关键词: 深度强化学习, 样本效率, 正则化层, 强化学习

Abstract: With the advancement of deep neural networks, combining them with reinforcement learning has led to deep reinforcement learning, enhancing decision-making in complex environments. However, these algorithms often suffer from low sample efficiency, needing extensive data interaction to learn effective strategies. Through experimental analysis, we have investigated the underlying causes of low sample efficiency and, based on these findings, we propose a Mixed Scaling Smoothing Regularization Layer that directly operates on the feature representations produced by the convolutional encoder. By randomly mixing different local regions of the same feature map to generate new feature representations, this method introduces multi-scale variations in the latent space, effectively preventing overfitting in the early stages of training and improving sample efficiency. Experimental results in various simulated environments show that our method demonstrates superiority over existing state-of-the-art methods, with an 18.13% improvement in sample efficiency. This method is simple to implement, highly generalizable, and provides a new approach for improving sample efficiency in deep reinforcement learning.

Key words: Deep Reinforcement Learning, Sample Efficiency, Regularization Layer, Reinforcement Learning