Integrating an explanation-based learning mechanism into a general problem-solver

Abstract

In this paper, we study the problem of integrating an Explanation-Based Learning mechanism into a general and industrial problem solver. During our investigations into SOAR we discovered a number of weaknesses related to its architecture and learning mechanism, known as Chunking. Using general concepts on which SOAR is based, we define a new learning system based, on the one hand, on a general and industrial problem solver, and, on the other hand, on an efficient learning mechanism known as EBG (Explanation-Based Generalization). Due to the fact that EBG sometimes learns production rules which are too general, we have introduced the possibility of restricting the generality of learned rules, in order to improve significantly the performances of industrial applications. We have tested this system on an industrial application at Thomson. In this application, we had to find correct trajectories for planes through a network of valleys. This type of problem is complex and the search combinational, as the valleys were not properly interconnected. However, learning 327 production rules made the resolution 78 times faster and the system sometimes even found better solutions.