Long-term collection of RL practice in the framework of Stable Baselines 3. I will make sure that each example is not the same, each has different processing details (except "hello world" enviroments: Cartpole and Lunarlander).

Lastest version of SB3: Installation.

Experimental features based on SB3: SB3 Contrib.

Anything about Gym enviroment: Gym Documentation.

Cartpole is a classical Gym enviroment (for details see here).

See codes here.

Episode length	Episode reward

Lunarlander is a classical Gym enviroment, aiming at rocket trajectory optimization (for details see here).

See codes here.

Episode length	Episode reward

Gridworld is modified from a custom Gym enviroment (for details see here), where an episode ends when the agent reaches the destination.

See codes here.

Random step	PPO agent

Maze is a 2d gridworld-like enviroment.

• Masking invalid actions greatly speeds up the training process of neural network.
• 1d observation is applied though it's natural to use a image-like observation. The reason is, SB3 CnnPolicy require image data to have a minimum size of 36x36. Lukily, a flattened observation still works well.

See codes here.

Episode length	Episode reward

Random step	PPO agent

Can the agent make correct decisions if the evaluation environment is different from the training environment? Applying data augment in training may help. In this case, the evaluation enviroment is a symmetric equivalent transformation of the training enviroment. An easy idea is generating equivalent batch datas based on the original batch data.

In order to achieve this goal, I build a custom callback and handle the rollout buffer data directly (See how to customize callback here) . We need to handle 9 kinds of data:

• Expand observations, actions and action_masks.
• Share episode_starts and rewards.
• Recompute values and log_probs.
• Recompute returns and advantages.

See codes here.

Training enviroment	Evaluation enviroment

Something interesting is, while the agent performed perfectly in the training environment, it got stuck in the evaluation environment. KL divergence explains this, large value of approx_kl seems to indicate that the network is not fully fitted:

total_steps = 8e4	total_steps = 15e4
approx_kl = 2.1942	approx_kl = 0.1253