1️⃣ First work to incorporate end-to-end vehicle routing model in a modern RL platform (CleanRL)

⚡ Speed up the training of Attention Model by 8 times (25hours $\to$ 3 hours)

🔎 A flexible framework for developing model, algorithm, environment, and search for operation research

• 13/04/2023: We release web demo on Hugging Face 🤗!
• 24/03/2023: We release our paper on arxiv!
• 20/03/2023: We release jupyter lab demo and pretrained checkpoints!
• 10/03/2023: We release our codebase!

We provide inference demo on colab notebook:

Environment	Search	Demo
TSP	Greedy
CVRP	Multi-Greedy

conda env create -n <env name> -f environment.yml
# The environment.yml was generated from
# conda env export --no-builds > environment.yml

It can take a few minutes.

wandb

Refer to their quick start guide for installation.

All the major implementations were under rlor folder.

./rlor
├── envs
│   ├── tsp_data.py # load pre-generated data for evaluation
│   ├── tsp_vector_env.py # define the (vectorized) gym environment
│   ├── cvrp_data.py 
│   └── cvrp_vector_env.py 
├── models
│   ├── attention_model_wrapper.py # wrap refactored attention model to cleanRL
│   └── nets # contains refactored attention model
└── ppo_or.py # implementaion of ppo with attention model for operation research problems

The ppo_or.py was modified from cleanrl/ppo.py. To see what's changed, use diff:

# apt install diff
diff --color ppo.py ppo_or.py

python ppo_or.py --num-steps 51 --env-id tsp-v0 --env-entry-point envs.tsp_vector_env:TSPVectorEnv --problem tsp

python ppo_or.py --num-steps 60 --env-id cvrp-v0 --env-entry-point envs.cvrp_vector_env:CVRPVectorEnv --problem cvrp

python ppo_or.py ... --track

Add --track argument to enable tracking with WandB.

It can be generated from the official repo of the attention-learn-to-route paper. You may modify the ./envs/tsp_data.py to update the path to data accordingly.

The neural network model is refactored and developed from Attention, Learn to Solve Routing Problems!.

The idea of multiple trajectory training/ inference is from POMO: Policy Optimization with Multiple Optima for Reinforcement Learning.

The RL environments are defined with OpenAI Gym.

The PPO algorithm implementation is based on CleanRL.