Beyond Simple Argumentation: Image Learnable Transformation For Efficient Reinforcement Learning

Deep reinforcement learning(RL) suffers from high complexity space exploration and a substantial amount of time calculation problem, when it handles tasks which need to process high-dimensional interaction image datasets. Many data augmentation algorithms alleviated these problems by applying basic image transformation methods. However, they did not quickly explore the effective policy space and requires manual design for different tasks. To alleviate this issue and improve sample-efficiency of RL algorithms, we propose a new data augmentation model called learnable image transformation RL, which contains affine transformation of autonomous learning and image embedding based on attention mechanism. It reduces half time of previous baseline and improve the accuracy and sample-efficiency in reinforcement learning tasks. We also show that the generation of countermeasure samples with noisy and image transformation improve the ability of network by extracting better semantical representations and the robustness of the model. Experimental results show the proposed method’s usefulness.