Explainable AI for Multi-Agent Control Problem

Abstract: This report presents research on the application of policy explanation techniques in the context of coordinated reinforcement learning (CRL) for mobile network optimization. The goal was to improve the interpretability and comprehensibility of decision-making processes in multi-agent environments, with a particular focus on the Remote Antenna Tilt (RET) problem. The task has been formulated as providing insight into the extension of policy explanation in a cooperative multi-agent reinforcement learning (MARL) environment, evaluating its applicability to a network use case, and revealing interactions between agents in such a setting. The results contribute to a better understanding of decision-making processes, dynamics of coordination, and aspects of interpretation in complex multi-agent systems, in particular in the context of mobile network optimization. This research is motivated by the need for transparency, accountability, and trust in AI-driven decision-making processes, especially in critical applications such as mobile networks. The study aimed to bridge the gap between the confusing behavior of many agents and the need for human-understandable explanations. The approach involved training a CRL agent and using a policy explanation method to generate explanations based on the observations and actions taken by the agent. The outcomes demonstrated the effectiveness of the policy explanation method in providing clear and robust interpretations in both single-agent and multi-agent environments. Furthermore, analysis of CRL Q-value functions revealed consistent patterns in some preferences and avoidance of certain interactions with neighboring agents. This insight allows for a better understanding of coordination dynamics in mobile network optimization. In conclusion, this study demonstrates the successful application of policy explanation methods in CRL to optimize mobile networks. Combining CRL and policy explanations improves the interpretation of agent behavior and increases accountability. The study contributes to the expansion of the explainable AI field and lays the foundation for future research on the optimization of complex multi-agent systems.