Introduction

The PyEPO (PyTorch-based End-to-End Predict-then-Optimize Tool) is an open-source software implemented in Python that facilitates the modeling and resolution of predict-then-optimize problems characterized by linear objective functions.

The primary function of PyEPO is to construct optimization models utilizing GurobiPy, Pyomo, or other solvers and algorithms, and then integrating the optimization model within an artificial neural network for end-to-end training. To achieve this, PyEPO incorporates the SPO+ loss, differentiable black-box optimizer, differentiable perturbed optimizers, and Fenchel-Young loss with Perturbation as PyTorch autograd modules.

End-to-End Predict-then-Optimize Framework

A labeled dataset \(\mathcal{D}\) comprising of feature-cost pairs \((x,c)\) or feature-solution pairs \((x,w)\) is utilized to train a machine learning model, specifically a deep neural network, which directly minimizes the decision error.