Adaptation and learning in software agents

Current software agent systems do only what they are programmed to do. Their capabilities extend very little beyond that. Our thesis is that software agents need to be driven by purposes in order to evolve the ability to adapt and learn. We have been developing agent systems whose intrinsic elements are purpose structures. Our goal is to extend both the theoretical foundation and the practical application of adaptive learning behavior in agent systems. We seek to work toward the development of systems that ultimately evolve learning as an emerging phenomenon. This paper describes our recent work under the DARPA TASK program in which we explored the use of purpose-based control structures as the conceptual framework underlying agent system behavior in a variety of complex tasks including adaptive traffic light control and adaptive UAV surveillance in heterogeneous dynamic operational environments. Our work in the latter area is presented here.