juwu-19 / reinforcement_learning_algo_evaluation Goto Github PK

We study the theory and mechanism of two reinforcement learning algorithms, A2C and PPO2 and their application for multiple tasks e.g., Cart pole pendulum and lunar landing task. To do so, we build up a pipeline able to train and manage our models and environments.

From the experimental results obtained, we compute multiple reliability metrics to be able to lead a discussion about the performance evaluation of the algorithms.

From the above statistics tables, we can observe that A2C and PPO2 have the similar ranks of short-term and long-term risks across time but risks of PPO2 have less variance than that of A2C, which means that they have the similar average performance for the worst-case scenarios on different tasks. However, PPO2 is more stable than A2C in this case across different tasks and runs with various settings. We can also see that PPO2 has better reliability performance than A2C with respect to dispersion across time and risk across runs, which infers that A2C has the larger average variance of performances and the worse average final performance across the training runs across training runs under different tasks. But A2C has better dispersion across time than PPO2, which means A2C has the smaller distribution for differential of training curves.