Gaussian Trust Region Policy Optimization (GTR) Enhances Reinforcement Learning in Non-Stationary Environments
This paper introduces Gaussian Trust Region Policy Optimization (GTR), a novel approach that addresses the limitations of Proximal Policy Optimization (PPO) in non-stationary environments. GTR utilizes a Gaussian-kernel-shaped trust region to create bounded, non-monotonic constraints, allowing for greater policy deviations when beneficial for adaptation. Additionally, a Mixture Gaussian Anchor reduces variance by adapting to recent trajectories, enhancing overall robustness. GTR has shown strong performance across various domains, suggesting geometry-aware trust region design is a promising direction for reinforcement learning in complex, dynamic settings.