This package implements techniques of machine learning for operations research to compute shortest paths on Warcraft terrain images. It is one of the applications of our paper Learning with Combinatorial Optimization Layers: a Probabilistic Approach.

This application was introduced in this paper, with the corresponding dataset and code. It was also considered in a Learning with Differentiable Perturbed Optimizers setting, with corresponding code.

We focus on two main frameworks: learning by imitation, involving Fenchel-Young losses and perturbed maximizers, and learning from experience. Both are based on the InferOpt.jl package. We also leverage the GridGraphs.jl package to compute shortest paths on grid graphs using Dijkstra algorithm.

1. Please download this dataset and place it in the data folder of the repo. You can unzip it manually or using decompress_dataset function.

2. You can then activate the environment locally with this command:

using Pkg
Pkg.activate(".")

Each point of the dataset is linked to a (12x12) Warcraft terrain grid. It is composed of:

• A color image of the Warcraft terrain of size (96x96).
• The cost labels for the corresponding grid of size (12x12).
• A 0-1 shortest path mask of size (12x12).

The two distinct frameworks we consider are:

1. Learning by imitation: given the images and labels, learn the cost such that the labelled shortest paths are close to the shortest path computed with the given cost using Dijkstra on the Warcraft grids.

2. Learning by experience: given the images and a black-box function that computes the cost of a path on any grid, learn the cost such that the true cost of the paths computed as shortest paths with respect to the learned costs are low.