Engineering Strategic Behavior in a Multi-Layered Behavior Based Agent

Intelligent autonomous agents need to be able to exhibit goal-directed behavior and reactivity to unexpected events in a complex, dynamic environment. Behavior based architectures decompose systems into behaviors, small independent processes each controlling some aspect of the agent's interaction with the world. Such systems have been shown to eeectively integrate goal-directed behavior and reactivity. However when a behavior based system has multiple complex goals, managing the interactions that occur between behaviors in the system and the world may become infeasible. We have implemented a multi-layered behavior based system that reduces the complexity of designing behavior based systems by allowing the designer to manipulate abstract behaviors directly. The use of abstract behaviors eeectively partition the system into a number of behavior based systems, each corresponding to a speciic complex goal or strategy. This report describes the architecture and, in particular, its application to Robocup, a simulated soccer competition being held at this year's International Joint Conference on Artiicial Intelligence, in Japan.