Supplemental observation acquisition for learning by observation agents

摘要

Learning by observation agents learn to perform a behaviour by watching an expert perform that behaviour. The ability of the agents to learn correctly is therefore related to the quality and coverage of the observations. This article presents two novel approaches for observation acquisition, mixed-initiative observation acquisition and delayed observation acquisition, that allow learning agents to identify problems they are having difficulty solving and ask the expert for assistance solving them. The observation approaches are presented in the context of a case-based learning by observation agent and empirically compared to traditional passive observation in the domain of Tetris. Our results show that not only do the mixed-initiative and delayed observation acquisition approaches result in observations that cannot be obtained in a passive manner, but they also improve the learning performance of an agent.