COMB: Scalable Concession-Driven Opponent Models Using Bayesian Learning for Preference Learning in Bilateral Multi-Issue Automated Negotiation

Learning an opponent’s preferences in bilateral multi-issue automated negotiations can lead to more favorable outcomes. However, existing opponent models can fail in negotiation contexts when their assumptions about opponent behaviors differ from actual behavior patterns. Although integrating broader behavioral assumptions into these models could be beneficial, it poses a challenge because the models are designed with specific assumptions. Therefore, this study proposes an adaptable opponent model that integrates a general behavioral assumption. Specifically, the proposed model uses Bayesian learning (BL), which can apply various behavioral assumptions by considering the opponent’s entire bidding sequence. However, this BL model is computationally infeasible for multi-issue negotiations. Hence, current BL models often impose constraints on their hypothesis space, but these constraints about the utility function’s shape significantly sacrifice accuracy. This study presents a novel scalable BL model that relaxes these constraints to improve accuracy while maintaining linear time complexity by separately learning each parameter of a utility function. Furthermore, we introduce a general assumption that the opponent’s bidding strategy follows a concession-based pattern to enhance adaptability to various negotiation contexts. We explore three likelihood function options to implement this assumption effectively. By incorporating these options into the proposed scalable model, we develop three scalable concession-driven opponent models using Bayesian learning (COMB). Experiments across 45 negotiation domains using 15 basic agents and 15 finalists from the automated negotiating agents competition demonstrate the proposed scalable model’s higher accuracy than existing scalable models. COMB models show higher adaptability to various negotiation contexts than state-of-the-art models.