Speedup Mechanisms for Large Learning Systems
Résumé
Eliminating combinatorics from the match in production systems is important for expert systems, real-time performance, machine learning, parallel implementation and cognitive modeling. We describe a way of managing the tradeoff between generality and efficiency in knowledge representation for large learning systems. We propose an architecture that enables to combine efficiency in problem solving to generality in learning. Our architecture combines generality and efficiency by using two problem solvers. The first one is interpreted and uses a general knowledge representation. It enables the system to learn general rules. The second one is compiled and uses a specialized knowledge representation. It enables the system to solve problems rapidly and to detect when learning can occur in order to decide to call the first problem solver. To speedup rules, we use two mechanisms which do not affect the generality of learned rules and three mechanisms that alter the learning abilities of the system and that are only used in the second problem solver. This approach has shown its efficiency in its application to the game of Go. The game of Go is the most complex two person complete information game.