Enhancing multi-agent communication through credibility and reward-based optimisation
Yi-Yu Lin, Xiao‐Jun Zeng · International Journal of General Systems · 2025
This paper presents the first formal formulation of the multi-agent communication problem in a partially observable environment, marking a crucial advancement in optimising communication by explicitly accounting for the associated costs and benefits. Building on this theoretical foundation, the paper introduces Credibility-Aware Reward-Optimised Communication (CAROC), a novel framework designed to address inefficiencies in existing methods, such as redundant information exchange. CAROC enables agents to communicate selectively, prioritising interactions that are both credible and valuable. The framework integrates credibility metrics, reward prediction, and informative entropy regularisation to enhance the efficiency and effectiveness of agent interactions. Credibility metrics rigorously assess the trustworthiness of peers, ensuring communication focuses on agents who consistently provide high-quality information. Reward prediction evaluates the potential impact on system objectives, guiding decisions that maximise overall performance. Informative entropy regularisation balances exploration and exploitation, allowing agents to adapt to dynamic environments while making informed, context-sensitive decisions. To thoroughly validate the effectiveness of CAROC, extensive testing was conducted in various environments, including the newly developed Territory Guardian and Invasion environment, which introduces complex, dynamic challenges. The results demonstrate that CAROC consistently outperforms existing methods in both communication efficiency and task performance, establishing a new benchmark for future research in multi-agent systems.