Approximating an auction mechanism by multi-issue negotiation

The main question addressed in this paper is whether a theoretical outcome determined by an auction mechanism can be reasonably approximated by negotiation among agents in order to drop some of the unrealistic constraints or assumptions presupposed by the mechanism. In particular, we are interested in whether the assumption that a buyer publicly announces her preferences in order to guarantee perfect knowledge of these preferences can be dropped if a negotiating agent is used that can learn preferences. We show how to setup a multiplayer multi-issue negotiation process where preferences are learned, and we investigate how the results of this process relate to the theoretical result of holding an auction in the case of complete knowledge about the preferences of the buyer. Experiments show that the outcomes obtained by negotiating agents that learn opponent preferences approximate the outcome predicted by the mechanism. It thus follows that the assumption of perfect knowledge about buyer preferences can be removed when players are equipped with proper learning capabilities. We also investigate whether the procedure dictated by the mechanism can be further relaxed but in that case experiments indicate that more complex considerations about the market need to be taken into account.