2A2-E07 Management of Experience Data for Rapid Adaption to New Preferences based on Bayesian Significance Evaluation

概要

論文の詳細を見る
An interactive Bayesian policy learning algorithm is presented that enables robots to adapt more rapidly to users' new preferences. Bayesian reasoning is used in many robotics applications when there is significant uncertainty accompanying perception and action. For Bayesian belief changes in query nodes, we are generally more interested in evidence that may lead to a change in decision. If an observation has little effect on a decision, it can be regarded as an insignificant observation for the learning process. To realize such an experience management system, we propose an algorithm for discarding insignificant data. The algorithm uses a Dirichlet-distribution-based significance test to discard insignificant data. The test is used to compare the change in distribution for consecutive data points. The algorithm can learn interactive behavior rapidly with very little data and can adapt rapidly to new situations. The time required for adaptation with the proposed algorithm depends on the discarding criteria, which are affected by the frequency with which interactive data are presented to the system-the higher the frequency, the more rapid the adaptation. In experiments with an interactive data frequency of 5 [Hz], an actual robot could adapt to a new situation within 3.75 [sec].
一般社団法人日本機械学会の論文
2009-05-25