The decision process in selective attention

Abstract

A learning system that can be trained to generate t h e necessary sequential decision policy for maneuvering through a multiple T-maze, is proposed. The .Selectively Attentive Environmental Learning System (SAELS) has a n architecture with two types of neurally-inspired learning mechanisms to model and act on its environment. The first mechanism, which has been shown to model 19 forms of classically conditioned behavior [l], generates an internal reinforcement signal that anticipates the external reinforcement signal that SAELS will receive. The second learning mechanism is a single-layer associative neural network with competitive learning. This mechanism uses the internal reinforcement of the first learning mechanism to learn the association between the state of SAELS and the decision that must b e generated in order to successfully complete the maze. It is shown that SAELS solves the temporal credit assignment problem that arises when reinforcement is only available at the end of a sequence of actions. The general requirements of a system of selective attention, and how SAELS meets these requirements is discussed.