PPO Agent Playing LunarLander-v3

This is a trained PPO agent playing LunarLander-v3, using a custom single-file PyTorch implementation (closely following the CleanRL PPO reference).

Evaluation

Mean reward over 10 episodes: 12.81 +/- 87.50

Hyperparameters

num_envs = 4
num_steps = 128
learning_rate = 0.00025
total_timesteps = 500000
gamma = 0.99
gae_lambda = 0.95
epochs = 4
clip_coef = 0.2
entropy_coef = 0.01
vf_coef = 0.5
max_grad_norm = 0.5
batch_size = 512
minibatch_size = 128

Usage

import torch
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="kapoorkrish/ppo-LunarLander-v3", filename="model.pt")
agent = Agent(envs)  # rebuild the same Agent class/architecture used in training
agent.load_state_dict(torch.load(model_path))
agent.eval()