From Semantic Decomposition to Human-Guided Control: Interpretable Multi-Agent Reinforcement Learning with Gated Expert

Multi-Agent Reinforcement Learning (MARL) has shown remarkable performance in collaborative perception and decision-making tasks. However, its highly opaque decisionmaking architecture hinders human understanding, intervention, and correction of policy behavior, thereby limiting its application in safety-critical scenarios. This paper proposes an interpretable MARL framework based on Mixture-of-Experts (MoE) and human preference-guided gating, aiming to achieve interpretable decisions and controllable human-agent collaboration without compromising performance. Using a multi-agent target tracking task as the case study, the framework first pretrains a set of semantic experts and forms a frozen library of behavioral primitives. A learnable gating network is then introduced to dynamically fuse the experts' outputs through weighted aggregation, where the gating weights directly quantify the relative contribution of different semantic behaviors to the current decision, thereby providing intrinsic interpretability at the structural level. To better align the policy with human intent, we propose a human preference-guided gating adaptation mechanism that models human feedback as a form of structured supervision over the expert weight distribution, rather than direct modification of the reward function or policy parameters. This mechanism enables humans to adjust decision-making intent in a controllable manner without destabilizing the underlying behavioral primitives. Experimental results demonstrate that the proposed method significantly enhances training efficiency and policy stability in multi-agent target tracking tasks, achieving semantically consistent and intervenable decision-making behavior. This work offers a novel pathway for human-agent collaboration in interpretable reinforcement learning.