Learning to Negotiate Optimally in a Multi-agent Based Negotiation System for Web Service Selection

Abstract

Growing interest in the Web service based selection systems has given importance to the negotiation as it appears as a relevant tool to resolve discrepancy between negotiators, and to improve the selection process success rate. Better results may be obtained if agents are equipped with the real life negotiators’ capabilities of learning from past interactions. Proofs from observing the real human behaviors confirm that negotiators learning from interactions are more likely to get better offers. In this paper, we focus on the chicken game negotiation strategy. In such a game, a correlation between the first provider’s adopted strategy and the negotiation process success or failure exists. To model this correlation, we adopt the decision tree, a supervised learning technique in order to guide the clients’ choices of their best strategies. The experimental results show that including learning capability to the client’s behavior model leads to optimal results in terms of the client’s utility and the CPU time.