A Computational Model of a Gazing Point Controlling Process in a Hierarchical Classifier System
Takahiro Yoshimi, Toshiharu Taura · Transactions of the Society of Instrument and Control Engineers · 2000
When a machine learning system solves problems, all of the ‘properties’ of the input information are not always required. Therefore, the authors assume that a certain mechanism which can select specific properties to be gazed. For the realization and discussion of this mechanism, the authors have focused on the Classifier System (CS). The CS also involves such “duality” that both the optimization processes of rules for problem solving and the generalization processes of input information are in a single process, which may lead to problems. In this paper, the authors propose a computational model in which these two processes are explicitly separated. The key concept of the proposed model is the Gazing point Controlling Process for the purpose of using rules for selecting properties to be gazed. This is separated from the standard rules for solving problems. A computer system is developed to evaluate the utility of this model. The results acquired by applying the model to an example problem are summarized as follows: (1) by means of using the model here, the process of selecting properties are explicitly observed which is hidden in the standard CS, (2) the model can handle another kind of input information of selecting properties and (3) it results in more effectively than the standard CS.