Reinforcement Learning of Multi-Issue Negotiation Dialogue Policies

We use reinforcement learning (RL) to learn a multi-issue negotiation dialogue policy. For training and evaluation, we build a hand-crafted agenda-based policy, which serves as the negotiation partner of the RL policy. Both the agendabased and the RL policies are designed to work for a large variety of negotiation settings, and perform well against negotiation partners whose behavior has not been observed before. We evaluate the two models by having them negotiate against each other under various settings. The learned model consistently outperforms the agenda-based model. We also ask human raters to rate negotiation transcripts between the RL policy and the agenda-based policy, regarding the rationality of the two negotiators. The RL policy is perceived as more rational than the agenda-based policy.