A clean, extensible reinforcement learning framework built with PyTorch. Implements classical RL algorithms with proper experiment tracking and visualization tools.

• Multiple RL Algorithms: DQN variants, Policy Gradient (REINFORCE), and extensible base classes for new algorithms
• Experiment Tracking: Structured logging with TensorBoard integration
• Modular Design: Clean separation between environments, models, and training logic
• Easy Extension: Add new algorithms with minimal boilerplate code

1. Install dependencies:

pip install -r requirements.txt

2. Run training:

python train.py configs/pong/dqn.json
python train.py configs/cartpole/pg.json

3. Monitor training:

tensorboard --logdir experiments/

├── models/              # RL algorithms and neural networks
│   ├── model.py        # Base classes (BaseAgent, etc.)
│   ├── base_trainer.py # Common training patterns
│   ├── network.py      # Neural network architectures
│   ├── value_based/    # DQN family algorithms
│   │   ├── base_value.py
│   │   ├── dqn.py
│   │   ├── ddqn.py
│   │   └── rainbow.py
│   ├── policy_based/   # Policy gradient algorithms
│   │   ├── base_policy.py
│   │   └── vanilla_pg.py
│   ├── actor_critic/   # Actor-critic algorithms
│   │   ├── base_actor_critic.py
│   │   ├── a2c.py
│   │   ├── a3c.py
│   │   └── adversarial_a2c.py
│   └── README.md      # Guide for adding new algorithms
├── games/              # Environment wrappers and utilities
│   ├── atari_env.py   # Atari environment preprocessing
│   ├── replay_buffer.py # Experience replay
│   ├── preprocessor.py # Frame preprocessing
│   └── experiment_logger.py # Structured experiment logging
├── configs/            # Algorithm configurations
│   ├── pong/
│   ├── breakout/
│   ├── skiing/
│   ├── asterix/
│   └── cartpole/
├── experiments/        # Training results and checkpoints
├── analysis/           # Experiment analysis and visualization tools
│   ├── analysis_utils.py # Analysis functions and ExperimentAnalyzer class
│   └── experiment_analysis.ipynb # Jupyter notebook for experiment comparison
└── train.py           # Main training script

• DQN: Deep Q-Network with experience replay
• Double DQN: Reduces overestimation bias
• Rainbow DQN: Multi-step learning, prioritized replay, noisy networks

• REINFORCE: Vanilla policy gradient with optional baseline

• A2C: Advantage Actor-Critic
• A3C: Asynchronous Actor-Critic
• Adversarial A2C: A2C with adversarial training

The framework provides base classes that make adding new algorithms straightforward. See models/README.md for a complete guide with examples.

Each algorithm uses JSON configuration files:

{
  "agent_type": "dqn",
  "env": {
    "id": "ALE/Pong-v5"
  },
  "model": {
    "input_shape": [4, 84, 84]
  },
  "train": {
    "num_episodes": 10000,
    "learning_rate": 0.00025,
    "gamma": 0.99
  }
}

Use the provided Jupyter notebook to analyze and compare experiment results:

jupyter notebook analysis/experiment_analysis.ipynb

The notebook provides:

• Training curve visualization
• Algorithm comparison
• Performance metrics analysis
• Hyperparameter sensitivity analysis

• Classic Control: CartPole, MountainCar, Acrobot
• Atari: All Atari games via Gymnasium
• Custom: Easy to add new environments through gym interface

Experiments are automatically logged with:

• Training metrics (rewards, losses, episode lengths)
• TensorBoard visualizations
• Model checkpoints
• Configuration snapshots