An extensive Reinforcement Learning (RL) for Combinatorial Optimization (CO) benchmark. Our goal is to provide a unified framework for RL-based CO algorithms, and to facilitate reproducible research in this field, decoupling the science from the engineering.
RL4CO is built upon:
- TorchRL: official PyTorch framework for RL algorithms and vectorized environments on GPUs
- TensorDict: a library to easily handle heterogeneous data such as states, actions and rewards
- PyTorch Lightning: a lightweight PyTorch wrapper for high-performance AI research
- Hydra: a framework for elegantly configuring complex applications
We offer flexible and efficient implementations of the following policies:
- Constructive: learn to construct a solution from scratch
- Autoregressive (AR): construct solutions one step at a time via a decoder
- NonAutoregressive (NAR): learn to predict a heuristic, such as a heatmap, to then construct a solution
- Improvement: learn to improve an pre-existing solution
We provide several utilities and modularization. For example, we modularize reusable components such as environment embeddings that can easily be swapped to solve new problems.
RL4CO is now available for installation on pip
!
pip install rl4co
To get started, we recommend checking out our quickstart notebook or the minimalistic example below.
This command installs the bleeding edge main
version, useful for staying up-to-date with the latest developments - for instance, if a bug has been fixed since the last official release but a new release hasn’t been rolled out yet:
pip install -U git+https://github.com/ai4co/rl4co.git
If you want to develop RL4CO we recommend you to install it locally with pip
in editable mode:
git clone https://github.com/ai4co/rl4co && cd rl4co
pip install -e .
We recommend using a virtual environment such as conda
to install rl4co
locally.
Train model with default configuration (AM on TSP environment):
python run.py
Tip
You may check out this notebook to get started with Hydra!
Change experiment settings
Train model with chosen experiment configuration from configs/experiment/
python run.py experiment=routing/am env=tsp env.num_loc=50 model.optimizer_kwargs.lr=2e-4
Here you may change the environment, e.g. with env=cvrp
by command line or by modifying the corresponding experiment e.g. configs/experiment/routing/am.yaml.
Disable logging
python run.py experiment=routing/am logger=none '~callbacks.learning_rate_monitor'
Note that ~
is used to disable a callback that would need a logger.
Create a sweep over hyperparameters (-m for multirun)
python run.py -m experiment=routing/am model.optimizer.lr=1e-3,1e-4,1e-5
Here is a minimalistic example training the Attention Model with greedy rollout baseline on TSP in less than 30 lines of code:
from rl4co.envs.routing import TSPEnv, TSPGenerator
from rl4co.models import AttentionModelPolicy, POMO
from rl4co.utils import RL4COTrainer
# Instantiate generator and environment
generator = TSPGenerator(num_loc=50, loc_distribution="uniform")
env = TSPEnv(generator)
# Create policy and RL model
policy = AttentionModelPolicy(env_name=env.name, num_encoder_layers=6)
model = POMO(env, policy, batch_size=64, optimizer_kwargs={"lr": 1e-4})
# Instantiate Trainer and fit
trainer = RL4COTrainer(max_epochs=10, accelerator="gpu", precision="16-mixed")
trainer.fit(model)
Other examples can be found on the documentation!
Run tests with pytest
from the root directory:
pytest tests
Installing PyG
via Conda
seems to update Torch itself. We have found that this update introduces some bugs with torchrl
. At this moment, we recommend installing PyG
with Pip
:
pip install torch_geometric
Have a suggestion, request, or found a bug? Feel free to open an issue or submit a pull request. If you would like to contribute, please check out our contribution guidelines here. We welcome and look forward to all contributions to RL4CO!
We are also on Slack if you have any questions or would like to discuss RL4CO with us. We are open to collaborations and would love to hear from you 🚀
If you find RL4CO valuable for your research or applied projects:
@misc{berto2024rl4co,
title={{RL4CO: an Extensive Reinforcement Learning for Combinatorial Optimization Benchmark}},
author={Federico Berto and Chuanbo Hua and Junyoung Park and Laurin Luttmann and Yining Ma and Fanchen Bu and Jiarui Wang and Haoran Ye and Minsu Kim and Sanghyeok Choi and Zepeda Gast and Andre Hottung and Jianan Zhou and Jieyi Bi and Yu Hu and Fei Liu and Hyeonah Kim and Jiwoo Son and Haeyeon Kim and Davide Angioni and Wouter Kool and Zhiguang Cao and Jie Zhang and Kijung Shin and Cathy Wu and Sungsoo Ahn and Guojie Song and Changhyun Kwon and Lin Xie and Jinkyoo Park},
year={2024},
eprint={2306.17100},
archivePrefix={arXiv},
primaryClass={cs.LG},
note={\url{https://github.com/ai4co/rl4co}}
}
Note that a previous version of RL4CO has been accepted as an oral presentation at the NeurIPS 2023 GLFrontiers Workshop. Since then, the library has greatly evolved and improved!
We invite you to join our AI4CO community, an open research group in Artificial Intelligence (AI) for Combinatorial Optimization (CO)!