Merging Neural Networks in a Multi-Agent System
January 30, 2000
The Connectionist Model Transfer (CMT) framework was proposed to allow an agent the ability to instantiate and execute neural networks received from other agents in order to maintain its learning or classification performance in a dynamic environment. One limitation of the CMT framework was that it assumed the existence of at least one network in the system that embodied the classification mappings required to exhibit satisfactory classification performance for each situation that an agent might encounter. In some cases, however, an agent required a single network that embodied classification mappings of several existing networks in order to maintain its performance. In such cases, a new network was manually trained and submitted to the system. In this paper, we extend the CMT framework for facilitating the dynamic merging of distributed networks. This extension allows agents to utilize new, automatically trained networks that embody the classification mappings of several distributed and dynamically selected networks. We apply this extension to a simulated aerial reconnaissance system in order to show how the merging of neural networks can help maintain the performance of agents tasked with recognizing images of mobile military objects.
Downloads: 26 downloads