Trust-Based Multiagent Credit Assignment (TMCA)

Abstract In Multiagent Reinforcement Learning (MARL), a single scalar reinforcement signal is the sole reliable feedback that members of a team of learning agents can receive from the environment around them. Hence, the distribution of the environmental feedback signal among learning agents, also known as the “ Multiagent Credit Assignment ” (MCA), is among the most challenging problems in MARL. In this paper, the authors propose an extended solution to the problem of MCA. In the proposed method, called “ Trust-based Multiagent Credit Assignment ” (TMCA), a trust and reputation based model is utilized to evaluate the trustworthiness of the learning agents. Unlike the existing methods, TMCA not only qualifies to benefit from the knowledge and expertise of the sole target agent (the agent for which the credit is being evaluated), but also from the knowledge and expertise of the whole as a team. To evaluate this method, the effect of different task types (e. g. AND vs. OR) are studied. Our simulations show the superiority of the proposed method in comparison to the prior investigated methods even in noisy environments, despite a reduction (caused by the noise) in the performance.

Authors

• Samira Nazari
• Mohammad Ebrahim Shiri

Keywords

No keywords are indexed for this paper.