DQN Crasher

dqn_crasher uses RL controlled NPC agents to cause collisions in the highway-env gymnaisum environment.

Installation

We recommend using Conda or Mamba for environment management.

git clone https://github.com/amarkoolk/dqn_crasher.git
cd dqn_crasher
pip install -e .

This will install all dependencies and make the dqn_crasher package (and its console scripts) available.

Usage

Running Experiments

You can run the main experiment either as a Python module or via the installed console script:

# Using the Python module interface
python -m dqn_crasher.main --config configs/default.yaml

# Or using the console script
dqn-crasher --config configs/default.yaml

Configuration

All experiments are configured via YAML files in the src/dqn_crasher/configs/ directory. Example:

env: "crash-v0"
seed: 42
episodes: 1000
dqn:
  learning_rate: 0.0005
  batch_size: 64
  epsilon_decay: 0.995
  ...

You can override config values via command-line flags:

python -m dqn_crasher.main --config configs/default.yaml dqn.learning_rate=0.001

WandB Sweeps

To launch a hyperparameter sweep with Weights & Biases:

wandb sweep configs/sweep.yaml
wandb agent <SWEEP_ID>

See the configs/ folder for ready-to-use sweep configs.

Examples

Train a DQN agent and analyze crash scenarios:

python -m dqn_crasher.main

Project Structure

dqn_crasher/
├── src/dqn_crasher/       # core package
│   ├── agents/            # DQN agent
│   ├── buffers/           # replay buffer, sum-tree
│   ├── configs/           # env + model configs (packaged)
│   ├── scenarios/         # scenarios & policies
│   ├── training/          # runner
│   ├── utils/             # helpers, logging
│   └── visualization/     # plotting
├── sweep_configs/         # WandB sweep configs (not packaged)
├── examples/              # experiment & plotting scripts
├── slurm_scripts/         # Slurm job scripts
├── models/                # trained models (git-ignored)
├── environment.yml        # Conda/Mamba env
└── pyproject.toml         # pip install entry

Contributing

Contributions are welcome! Please open an issue or submit a pull request. See CONTRIBUTING.md for guidelines.

License

MIT License. See the LICENSE file for details.

Contact

For questions or suggestions, please open an issue or contact the maintainer at amar@example.com.