Towards Larger Receptive Field: Non-Local Reinforcement Learning
Changhao Zhao · 2022 International Conference on Big Data, Information and Computer Network (BDICN) · 2022
As a classical reinforcement learning algorithm, DQN combines traditional Q-learning with the deep neural network. When the input is an image, convolutional neural networks are frequently employed to extract features and the long-range relationships commonly occur in images. However, the receptive field of traditional convolutional neural networks is limited, failing to capture the long-range relationship. Hence, this work aims to address this problem and model the long-range relationship in the image by the proposed non-local reinforcement learning. Eventually, the performance of this method in the environment of angry Birds is evaluated, and the results demonstrate the superiority of this method.