muhrivandysetiawan/ppo-LunarLander-v3

🚀 PPO Agent — LunarLander-v3

Trained with Stable-Baselines3 PPO on LunarLander-v3. Hyperparameters optimized via Optuna TPE — 20 trials. Part of a 12-project Deep RL portfolio.

📊 Performance

Version	Score	Notes
Baseline	210.15	Default SB3 params
Optuna v1	259.15	20-trial HPO sweep
Improvement	+49.00	clip_range ↓, more stable

Target: ≥ 200 — cleared by 59.1 points.

⚡ Reproduce

from huggingface_sb3 import load_from_hub
from stable_baselines3 import PPO
import gymnasium as gym

# Load model
checkpoint = load_from_hub(
    repo_id  = "muhrivandysetiawan/ppo-LunarLander-v3",
    filename = "ppo-LunarLander-v3.zip",
)
model = PPO.load(checkpoint)

# Run episode
env = gym.make("LunarLander-v3", render_mode="human")
obs, _ = env.reset()
done   = False
while not done:
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, _ = env.step(action)
    done = terminated or truncated
env.close()

🏗️ Architecture

OOP class-based pipeline — 6 classes, 3 entry points. Zero hardcoded hyperparameters — config-driven throughout. LunarLanderEnv → env wrapping + seed control PPOAgent → model lifecycle Trainer → training loop HyperparamSweeper → Optuna HPO Evaluator → separate eval pipeline WandBLogger → centralized tracking

🔬 Training Details

Parameter	Baseline	Optuna Best
learning_rate	3e-4	optimized
n_steps	1024	optimized
clip_range	0.20	0.161
batch_size	64	optimized
n_epochs	4	optimized
total_timesteps	1M	1M
seed	42	42
n_envs	16	16

📊 Experiment Tracking

W&B Dashboard → https://wandb.ai/muhrivandysetiawan-muh-rivandy-setiawan/deep-rl-portfolio 22 tracked runs — baseline, 20 Optuna trials, final.

🗂️ Portfolio

Project 1/12 — HuggingFace Deep RL Curriculum Author: muhrivandysetiawan