same domain-specific Advisors, learning methods, and behavioral script are used on every problem class; problem-specific information is represented only in the data that initialize the problem frame and the data learned to instantiate the useful knowledge frame. Although several FORR-based programs are in development, the most accomplished is Hoyle, for the domain of two-person, perfect information, finite board games (Epstein 1992~ In Hoyle, a problem class is a game, such as chess or tic-tac-toe. Each game is defined by a different instantiation of the problem flame that describes the board, the playing pieces, and the rules. The behavioral script, for all games, describes how contestants take turns, make legal moves, and so on. The useful knowledge frame holds data worth learning, such as average contest length or good openings. Hoyle has 23 game-independent Advisors. These include one that concerns mobility, another that considers material (number of pieces on the board), and another that monitors openings. acquired useful knowledge current state Tier 1: Shallow search and inference based on perfect knowledge legal moves