PPO Agent Playing LunarLander-v3
This repository contains a PPO agent trained on LunarLander-v3 using PyTorch and Gymnasium.
Evaluation
The final policy was evaluated deterministically over 10 episodes.
- Mean reward: 226.17
- Standard deviation: 19.13
- Minimum reward: 202.36
- Maximum reward: 260.80
- Mean episode length: 516.90
Replay
replay.mp4 contains one deterministic evaluation episode.
- Replay reward: 213.39
- Replay episode length: 481
- Recorded frames: 482
Hyperparameters
- exp_name:
ppo - seed:
1 - torch_deterministic:
True - cuda:
True - capture_video:
False - track:
False - wandb_project_name:
cleanRL - wandb_entity:
None - env_id:
LunarLander-v3 - total_timesteps:
1000000 - learning_rate:
0.00025 - num_envs:
16 - num_steps:
1024 - anneal_lr:
True - gae:
True - gamma:
0.999 - gae_lambda:
0.98 - num_minibatches:
256 - update_epochs:
4 - norm_adv:
True - clip_coef:
0.2 - clip_vloss:
True - ent_coef:
0.01 - vf_coef:
0.5 - max_grad_norm:
0.5 - target_kl:
None - repo_id:
roshana1s/ppo-LunarLander-v3-2 - push_to_hub:
True - eval_episodes:
10 - video_fps:
30 - batch_size:
16384 - minibatch_size:
64
Repository Files
model.pt— checkpoint containing the trained state dictionary and model informationresults.json— final evaluation metricsreplay.mp4— deterministic replayhyperparameters.json— training configurationppo.py— training implementationlogs/— TensorBoard logs
Usage
This model uses the same actor/critic architecture defined in ppo.py.
Evaluation results
- Mean Reward on LunarLander-v3self-reported226.17 +/- 19.13