Acquiring Classifiers for Bipolarized Reward by XCS in a Continuous Reward Environment

Abstract

In data mining, it is important to clarify how effective the acquired rules are and which elements are affected by rule evaluation. Extended learning classifier system (XCS) reveals factors that affect the classifier (rule) evaluation by generalizing the multiple classifiers that acquire the same reward (evaluation value) into a generalized classifier. In a real-world problem, because the reward of the classifier varies, XCS cannot acquire the generalized classifier. As useful classifiers with a narrow range of the acquired rewards are required, this paper proposes a new XCS (XCS based on reward bipolarization: XCS-RB) that acquires the classifiers that acquire only high rewards and classifiers that acquire only low rewards. XCS-RB was applied to the problems such as predicting the ratio of the deep sleep time of the night of the day from the care plan implemented in the care house and predicting the calculation time of a matrix-matrix product using SGEMM GPU kernel. XCS-RB acquired rules indicating care plan that leads to deep sleep and parameter settings with short calculation time. XCS-RB was able to acquire the generalized classifiers so as not to conflict with the input data; in this paper, the potential advantages of XCS-RB have been demonstrated.

Keywords

• learning classifier system
• accuracy criteria
• reward