Management of Experience Data for Rapid Adaptation to New Preferences Based on Bayesian Significance Evaluation

Abstract

This paper presents a rapid adaptation method of behavior preference based on Bayesian significance evaluation of experience data. Rapid adaptation to user preferences cannot be achieved when data from every process cycle are used for learning because significant data are not differentiated from insignificant data. We propose a method to solve this problem by selecting significant data for the learning based on the change in the degree of confidence of the behavior decision. A small change in the degree of confidence can be regarded as reflecting insignificant data for learning, so that data can be discarded. Accordingly, the system can avoid having to store too frequent experience data and the robot can adapt rapidly to changes in the user preferences. We discuss the experimental results of two experiments in which user preference changes among three preferences on a mobile robot. In an interactive experiments with a robot following its user preference with a data frequency of 5 Hz, the robot could adapt to a new preference within 3.75 s.