A Data-Efficient Training Method for Deep Reinforcement Learning

Data inefficiency is one of the major challenges for deploying deep reinforcement learning algorithms widely in industry control fields, especially in regard to long-horizon sparse reward tasks. Even in a simulation-based environment, it is often prohibitive to take weeks to train an algorithm. In this study, a data-efficient training method is proposed in which a DQN is used as a base algorithm, and an elaborate curriculum is designed for the agent in the simulation scenario to accelerate the training process. In the early stage of the training process, the distribution of the initial state is set close to the goal so the agent can obtain an informative reward easily. As the training continues, the initial state distribution is set farther from the goal for the agent to explore more state space. Thus, the agent can obtain a reasonable policy through fewer interactions with the environment. To bridge the sim-to-real gap, the parameters for the output layer of the neural network for the value function are fine-tuned. An experiment on UAV maneuver control is conducted in the proposed training framework to verify the method. We demonstrate that data efficiency is different for the same data in different training stages.