Active learning from cross perceptual aliasing caused by directteaching

摘要

Proposes an active learning method by which the learner hascapabilities of self-learning and understanding the instructions bycoping with cross perceptual aliasing problem caused by the state spacedifference between the learner and the teacher. The learner asks theteacher to give an appropriate instruction when necessary to reduce theinstruction frequency. Further, the learner finds taught datainconsistent with learner's state space caused by cross perceptualaliasing, and modifies its state space based on the clustering methodC4.5 so that it can successfully achieve the goal by reinforcementlearning. The method is applied to the domain of RoboCup, where alearning robot attempts to approach and shoot a ball into the goal withthe help of instructions given by teacher on request. The experimentalresults are shown and a discussion is given