Developing agent models with a neural reinforcement technique

Abstract

A reinforcement training procedure was developed for sequential back-propagation networks and applied in several studies demonstrating interaction between agents in multiple-agent networks. In the first study, a network was trained to predict the next position of an agent which was moving in a complex pattern around the corners of a square. The network quickly learned to predict the position without error. In particular, the network may be said to have developed an agent or user model of the moving agent. In two additional studies, a joint contingency was applied to two agents and limited cooperation was developed between them. Overall, the results provide support for the application of neural networks in distributed AI (artificial intelligence).<>