We use attention model to solve SBRP considering number of routes. Taking number of routes (reprensenting the bus on SBRP) into account, we trained our own model instance with new objective function.
Our implementatin is mainly on the problems/vrp/problem_vrp.py and the class CVRP_BUS inside it.
We also implement more real-world like validation data generation on generate_data_cvrp_bus.py.
Our train result is on team23_model directory which are used on result_team23.ipynb (for epoch 99).
Most of code implementation is based on the source code from repository of the paper Attention, Learn to Solve Routing Problems! which has been accepted at ICLR 2019
Training data is generated on the fly. To generate validation as used in result_team23.ipynb
python generate_data_cvrp_bus.py -f --graph_sizes 20 Generated data will be saved as data/cvrp_bus/cvrp_bus20_validation_seed1234.pkl
For training our model instance with new objective function:
python run.py --problem 'cvrp_bus' --graph_size 20 --baseline rollout --run_name 'cvrp_bus_20_rollout'The output will be saved on outputs/cvrp_bus_20 directory.
See result_team23.ipynb for an example of loading a pretrained model and plotting the result for Capacitated VRP with 20 nodes.
We compared our model with new objective function and previous pretrained model with naive distance objective on this jupyter notebook with 10000 validation data.
Below is the one of the comparisons. You can check the result_team23.ipynb for more cases.
python run.py -h
python eval.py -hThanks to wouterkool/attention-learn-to-route for making us to reproduce our product.