PPO LunarLander-v3
PPO agent trained for 1,000,000 timesteps with Stable-Baselines3 on 16 parallel Gymnasium LunarLander-v3 environments. The Hugging Face Deep RL Course notebook still refers to LunarLander-v2; that ID is no longer available in current Gymnasium releases, so this run uses the maintained v3 equivalent.
Evaluation
The official Unit 1 scoring rule is mean_reward - std_reward over 10 deterministic episodes. This run scored 242.15 (mean 263.61, standard deviation 21.46), above the course's 200-point threshold.
| Environment | Episodes | Mean | Std. dev. | Mean − std. | Threshold |
|---|---|---|---|---|---|
| LunarLander-v3 | 10 | 263.61 | 21.46 | 242.15 | 200 |
The evaluation ran in a separate monitored environment. results.json records the exact values; replay.mp4 shows one deterministic episode. This checkpoint uses the course PPO setup (n_steps=1024, batch_size=64, n_epochs=4, gamma=0.999, gae_lambda=0.98, ent_coef=0.01).
Use
from stable_baselines3 import PPO
agent = PPO.load("ppo-LunarLander-v3.zip")
Model weights are saved as ppo-LunarLander-v3.zip.
- Downloads last month
- 20