A5omic/pong-dqn

🕹️ DQN Agent for Pong (PongNoFrameskip-v4)

This repository contains a DQN (Deep Q-Network) agent trained to play Pong using the PongNoFrameskip-v4 environment from the Atari Learning Environment.

The agent was trained with Stable-Baselines3 and uploaded via the huggingface_sb3 integration as part of a reinforcement learning project. 📊 Training Details

Environment: PongNoFrameskip-v4

Algorithm: DQN (Deep Q-Network)

Library: Stable-Baselines3

Timesteps: 100,000

Evaluation Mean Reward: ~-21.00

Frame Stack: 4

Replay Video: (Not available due to environment limitation)

Training was performed on a single environment with default hyperparameters. 💾 Load the Model from the Hub

To use this model directly in your own project:

from stable_baselines3 import DQN from huggingface_sb3 import load_from_hub

Load checkpoint from the Hub

repo_id = "A5omic/pong-dqn" filename = "pong-dqn.zip"

checkpoint = load_from_hub(repo_id, filename) model = DQN.load(checkpoint)

Now you can use model.predict() or evaluate

⚠️ Notes

The model was trained for demonstration purposes with limited training time (100k steps).

Evaluation scores may improve with longer training (e.g., 1M+ steps).

Replay video failed to render due to a known issue with DummyVecEnv. Consider switching to VecVideoRecorder with SubprocVecEnv for future uploads.

🧠 About DQN

DQN is one of the foundational deep RL algorithms that combines Q-learning with deep neural networks. It uses experience replay and target networks to stabilize training and is well-suited for environments like Atari games. 📬 Citation / Attribution

This work builds on:

Stable-Baselines3

Hugging Face 🤗 RL Course

Atari Learning Environment