Deep Reinforcement Learning Guided by Game-Theoretic Structure for Multi-Agent Resource Allocation and Scheduling

This study addresses the challenge of balancing individual rationality and system-level efficiency in multi-agent collaborative decision making under complex and dynamic environments. It proposes a multi-agent collaborative optimization method that integrates game theory with deep reinforcement learning. The interaction process among agents is formulated as a Markov game. Strategy coupling and environmental dynamics are described within a unified sequential decision framework. By introducing a cooperative potential function and consistency constraints, system-level objectives are incorporated into individual rewards in a controlled manner. This guides the emergence of stable coordinated behavior under decentralized execution. In terms of learning design, a centralized training and decentralized execution paradigm is adopted. Centralized value evaluation is used to alleviate non-stationarity in multi-agent learning, while preserving autonomy and scalability of individual decisions. The framework can learn joint policies that balance immediate responsiveness and long-term coordination without relying on explicit environment models. As a result, multi-agent systems exhibit more consistent behavior patterns in tasks such as resource allocation, route selection, and scheduling. Comparative results against representative methods show that the proposed approach achieves superior overall performance on system-level metrics, including service efficiency, operational cost, and resource utilization. This confirms the effectiveness and practical value of deep reinforcement learning guided by a game-theoretic structure for multi-agent collaborative optimization.