Assess Your Opponent: A Bayesian Process for Preference Observation in Multi-attribute Negotiations

In an agent based multi-attribute negotiation, preferences are private knowledge. Once public, an agent’s preferences will be exploited by opposing parties trying to improve their own utility gains. Therefore, an agent will strictly hide its attribute weights from any competitor. The paper presents a method of learning an opposing party’s preferences by observing its behavior in a negotiation. The method includes a Bayesian learning process where the agent forms hypotheses based on its analysis of offers it receives. As a result, the agent improves estimation of the opponent’s preferences. Our experimental study on the performance of this negotiation strategy shows that not only does it improve the chances of the negotiation being successful at all (by a higher chance of finding the “area of agreement”), but it will also improve the individual success of the agent that applies it. The Bayesian agents perform superior to agents with static knowledge. In an all-Bayesian negotiation, there is a higher probability of a contract being reached than in a negotiation with static participation or in an all-static negotiation. Limited information about the opponent is one of the key limiting factors to successful automated negotiations. The idea of gaining knowledge that is intently kept hidden by another party is therefore a crucial basis for future automated negotiation systems that are trusted to succeed.