Efficient DDPG via the Self-Supervised Method

In this paper, embedded with self-supervised learning network, an efficient DDPG(Deep Deterministic Policy Gradient) RL algorithm is investigated. With more essential characteristics of observing data included, the inputs of actor network and critic network of DDPG are replaced by the high-dimensional outputs from feature extracting layers and forward network respectively. Additionally, the parameters of these auxiliary layers are optimized with a self-supervised method by minimizing predicting errors, and thus both optimizing progresses can run parallelly and simultaneously. Lastly, an antagonistic air-fight simulation with a novel customized training index is introduced to perform the effectiveness and rising efficiency of our self-supervised DDPG RL algorithm.