Reinforcement Learning Planning for Resilient Multi-Agent Coordination via Distributed Non-Convex Optimization

The coordination of multi-agent systems in complex, dynamic environments presents a fundamental challenge in modern control theory and artificial intelligence. This paper introduces a novel framework that integrates deep reinforcement learning with distributed non-convex optimization to enhance resilience and efficiency in multi-agent coordination. Traditional methods often struggle with the non-convex nature of real-world constraints and the computational burden of centralized planning. Conversely, pure learning-based approaches frequently lack the stability guarantees required for safety-critical applications. We propose a hybrid architecture where a reinforcement learning agent acts as a high-level planner, generating warm-start initializations and heuristic constraints for a low-level distributed optimization solver. This synergy allows the system to navigate non-convex landscapes effectively while maintaining robustness against agent failures and communication disruptions. We evaluate our approach across various simulation environments, demonstrating superior convergence rates and fault tolerance compared to state-of-the-art baselines. The results corroborate that the proposed method effectively mitigates local minima entrapment and ensures continuous operation under significant network degradation.