Multi-agent modelling using intelligent agents in competitive games

Abstract

Summary Multi-Agent systems typically utilise simple, predictable agents. The usage of such agents in large systems allows for complexity to be achieved through the interaction of these agents. It is feasible, however, to utilise intelligent agents in smaller systems, allowing for more agent complexity and hence a higher degree of realism in the multi-agent model. By utilising the TD( ) Algorithm to train feedforward neural networks, intelligent agents were successfully trained within the reinforcement learning paradigm. A methodology for stabilising this typically unstable neural network training was found through first looking at the relatively simple problem of Tic-Tac-Toe. Once a stable training methodology was arrived at, the more complex task of tackling a multi-player, multi-stage card-game was tackled. The results illustrated that a variety of scenarios can be realistically investigated through the multi-agent model, allowing for solving of situations and better understanding of the game itself. Yet more startling, owing to the agent’s design, the agents learned on their own to bluff, giving much greater insight into the nature of bluffing in such games that lend themselves to the act.