Composer: A decision-theoretic approach to adaptive problem solving

• Authors: Jonathan M Gratch
• Venue: Composer: A decision-theoretic approach to adaptive problem solving
• Year: 1993

While the general formalization of planning and scheduling problems is NP-hard in worst-case complexity, in practice, for specific distributions of problems and constraints, domain-specific solutions have been shown to perform in much better than exponential time. Unfortunately, constructing and improving custom techniques is a knowledge-intensive and time-consuming process that requires a deep understanding of the application. The goal of our work is to develop techniques to allow for the automatic transformation of a general problem solver into one customized for framework with which to describe the problem. We then present a statistically sound learning algorithm built upon this framework. Our COMPOSER method implements a hill-climbing search through a space of possible problem solving modifications. The learning system attempts to identify a sequence of modifications that improve problem solving performance with respect to the problem distribution and performance requirements for a given application. The thesis includes two applications that illustrate the technique''s power and flexibility. In the first, we show how COMPOSER can improve the performance of the PRODIGY planning system, and that it outperforms several state of the art learning techniques. We then discuss an application of the approach to the real-world domain of scheduling satellite communications. Using problem distributions based on NASA mission requirements, COMPOSER identified strategies that both decrease the amount of CPU time required to produce schedules, and increase the percentage of problems that are solvable within computational resource limitations. We also provide formal proofs of the performance characteristics of the system.