Learning to sense selectively in physical domains

Abstract

In this paper we describe an approach to representing, using, and improving sensory skills for physical domains. We present ICARUS, an architecture that represents control knowledge in terms of durative states and sequences of such states. The system operates in cycles, activating a state that matches the environmental situation and letting that state control behavior until its conditions fail or until finding another matching state with higher priority. Information about the probability that conditions will remain satisfied minimizes demands on sensing, as does knowledge about the durations of states and their likely successors. Three statistical learning methods let the system gradually reduce sensory load as it gains experience in a domain. We report experimental evaluations of this ability on three simulated physical tasks: flying an aircraft, steering a truck, and balancing a pole. Our experiments include lesion studies that identify the reduction in sensing due to each of the learning mechanisms and others that examine the effect of domain characteristics.