Predicting people's bidding behavior in negotiation

This paper presents a statistical learning approach to predicting people's bidding behavior in negotiation. Our study consists of multiple 2-player negotiation scenarios where bids of multi-valued goods can be accepted or rejected. The bidding task is formalized as a selection process in which a proposer player chooses a single bid to offer to a responder player from a set of candidate proposals. Each candidate is associated with features that affect whether or not it is the chosen bid. These features represent social factors that affect people's play. We present and compare several algorithms for predicting the chosen bid and for learning a model from data. Data collection and evaluation of these algorithms is performed on both human and synthetic data sets. Results on both data sets show that an algorithm that reasons about dependencies between the features of candidate proposals is significantly more successful than an algorithm which assumes that candidates are independent. In the synthetic data set, this algorithm achieved near optimal performance. We also study the problem of inferring the features of a proposal given the fact that it was the chosen bid. A baseline importance sampling algorithm is first presented, and then compared with several approximations that attain much better performance.