PPO Agent playing LunarLander-v2

This is a trained model of a PPO agent playing LunarLander-v2 using the stable-baselines3 library.

Usage (with Stable-baselines3)

from huggingface_sb3 import load_from_hub<br /> from stable_baselines3 import PPO<br /> from stable_baselines3.common.env_util import make_vec_env<br /> from stable_baselines3.common.evaluation import evaluate_policy<br /> # Download checkpoint<br /> checkpoint = load_from_hub("araffin/ppo-LunarLander-v2", "ppo-LunarLander-v2.zip")<br /> # Load the model<br /> model = PPO.load(checkpoint)<br /> env = make_vec_env("LunarLander-v2", n_envs=1)<br /> # Evaluate<br /> print("Evaluating model")<br /> mean_reward, std_reward = evaluate_policy(<br /> model,<br /> env,<br /> n_eval_episodes=20,<br /> deterministic=True,<br /> )<br /> print(f"Mean reward = {mean_reward:.2f} +/- {std_reward:.2f}")<br /> # Start a new episode<br /> obs = env.reset()<br /> try:<br /> while True:<br /> action, _states = model.predict(obs, deterministic=True)<br /> obs, rewards, dones, info = env.step(action)<br /> env.render()<br /> except KeyboardInterrupt:<br /> pass<br />

Training code (with SB3)

from stable_baselines3 import PPO<br /> from stable_baselines3.common.env_util import make_vec_env<br /> from stable_baselines3.common.callbacks import EvalCallback<br /> # Create the environment<br /> env_id = "LunarLander-v2"<br /> n_envs = 16<br /> env = make_vec_env(env_id, n_envs=n_envs)<br /> # Create the evaluation envs<br /> eval_envs = make_vec_env(env_id, n_envs=5)<br /> # Adjust evaluation interval depending on the number of envs<br /> eval_freq = int(1e5)<br /> eval_freq = max(eval_freq // n_envs, 1)<br /> # Create evaluation callback to save best model<br /> # and monitor agent performance<br /> eval_callback = EvalCallback(<br /> eval_envs,<br /> best_model_save_path="./logs/",<br /> eval_freq=eval_freq,<br /> n_eval_episodes=10,<br /> )<br /> # Instantiate the agent<br /> # Hyperparameters from https://github.com/DLR-RM/rl-baselines3-zoo<br /> model = PPO(<br /> "MlpPolicy",<br /> env,<br /> n_steps=1024,<br /> batch_size=64,<br /> gae_lambda=0.98,<br /> gamma=0.999,<br /> n_epochs=4,<br /> ent_coef=0.01,<br /> verbose=1,<br /> )<br /> # Train the agent (you can kill it before using ctrl+c)<br /> try:<br /> model.learn(total_timesteps=int(5e6), callback=eval_callback)<br /> except KeyboardInterrupt:<br /> pass<br /> # Load best model<br /> model = PPO.load("logs/best_model.zip")<br />