gaocrr/elg

Official implementation of IJCAI'24 paper "Towards Generalizable Neural Solvers for Vehicle Routing Problems via Ensemble with Transferrable Local Policy"

ELG (Ensemble of Local and Global policies)

This repository is the code of the https://arxiv.org/abs/2308.14104, which ensembles a transferrable local policy to boost generalization. We provide the trained models to reproduce the test results in the paper.

Test ELG-POMO on VRPLIB

Under the ELG/CVRP folder, use the default settings in config.yml, run

python test_vrplib.py

You can choose the vrplib_set config from {X, XXL} to test on two different VRPLIB sets.

Test ELG-POMO on TSPLIB

Under the ELG/TSP folder, use the default settings in config.yml, and run

python test_tsplib.py

Train ELG-POMO on CVRP or TSP

First, generate the validation sets by

python generate_data.py

Modify the load_checkpoint config in config.yml to Null (i.e., load_checkpoint: ), and run

python train.py

elg's Issues

the solution of a280 is not valid.

Hello, when I try to run the TSPLib dataset with the code and model that you have provided, I find that the solution of a280 is not valid, I would like the author to check it, thanks!

Code for checking validity:
if done: assert (self.step_state.ninf_mask == float('-inf')).all(), 'ninf_mask is not all -inf, but done is True, so the solution is not valid.'

Best regards!