ELG (Ensemble of Local and Global policies)

This repository is the code of https://arxiv.org/abs/2308.14104, which ensembles a transferrable local policy to boost generalization. Our code is built on the code of POMO[1]. We provide the trained models to reproduce the test results in the paper.

Test ELG* on VRPLIB[2, 4]

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.

Train ELG* on CVRP

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

Test ELG* on TSPLIB[4]

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

python test_tsplib.py

Train ELG* on TSP

First, generate the validation sets by

python generate_data.py

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

python train.py

Reference:

[1] Kwon, Y.-D.; Choo, J.; Kim, B.; Yoon, I.; Gwon, Y.; and Min, S. 2020. POMO: Policy optimization with multiple optima for reinforcement learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 21188–21198. Virtual.

[2] Uchoa, E.; Pecin, D.; Pessoa, A.; Poggi, M.; Vidal, T.; and Subramanian, A. 2017. New benchmark instances for the capacitated vehicle routing problem. European Journal of Operational Research, 257(3): 845–858.

[3] Reinelt, G. 1991. TSPLIB - A traveling salesman problem library. ORSA Journal on Computing, 3(4): 376–384.

[4] Arnold, F.; Gendreau, M.; and S¨orensen, K. 2019. Efficiently solving very large-scale routing problems. Computers & Operations Research, 107: 32–42.