| In many agent-based systems, neural networks are used to implement the mechanisms by which agents learn and classify patterns. In such systems, agents are typically embedded with a connectionist-based model during implementation with no means for overriding this model with new models during run-time. Thus, the way in which an agent learns or classifies patterns often remains static during the life of the agent. This is due, in part, to the lack of research in communicating subsymbolic knowledge (i.e. knowledge encoded in a neural network) between agents. As a result, an agent may be unable to adapt its learning or classification behavior to changes in its environment (especially if such changes are highly dynamic), and thus experience a degradation in performance. By allowing agents to override existing neural network models for learning or pattern classification with new models in real-time, however, they may potentially maintain or increase their performance in a dynamic environment.; In this dissertation, we present a framework for communicating neural network knowledge between agents in order to modify an agents learning or pattern classification. behavior. This framework is comprised of a specification for modeling neural network knowledge, a protocol for communicating neural network knowledge between agents, and a specification of a multi-agent architecture for managing and using neural network knowledge among a set of distributed agents. This framework is applied to a simulated aerial reconnaissance system in order to show how the communication of neural network knowledge can help maintain the performance of agents tasked with recognizing images of mobile military objects. |
