Effective Adaptive Strategy Selection Using Extended Fine-Tuning and CNN-Based Surrogate Model in Repeated-Encounter Bilateral Automated Negotiation

In this study, we tackle the challenges of repeated-encounter bilateral automated negotiation (RBAN) by introducing a fine-tuning approach to improve surrogate model-based strategy selection methods. To make this fine-tuning process more effective, we consider two policies: firstly, the aggregating policy, which reduces training parameters, and secondly, the complete-preference integration policy, which improves the use of past negotiation information. Moreover, we propose a convolutional neural network (CNN)-based surrogate model (CSM) to predict a strategy’s performance by analyzing the distribution of utility values in negotiation outcomes. We evaluate the prediction capability of the CSM and the impact of the two policies on both CSM and fine-tuning approach. The experimental results demonstrate that the CSM outperforms existing expert-feature-based opponent models in terms of prediction accuracy. Ablation studies in RBANs reveal the superiority of combining the fine-tuning approach with both policies over using fine-tuning alone or with just one policy. Ablation studies in independent negotiations show that applying either or both policies on the CSM also improves the CSM’s performance in independent negotiations, although the two policies are not initially designed for this context.