An Analytic Learning System for Specializing Heuristics

This paper describes how meta-level theories are used for analytic learning in MULTI-TAC. MULTI-TAC operationalizes generic heuristics for constraint-satisfaction problems, in order to create programs that are tailored to specific problems. For each of its generic heuris-tics, MULTI-TAC has a meta-theory specifically designed for operationalising that heuris-tic. We present examples of the specialisation process and discuss how the theories influence the tractability of the learning process. We also describe an empirical study showing that the specialised programs produced by MULTI-TAC compare favorably to hand-coded programs. 1 Introduction MULTI-TAC (Multi-Tactic Analytic Compiler) is a learning system for constraint-satisfaction problems (CSPs). The system operationalises generic heuristics[ll], producing problem-specific versions of these heuristics, and then attempts to find the most useful combination of these heuristics on a set of training problems. This paper focuses on the knowledge that MULTI-TAC uses in order to operationalize generic heuristics, and how this approach differs from previous work in "ana-lytic" (or knowledge-based) speed-up learning. Previous analytic speed-up methods, such as EBL, chunking, and derivational analogy, have been used primarily for caching problem-solving experience[9]. Typical EBL systems , for example, learn from problem-solving successes and/or failures by caching a knowledge structure sum-marising the experience (e.g., a chunk) and then reusing that knowledge during subsequent problem solving. In retrospect, relatively little attention has been paid to the theories (i.e., the knowledge) used in the learning process. However, this subject deserves more attention since the choice of theories determines what is learned. MULTI-TAC employs meta-level theories in the learning process. These enable the system to reason about the problem solver's base-level theory, as opposed to simply caching the generalised results of the problem solver's search. We argue that this approach is particularly appropriate when solving combinatorial problems, such as scheduling problems. The system employs a rich variety of meta-level theories. This reflects a shift in research focus from "learn-ing as a caching process" to "learning as an inferential process". The key is to find tractable meta-theories for generating useful search control knowledge. We outline two such theories, and discuss how the representation of the meta-level theories and the representation of the underlying constraint-satisfaction task influences the utility of the learning process. We also describe an empirical study in which MULTI-TAC compared favorably with hand-coded programs. 2 Theories and Analytic Learning Informally, analytic learning systems are characterized by a "theory-driven" component that generates hypotheses by analyzing a domain. Several analytic approaches …