Learning to branch for the Crew Pairing Problems is an article currently under submission at Operations Research Forum. This repository contains the code used to train the branching heuristics based on the strong branching policy.
You will also find our collected dataset, trained models and results of our experiements, in the folder data_results.
We trained 3 models :
- Linear model
- Multilayer Perceptron (MLP)
- Transformer encoder
Considering our models have to select the branching candidate among 3 possible candidates, we would be in a situation like this:
Those models evaluate each candidate separately. It could be represented like this (for the MLP):
This model is able to evaluate all candidates at the same time.
By using the attention module, it is able to ponderate each candidate predictions based on the other candidates.
It could be represented like this:
