Using Transfer Learning to Model Unknown Opponents in Automated Negotiations

摘要

Modeling unknown opponents is known as a key factor for the efficiency of automated negotiations. The learning processes are however challenging because of (1) the indirect way the target function can be observed, and (2) the limited amount of experience available to learn from an unknown opponent at a single session. To address these difficulties we propose to adopt two approaches from transfer learning. Both approaches transfer knowledge from previous tasks to the current negotiation of an agent to aid learn the latent behavior model of an opposing agent. The first approach achieves knowledge transfer by weighting the encounter offers of previous tasks and the ongoing task, while the second one by weighting the models learnt from the previous negotiation tasks and the model learnt from the current negotiation session. Extensive experimental results show the applicability and effectiveness of both approaches. Moreover, the robustness of the proposed approaches is evaluated using empirical game theoretic analysis.