Stable-Baselines3 Tutorial#

These tutorials show you how to use the Stable-Baselines3 (SB3) library to train agents in PettingZoo environments.

For environments with visual observation spaces, we use a CNN policy and perform pre-processing steps such as frame-stacking and resizing using SuperSuit.

For non-visual environments, we use MLP policies and do not perform any pre-processing steps.


Note: SB3 is designed for single-agent RL and does not directly support multi-agent algorithms or environments. These tutorials are only intended for demonstration purposes, to show how SB3 can be adapted to work with PettingZoo.


These tutorials utilize PPO with parameter sharing, allowing a single model to control all the agents in an environment.

For more information on PPO implementation details and multi-agent environments, see

For example, if there is a two-player game, we can create a vectorized environment that spawns two sub-environments. Then, the vectorized environment produces a batch of two observations, where the first observation is from player 1 and the second observation is from player 2. Next, the vectorized environment takes a batch of two actions and tells the game engine to let player 1 execute the first action and player 2 execute the second action. Consequently, PPO learns to control both player 1 and player 2 in this vectorized environment.

Stable-Baselines Overview#

Stable-Baselines3 (SB3) is a library providing reliable implementations of reinforcement learning algorithms in PyTorch. It provides a clean and simple interface, giving you access to off-the-shelf state-of-the-art model-free RL algorithms. It allows training of RL agents with only a few lines of code.

For more information, see the Stable-Baselines3 v1.0 Blog Post

SB3 Logo