Perceptual and functional representations: Bridging the gap

An agent operating in an unpredictable world must be prepared to adapt its plans to the circumstances that it actually encounters. To do this, it must recognise when unpredicted situations aaect its goals. In general, because of the complexity of both the world and the agent's plans, it is infeasible to consider every element of every situation that the agent encounters in the context of each of its goals. A productive approach is to use a ltering mechanism to indicate those elements that are likely to be signiicant. Such a mechanism must be computationally eecient, so that the agent can respond in a timely manner, and highly predictive, so that the responses are appropriate. The eeciency requirement indicates that the mechanism must be based on easily perceivable features, and the prediction requirement indicates that it must be based on functional features. In this paper I present a type of feature that combines these two requirements and that can be used as the basis of an eeective ltering mechanism. I discuss the prevalence of these features, reference features, in everyday life and consider how they can form the basis of expertise.