EnTADRL: Deep Reinforcement Learning System-on-Chip for End-to-End Training Acceleration

For sequential decision-making problems such as gaming agents and autonomous driving, deep reinforcement learning (DRL) achieves human-level performance by training agents through interactions with the environment [1]. In detail, an agent represented in a deep neural network (DNN) provides action to the environment and the environment offers rewards and states to the agent according to action. By repeating this process, agents are trained to maximize rewards and demonstrate appropriate actions based on the environment. In real-world, environments are dynamic because of new users, new risks and changeable elements, which means the trained agent may fail to demonstrate appropriate actions. Therefore, continuous and efficient DRL training accelerator is necessary for real-world applications.