Implementing Deep Reinforcement Learning Algorithms on the Tennis Environment

In this paper, different deep reinforcement learning algorithms including cross entropy, DQN, and DDPG are applied to the tennis environment (from Unity ML-Agents). This environment includes 2 agents, each observing 8 states corresponding to the position and the velocity of the racket and the ball. For each agent, two continuous actions are available corresponding to the movement toward or away from the net and jumping. The purpose is to make the game last as long as possible, and agents collaborate to prevent the ball from reaching the ground. The number of episodes and time required for training and reliability are the important factors for evaluating and comparing the performance of different algorithms. DDPG algorithm was trained in fewer episodes, but the time required for training was longer due to the complexity.