Dynamical Categories and Control Policy Selection

Every autonomous agent operating in realistic settings must deal with incomplete state information. Perceptual limitations often compromise system observability, and may prevent the acquisition of optimal control policies for a given task. This paper addresses the observability problem within the control composition framework, where the agent selects which policy to adopt next from a set of pre-defined control policies. The idea is to treat the agent in its environment as a dynamical system, and augment the perceived state space (situation space) with contextual cues extracted empirically as the agent exercises each of the existing control policies. Contextual cues are provided by the correlation between dynamic features of the agentenvironment interaction and agent performance. Initial experiments involving an agent with impoverished sensing capabilities in a simulated, dynamic environment demonstrate that relevant contextual information can be extracted and used to enhance the agent’s performance.