A novel multi-agent reinforcement learning approach based on state adaptive weighting and exploration path sampling

In Multi-Agent Reinforcement Learning (MARL), the relationship between the current state and the reward is crucial to the model's performance. A good weight adaptation leads to a more reasonable reward, boosting the spatial exploration efficiency and stable control of all agents. This paper proposes State Adaptive Weighting (SAW) and Exploration Path Sampling (EPS), a powerful design to achieve state weight adaptation. SAW employs a dynamic state weighting network to prioritize informative and decision-critical states during training. It allows agents to focus on task-relevant regions of the state space and improves overall sample efficiency. EPS introduces a multi-faceted exploration strategy composed of three parts: a curiosity-driven module that uses prediction errors to generate intrinsic rewards, a path diversity tracker that encourages visits to novel states through visitation-based bonuses, and an adaptive noise mechanism that modulates exploration intensity based on environmental novelty. This paper conducts a series of experiments to demonstrate the improvements in learning speed and exploration quality. Our work offers a useful solution to address both effectiveness and stability in multi-agent systems. Abstract text.