This is the implementation of the paper Automatically Learning Compact Quality-aware Surrogates for Optimization Problems accepted as a spotlight presentation at NeurIPS 2020. The paper includes three examples: adversarial modeling in network security games (NSG folder), movie recommendation with a submodular objective (movie folder), and a convex portfolio optimization (portfolio folder). Among these three, NSG uses synthetic data, movie recommendation uses the data from MovieLens (ml-25m), and portfolio optimization uses the data downloaded from Quandl using quandl API.
The commands to run each example are included in each folder. You will have to download the data from MovieLens in movie recommendation and apply for an API key from Quandl in portfolio optimization before running the code.
All the implementations are written in Python3.
Here is a list of dependency:
- qpth 0.0.15
- cvxpylayers 0.1.4
- scipy
- tqdm
- autograd 1.3
- quandl 3.6.0
- pytorch-geometric 1.6.3
[Update 2021/2/7]
I used a fresh conda environment to test the code (all three domains) and fixed some dependency issues.
More specifically, I updated some networkx syntax to accomodate the latest networkx version. I also removed the dependency of gurobipy (and localqpth), which was not used in the final implementation.
They are some other misc changes (disabling qpth verbose etc.), but the main part of the implementation was not involved. You can find the updated package-list.txt in the repo.
Across all three domains, the epoch -1 refers to the optimal performance, where a perfect prediction is used to compute the loss (should be 0) and the corresponding performance. The epoch 0 instead computes the untrained performance, where no training is performed in this epoch but just evaluation. The training only starts from epoch 1.
Sometimes qpth would have some convergence issue. I switched to use cvxpylayers, a newer implementation of differentiable convex optimization layer, in the portifolio domain.