masterdezign/dqn2-SpaceInvadersNoFrameskip-v4

DQN Agent playing SpaceInvadersNoFrameskip-v4

This is a trained model of a DQN agent playing SpaceInvadersNoFrameskip-v4 using the stable-baselines3 library.

Usage (with Stable-baselines3)

 from stable_baselines3.common.env_util import make_atari_env
from stable_baselines3.common.vec_env import VecFrameStack
from stable_baselines3 import DQN
from stable_baselines3.common.evaluation import evaluate_policy
from huggingface_sb3 import load_from_hub, package_to_hub
from stable_baselines3.common.utils import set_random_seed

env_id = "SpaceInvadersNoFrameskip-v4"

env = make_atari_env(env_id,
                     n_envs=12,
                     # Improving reproducibility
                     seed=1)
env = VecFrameStack(env, n_stack=4)  # Stack last four images

# Improving reproducibility
set_random_seed(42)

# Using these parameters as default: https://huggingface.co/micheljperez/dqn-SpaceInvadersNoFrameskip-v4
model = DQN(policy = "CnnPolicy",
            env = env,
            batch_size = 32,
            buffer_size = 100_000,
            exploration_final_eps = 0.01,
            exploration_fraction = 0.025,
            gradient_steps = 1,
            learning_rate = 1e-4,
            learning_starts = 100_000,
            optimize_memory_usage = True,
            replay_buffer_kwargs = {"handle_timeout_termination": False},
            target_update_interval = 1000,
            train_freq = 4,
            # normalize = False,
            tensorboard_log = "./tensorboard",
            verbose=1
           )
           
f = load_from_hub('masterdezign/dqn2-SpaceInvadersNoFrameskip-v4', 'dqn-SpaceInvadersNoFrameskip-v4.zip')
model = model.load(f)

mean_reward, std_reward = evaluate_policy(model, env)
print(f"Mean reward = {mean_reward:.2f} +/- {std_reward:.2f}")