STRATEGY MANAGEMENT IN A MULTI-AGENT SYSTEM USING NEURAL NETWORKS FOR INDUCTIVE AND EXPERIENCE-BASED LEARNING

Intelligent agents and multi-agent systems prove to be a promising paradigm for solving problems in a distributed, cooperative way. Neural networks are a classical solution for ensuring the learning ability of agents. In this paper, we analyse a multi-agent system where agents use different training algorithms and different topologies for their neural networks, which they use to solve classification and regression problems provided by a user. Out of the three training algorithms under investigation, Backpropagation, Quickprop and Rprop, the first demonstrates inferior performance to the other two when considered in isolation. However, by optimizing the strategy of accepting or rejecting tasks, Backpropagation agents succeed in outperforming the other types of agents in terms of the total utility gained. This strategy is learned also with a neural network, by processing the results of past experiences. Therefore, we show a way in which agents can use neural network models for both external purposes and internal ones.