Sampling Rate Decay in Hindsight Experience Replay for Robot Control.

Training agents via deep reinforcement learning with sparse rewards for robotic control tasks in vast state space are a big challenge, due to the rareness of successful experience. To solve this problem, recent breakthrough methods, the hindsight experience replay (HER) and aggressive rewards to counter bias in HER (ARCHER), use unsuccessful experiences and consider them as successful experiences achieving different goals, for example, hindsight experiences. According to these methods, hindsight experience is used at a fixed sampling rate during training. However, this usage of hindsight experience introduces bias, due to a distinct optimal policy, and does not allow the hindsight experience to take variable importance at different stages of training. In this article, we investigate the impact of a variable sampling rate, representing the variable rate of hindsight experience, on training performance and propose a sampling rate decay strategy that decreases the number of hindsight experiences as training proceeds. The proposed method is validated with three robotic control tasks included in the OpenAI Gym suite. The experimental results demonstrate that the proposed method achieves improved training performance and increased convergence speed over the HER and ARCHER with two of the three tasks and comparable training performance and convergence speed with the other one.

Two-stage training algorithm for AI robot soccer.

In multi-agent reinforcement learning, the cooperative learning behavior of agents is very important. In the field of heterogeneous multi-agent reinforcement learning, cooperative behavior among different types of agents in a group is pursued. Learning a joint-action set during centralized training is an attractive way to obtain such cooperative behavior; however, this method brings limited learning performance with heterogeneous agents. To improve the learning performance of heterogeneous agents during centralized training, two-stage heterogeneous centralized training which allows the training of multiple roles of heterogeneous agents is proposed. During training, two training processes are conducted in a series. One of the two stages is to attempt training each agent according to its role, aiming at the maximization of individual role rewards. The other is for training the agents as a whole to make them learn cooperative behaviors while attempting to maximize shared collective rewards, e.g., team rewards. Because these two training processes are conducted in a series in every time step, agents can learn how to maximize role rewards and team rewards simultaneously. The proposed method is applied to 5 versus 5 AI robot soccer for validation. The experiments are performed in a robot soccer environment using Webots robot simulation software. Simulation results show that the proposed method can train the robots of the robot soccer team effectively, achieving higher role rewards and higher team rewards as compared to other three approaches that can be used to solve problems of training cooperative multi-agent. Quantitatively, a team trained by the proposed method improves the score concede rate by 5% to 30% when compared to teams trained with the other approaches in matches against evaluation teams.