Research on Policy Stability of Reinforcement Learning in Complex Dynamic Decision-Making Environments

Abstract

The policy stability of reinforcement learning is critical to its practical application when addressing sequential decision-making problems in complex dynamic environments. This paper systematically analyzes the impact mechanisms of environmental dynamics, algorithm sensitivity, and state-space complexity on policy stability, and it constructs an enhanced framework that integrates constrained optimization, probabilistic environment modeling, meta-learning, and multi-agent coordination. The research provides systematic theoretical support and methodological pathways for enhancing the robustness and adaptability of reinforcement learning in non-stationary environments.