yd-kwon / sgbs

codes for the paper "Simulation-guided Beam Search for Neural Combinatorial Optimization"

Simulation-Guided Beam Search for Neural Combinatorial Optimization

This repository is the official implementation of Simulation-Guided Beam Search for Neural Combinatorial Optimization (NeurIPS 2022).

https://arxiv.org/abs/2207.06190

Inference with SGBS/SGBS+EAS method

📋 TSP

SGBS

cd ./TSP/2_SGBS  
python3 test.py

SGBS+EAS

cd ./TSP/3_SGBS+EAS  
python3 test.py

📋 CVRP

SGBS

cd ./CVRP/2_SGBS  
python3 test.py

SGBS+EAS

cd ./CVRP/3_SGBS+EAS  
python3 test.py

📋 FFSP

To run SGBS and SGBS+EAS for FFSP, you have to download and unpack FFSP.tar.gz first.
See Requirements - FFSP trained model & dataset section.

SGBS

cd ./FFSP/2_SGBS  
python3 test_ffsp20.py
python3 test_ffsp50.py
python3 test_ffsp100.py

SGBS+EAS

cd ./FFSP/3_SGBS+EAS  
python3 test_ffsp20.py
python3 test_ffsp50.py
python3 test_ffsp100.py

Inference with Greedy, Sampling, Beam Search, MCTS or SGBS method for CVRP

cd ./CVRP/2_SGBS  
python3 test.py -disable_aug --mode greedy
python3 test.py -disable_aug --mode sampling
python3 test.py -disable_aug --mode obs
python3 test.py -disable_aug --mode mcts
python3 test.py -disable_aug --mode sgbs

-disable_aug: disables instance augmentation
--mode: specifies inference method, (obs means original beam search method)

If you want test small number of test episodes, you can use --ep option. For example, if you want test just 10 episodes with greedy method,

python3 test.py -disable_aug --mode greedy --ep 10

Note: MCTS takes a lot of time compared to other methods.

Requirements - FFSP trained model & dataset

Download FFSP trained model and dataset file from https://drive.google.com/file/d/1TdkeErG1FCUMxoe8ENpxiWwUopPIUikb/view?usp=sharing.
Move FFSP.tar.gz into your root folder and unpack it.

tar -xvzf FFSP.tar.gz

Language and Libraries

python 3.8.6
torch 1.11.0