The influence of different environments on reinforcement learning

The environment of the reinforcement learning is important to training learning. A good environment can speed up training, reduce costs, and better test new RL algorithms. Arcade Learning Environment, MinGrid and MinAtar are analyzed and summarize the characteristics of the three platforms and their respective advantages. Three reinforcement learning algorithms PPO, DQN and AC were tested in different environments. The conclusion is that, under certain constraints, simplifying the representation of the environment can speed up the learning of the agent and it can still show the characteristics of different algorithms very well. However, at the same time, inappropriate simplification may also lead to instability in training. This work investigates and demonstrates the importance of environments for new algorithm testing, as well as the strengths and weaknesses of existing environments.