Representations for Action Selection Learning from Real-Time Observation of Task Experts

The association of perception and action is key to learning by observation in general, and to programlevel task imitation in particular. The question is how to structure this information such that learning is tractable for resource-bounded agents. By introducing a combination of symbolic representation with Bayesian reasoning, we demonstrate both theoretical and empirical improvements to a general-purpose imitation system originally based on a model of infant social learning. We also show how prior task knowledge and selective attention can be rigorously incorporated via loss matrices and Automatic Relevance Determination respectively.