LDQN: A Lightweight Deep Reinforcement Learning Model

Abstract

Deep Reinforcement Learning (DRL) integrates Reinforcement Learning (RL) with Deep Learning and has made remarkable strides in several domains, such as gaming, robotic control, autonomous driving, and natural language processing. Yet, in real-world applications, many deep learning models utilizing Deep Q-learning Networks (DQN) continue to encounter issues like slow training times and significant computational expenses. To address these challenges, this paper proposes a lightweight deep reinforcement learning model based on DQN, called Lightweight Deep Q-learning Network (LDQN). LDQN incorporates design concepts from the lightweight convolutional neural network MobileNet, adapting the first two standard convolutional layers into depthwise separable convolutions to minimize computational costs and model size. Additionally, the network employs the efficient and high-performance H-Swish activation function to boost image feature extraction abilities while minimizing computational overhead. A series of experimental results demonstrate that the proposed LDQN model not only reduces computational costs and model parameters but also delivers exceptional performance in Atari games.