End-to-End Learning for Optimization via Constraint-Enforcing Approximators

In many real-world applications, predictive methods are used to provide inputs for downstream optimization problems. It has been shown that using the task-based objective learn intermediate model is often better than only task objectives, such as prediction error. The learning in former approach referred end-to-end learning. difficulty lies differentiating through problem. Therefore, we propose a neural network architecture can approximately solve these problems, particularly ensuring its output satisfies feasibility constraints via alternate projections. We show projections converge at geometric rate exact projection. Our more computationally efficient existing do not need original problem each iteration. Furthermore, our be applied wider range of apply this shortest path which first stage forecasting computer vision predicting edge costs from terrain maps, capacitated multi-product newsvendor problem, and maximum matching method out-performs approaches terms final loss training time.