Analysis of Training Performance of Deep Reinforcement Learning on Single Board Computers

Single board computers (SBCs), such as the Raspberry Pi, are widely used in education, industry, and personal projects due to their affordability, compact size, and low power consumption. Recent research has focused on applying deep learning and deep reinforcement learning (DRL) to SBCs to enhance tasks like pattern recognition and autonomous control. However, due to SBCs' limited computational power, deep learning models are typically trained on high-performance machines and then optimized for deployment on SBCs. DRL, which involves continuous interaction with the environment, often requires SBCs to participate directly in the training process, posing challenges in terms of training time and communication latency. This paper analyzes the performance of training DRL algorithms on SBCs, particularly the Raspberry Pi 3B, 3B+ and 4B, evaluating the feasibility and overhead of different training architectures.