Meta-SAGE: Scale Meta-Learning Scheduled Adaptation with Guided Exploration for Mitigating Scale Shift on Combinatorial Optimization

This repository is the official implementation of Meta-SAGE: Scale Meta-Learning Scheduled Adaptation with Guided Exploration for Mitigating Scale Shift on Combinatorial Optimization (ICML 2023).

https://arxiv.org/abs/2306.02688

Dependencies

• Python>=3.8
• PyTorch>=1.7

Training Scale Meta Learner (SML)

POMO-TSP

1. Generate data

cd ./data_generation  

python generate_data.py --name train --problem tsp --dataset_size 3000 --seed 12345 --graph_size 200 300 400 --data_dir data/train

2. Generate target embedding label

cd ./train/POMO/TSP/2_Meta

python Generate_label.py --ep 3000 --problem_size 200 --eas_batch_size 50 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 300 --eas_batch_size 25 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 400 --eas_batch_size 10 --eas_num_iter 100 --seed 12345

3. Training Scale Meta Learner (SML)

python SML_train.py --ep 3000 --eas_batch_size 5

POMO-CVRP

1. Generate data

cd ./data_generation  

python generate_data.py --name train --problem vrp --dataset_size 3000 --seed 12345 --graph_size 200 300 400 --data_dir data/train

2. Generate target embedding label

cd ./train/POMO/CVRP/2_Meta

python Generate_label.py --ep 3000 --problem_size 200 --eas_batch_size 50 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 300 --eas_batch_size 25 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 400 --eas_batch_size 10 --eas_num_iter 100 --seed 12345

3. Training Scale Meta Learner (SML)

python SML_train.py --ep 3000 --eas_batch_size 5

Sym_NCO-TSP

1. Generate data

cd ./data_generation  

python generate_data.py --name train --problem tsp --dataset_size 3000 --seed 12345 --graph_size 200 300 400 --data_dir data/train

2. Generate target embedding label

cd ./train/Sym-NCO/Sym-NCO-POMO/TSP/2_Meta

python Generate_label.py --ep 3000 --problem_size 200 --eas_batch_size 50 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 300 --eas_batch_size 25 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 400 --eas_batch_size 10 --eas_num_iter 100 --seed 12345

3. Training Scale Meta Learner (SML)

python SML_train.py --ep 3000 --eas_batch_size 5

Sym_NCO-CVRP

1. Generate data

cd ./data_generation  

python generate_data.py --name train --problem vrp --dataset_size 3000 --seed 12345 --graph_size 200 300 400 --data_dir data/train

2. Generate target embedding label

cd ./train/Sym-NCO/Sym-NCO-POMO/CVRP/2_Meta

python Generate_label.py --ep 3000 --problem_size 200 --eas_batch_size 50 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 300 --eas_batch_size 25 --eas_num_iter 100 --seed 12345

python Generate_label.py --ep 3000 --problem_size 400 --eas_batch_size 10 --eas_num_iter 100 --seed 12345

3. Training Scale Meta Learner (SML)

python SML_train.py --ep 3000 --eas_batch_size 5

Testing Meta-SAGE

POMO-TSP

SAGE

cd ./test/POMO/TSP/2_SAGE  

python test.py --ep 1000 --problem_size 200 --sage_batch_size 50 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 500 --sage_batch_size 10 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 1000 --sage_batch_size 4 --iter 200 --use_bias 

POMO-CVRP

SAGE

cd ./test/POMO/CVRP/2_SAGE  

python test.py --ep 1000 --problem_size 200 --sage_batch_size 50 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 500 --sage_batch_size 10 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 1000 --sage_batch_size 4 --iter 200 --use_bias 

Sym_NCO-TSP

SAGE

cd ./test/Sym-NCO/TSP/2_SAGE  

python test.py --ep 1000 --problem_size 200 --sage_batch_size 50 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 500 --sage_batch_size 10 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 1000 --sage_batch_size 4 --iter 200 --use_bias 

Sym_NCO-CVRP

SAGE

cd ./test/Sym-NCO/CVRP/2_SAGE  

python test.py --ep 1000 --problem_size 200 --sage_batch_size 50 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 500 --sage_batch_size 10 --iter 200 --use_bias 

python test.py --ep 128 --problem_size 1000 --sage_batch_size 4 --iter 200 --use_bias