The Prediction of Partners’ Behaviors in Self-Interested Agents

Prediction partners’ behaviors in negotiation has been an active research direction in recent years. By employing the estimation results, agents can modify their negotiation strategies in order to achieve an agreement much quicker or to gain higher benefits. Some of estimation strategies have been proposed by researchers, and most of them are based on machine learning mechanisms. However the machine learning based approach may not work well in some open and dynamic domains for the reasons of (1) lacking of sufficient data to train the system, and (2) requesting plenty of resources in each training process. Furthermore, because the estimation results may have errors, so single result maybe not accurate and practical enough in most situations. In order to address these issues mentioned above, we propose a quadratic regression analysis approach to predict partners’ behaviors in this paper. The proposed approach is based only on the history of the offers during the current negotiation and does not require any training process in advance. This approach can estimate an interval of behaviors according to an accuracy requirement. The experimental results illustrate that by employing the proposed mechanism, agents can gain more accurate estimation results on partners’ behaviors by comparing with other two estimation functions.