Chris1/ppo_lstm-BipedalWalker-v3-1

RecurrentPPO Agent playing BipedalWalker-v3

This is a trained model of a RecurrentPPO agent playing BipedalWalker-v3 using the stable-baselines3 library and the RL Zoo.

The RL Zoo is a training framework for Stable Baselines3 reinforcement learning agents, with hyperparameter optimization and pre-trained agents included.

Usage (with SB3 RL Zoo)

RL Zoo: https://github.com/DLR-RM/rl-baselines3-zoo
SB3: https://github.com/DLR-RM/stable-baselines3
SB3 Contrib: https://github.com/Stable-Baselines-Team/stable-baselines3-contrib

# Download model and save it into the logs/ folder
python -m utils.load_from_hub --algo ppo_lstm --env BipedalWalker-v3 -orga Chris1 -f logs/
python enjoy.py --algo ppo_lstm --env BipedalWalker-v3  -f logs/

Training (with the RL Zoo)

python train.py --algo ppo_lstm --env BipedalWalker-v3 -f logs/
# Upload the model and generate video (when possible)
python -m utils.push_to_hub --algo ppo_lstm --env BipedalWalker-v3 -f logs/ -orga Chris1

Hyperparameters

OrderedDict([('batch_size', 256),
             ('clip_range', 0.18),
             ('ent_coef', 0.0),
             ('gae_lambda', 0.95),
             ('gamma', 0.999),
             ('learning_rate', 0.0003),
             ('n_envs', 32),
             ('n_epochs', 10),
             ('n_steps', 256),
             ('n_timesteps', 5000000.0),
             ('normalize', True),
             ('policy', 'MlpLstmPolicy'),
             ('policy_kwargs',
              'dict( ortho_init=False, activation_fn=nn.ReLU, '
              'lstm_hidden_size=64, enable_critic_lstm=True, '
              'net_arch=[dict(pi=[64], vf=[64])] )'),
             ('normalize_kwargs', {'norm_obs': True, 'norm_reward': False})])