Code for paper "Differentiable Distributionally Robust Optimization Layers"

This repository contains the source code to reproduce the experiments of our paper "Differentiable Distributionally Robust Optimization Layers" presented in ICML 2024.

Dependency

Except for commonly used packages like torch and numpy, other important packages our code is built on are listed as follows.

• cvxpylayers
• cvxpy
• gurobipy

Moreover, make sure that Gurobi solver can be called in Python environment. This tutorial may be helpful for building Gurobi in Python.

Usage

Code for experiments on the multi-item newsvendor problem is in folder 'Toy_Example_Newsvendor', and running the main file in this folder will conduct this experiment.

Code for experiments on the portfolio management problem is in folder 'Portfolio_Management_Problem'. Under this folder, the 'Pure_Continous_Decision_Case' folder contains experiments where all decision variables are continuous, and 'Mixed_Integer_Decision_Case' folder contains experiments where decision variables are mixed-integer. Simply running the main file in folders 'Portfolio_Management_Problem/Pure_Continous_Decision_Case' and 'Portfolio_Management_Problem/Mixed_Integer_Decision_Case' will conduct corresponding experiments.

For anyone who wants to construct a DRO Layer for their own problems, using the cvxpylayers package directly when all decision variables are continuous. When the decision variables are mixed-integer, we have provided MIDROLayer class (in 'Portfolio_Management_Problem\Mixed_Integer_Decision_Case\my_layer_MI.py') for computing and back-propagating the gradient.