Reinforcement Learning of Two-Issue Negotiation Dialogue Policies

We use hand-crafted simulated negotiators (SNs) to train and evaluate dialogue policies for two-issue negotiation between two agents. These SNs differ in their goals and in the use of strong and weak arguments to persuade their counterparts. They may also make irrational moves, i.e., moves not consistent with their goals, to generate a variety of negotiation patterns. Different versions of these SNs interact with each other to generate corpora for Reinforcement Learning (RL) of argumentation dialogue policies for each of the two agents. We evaluate the learned policies against hand-crafted SNs similar to the ones used for training but with the modification that these SNs no longer make irrational moves and thus are harder to beat. The learned policies generally do as well as, or better than the hand-crafted SNs showing that RL can be successfully used for learning argumentation dialogue policies in twoissue negotiation scenarios.