Dynamic Parameter Change in Learning by Switching Rule-Generation Method

Abstract

When we solve a problem, we firstly have no knowledge. With observing data, we are constructing knowledge for solving the problem and finally have complete one. We proposed the system that may switch a reasoning method to another, reconstruct the rule set, or switch a knowledge representation to another according to the number of incorrect reasoning results. When a rule reconstruction does not improve an evaluation, parameters in a rule generation method are changed to generate a better rule set. If the parameter change also does not improve the evaluation, then the system switches a knowledge representation. We apply the system to Stefan Aeberhard's wine data.