Modelling Learning for Intelligent Software Agents: A Connectionist Approach
Dan W. Joyce, Paul H. Lewis · ePrints Soton (University of Southampton) · 1999
This paper aims to show how a connectionist model can provide a form of adaptive action selection mechanism (ASM) for reactive virtual agents. By adopting a horizontally layered control architecture, we can build an agent with the ability to learn associations between sensory input and internal state to produce and adapt predictions or responses. At the lowest level, stimuli are categorised by a plastic self-organising mechanism which then activates a prediction module. Subsequently, if the prediction module's action results in a harmful environmental consequence, a conditioning network (reflecting internal state) modifies the agent's choice of prediction during the remainder of its attempt to find the optimal action. This acquisition of behaviour is regulated by a control layer and finally, an application-specific layer.