An Entropy-Based Approach to the Hierarchical Acquisition of Perception-Action Capabilities

摘要

We detail an approach to the autonomous acquisition of hierarchical perception-action competences in which capabilities are bootstrapped using an information-based saliency measure. Our principle aim is hence to accelerate learning in embodied autonomous agents by aggregating novel motor capabilities and their corresponding perceptual representations using a subsumption-based strategy. The method seeks to allocate affordance parameterizations according to the current (possibly autonomously-determined) learning goal in a manner that eliminates redundant percept-motor context, thereby obtaining maximal parametric efficiency. Experimental results within a simulated environment indicate that doing so reduces the complexity of a multistage perception-action learning problem by several orders of magnitude.