Learning negotiation policies using IB3 and bayesian networks

Abstract

This paper presents an intelligent offer policy in a negotiation environment, in which each agent involved learns the preferences of its opponent in order to improve its own performance. Each agent must also be able to detect drifts in the opponent's preferences so as to quickly adjust itself to their new offer policy. For this purpose, two simple learning techniques were first evaluated: (i) based on instances (IB3) and (ii) based on Bayesian Networks. Additionally, as its known that in theory group learning produces better results than individual/single learning, the efficiency of IB3 and Bayesian classifier groups were also analyzed. Finally, each decision model was evaluated in moments of concept drift, being the drift gradual, moderate or abrupt. Results showed that both groups of classifiers were able to effectively detect drifts in the opponent's preferences. © 2010 Springer-Verlag Berlin Heidelberg.