The LSTM-PER-TD3 Algorithm for Deep Reinforcement Learning in Continuous Control Tasks

Abstract

Deep reinforcement learning has shown great potential in the field of robot control, but it still faces challenges in continuous control tasks. Traditional reinforcement learning algorithms perform poorly when dealing with high-dimensional state spaces and continuous action spaces, resulting in low learning efficiency and degraded performance. To solve this issue, we propose a novel reinforcement learning algorithm named LSTM-PER-TD3 (LPT3), which combines Long Short-Term Memory networks (LSTM), prioritized experience replay (PER), and Twin Delayed Deep Deterministic Policy Gradient algorithm (TD3). The LPT3 algorithm utilizes LSTM networks to process sequential state information and combines PER and TD3 methods to achieve efficient continuous control. It is capable of learning accurate policies in high-dimensional state spaces, improving learning efficiency and performance through experience replay and policy evaluation. We evaluate the LPT3 algorithm on several standard continuous control tasks and compare it with traditional reinforcement learning algorithms. Experimental results demonstrate the superiority of LPT3 in terms of learning efficiency and performance. Compared to traditional algorithms, LPT3 converges faster and achieves higher control accuracy, effectively addressing the challenges in continuous control tasks.