Stable Baselines3

Basic usage

1. Create environment: gym.make() / make_vec_env()
2. Define the model, instantiate this model: model = PPO("MlpPolicy)
3. Train the agent by define the traning timesteps: model.learn(total_timesteps=int(n))
4. Wrap the environment in a Monitor: eval_env = Monitor(env)
5. Evaluate policy: evaluate_policy(model, eval_env, m_eval_episodes, deterministics)

Functions

Training

• make_vec_env("EnvName", n_envs=n): Create a vectorized environment of n environment, a method for stacking multiple independent environment into a single environment.
• model.save("filename"): Specify file name for model and save the model to file

Evaluation

• evaluate_policy(model, env, n_eval_episodes, deterministics): Run the policy foe n_eval_episodes and outputs the average return per episode.
  • deterministics: Whether to use deterministic or stochastic actions
• Monitor: It is used to know the episode reward, length, time and other data

Load model

• methods.load(checkpoint, custom_objects=custome_objects, print_system_info=True): models from huggingface hub
  • checkpoint: huggingface_sb3.load_from_hub()
  • custom_objects: Dictionary of objects to replace upon loading, such as learning rate