Multi-Strategy Distillation Based on CTCE and CEDE

This study proposes a multi-agent reinforcement learning algorithm that integrates the Centralized Training with Centralized Execution (CTCE) and Centralized Training with Decentralized Execution (CTDE) frameworks to enhance distributed agents' collaboration efficiency and training performance through customized knowledge distillation strategies. To address the challenges of partial observability and credit assignment in multi - agent systems, we design a distillation framework that employs the Multi - Agent Transformer (MAT) as the teacher model and QMIX as the student model, and incorporate dynamic intervention mechanisms (“Early Guidance”, “On Demand Assistance”, and “Key - Point Emphasis” strategies) to optimize the frequency and intensity of teacher guidance during critical decision - making phases. Additionally, we innovatively introduce distributional reinforcement learning methods. We measure the student model's confidence level through the multipeak concentration index (MCI) to facilitate effective teacher intervention. Experimental results demonstrate that our method significantly improves the success rate and reward values of student models in StarCraft II mini - games, validating the effectiveness of dynamic distillation strategies in balancing global collaboration with local adaptability. This approach provides novel insights for multi - agent coordination in complex tasks.