Work

CG-MARL: Human-in-the-loop cognition-guided credit assignment for distributed multi-agent reinforcement learning

Year: 2026

Type: article

Abstract: Credit assignment remains a key challenge in cooperative multi-agent reinforcement learning (MARL), as existing methods infer agent contributions indirectly from global rewards through restrictive structural assumptions. We propose Cognition Guided MARL (CG-MARL), a preference-driven credit assignment framework that learns agent-wise contributions directly from trajectory-level human preferences. Human feedback is provided through pairwise trajectory comparisons with both global and agent-level annotations, capturing holistic judgments over cooperative behavior without requiring dense environment-defined rewards. We model such feedback using an agent-aware preference predictor under a Bradley–Terry formulation, which induces a preference-aligned cooperative value landscape. On this induced value, we perform Shapley value-based credit attribution to systematically decompose preference-aligned cooperation into agent-level contributions via counterfactual coalition reasoning. Crucially, human preference signals are further used to supervise and regularize the resulting Shapley credits, grounding credit attribution beyond architectural assumptions. Extensive experiments on SMAC demonstrate consistent performance gains and improved coordination, with ablations validating the effectiveness of cognition-guided credit assignment. Besides, real-world multi-agent experiments further confirm its generalization and robustness beyond simulation.

Source: Biomimetic Intelligence and Robotics

DOI: 10.1016/j.birob.2026.100362