liu-yang-maker / LODLs

Code for "Decision-Focused Learning without Differentiable Optimization: Learning Locally Optimized Decision Losses"

Code for "Decision-Focused Learning without Decision-Making: Learning Locally Optimized Decision Losses"

Set up environment

To run the code, first you have to set up a conda environment. Once you have Anaconda installed, run the following command:

conda env create --name lodlenv --file=environment.yml

Once the environment has been created, load it using the command

conda activate lodlenv

You'll also need to manually install one of the libraries using the following command

conda install metis

Running Different Domains

We now present the default parameters required to run our DirectedMSE approach on the different domains from the paper below.

Linear Model

python main.py --problem=cubic --loss=weightedmse++ --seed=1 --instances=400 --testinstances=400 --valfrac=0.5 --numitems=50 --budget=1 --lr=0.1 --layers=1 --sampling=random_hessian --samplingstd=1 --numsamples=50000 --losslr=0.01 --serial=False

Web Advertising

python main.py --problem=budgetalloc --loss=weightedmse++ --seed=1 --instances=100 --testinstances=500 --valfrac=0.2 --numitems=5 --budget=1 --sampling=random_hessian --numsamples=5000 --losslr=0.1 --serial=False

Portfolio Optimization

python main.py --problem=portfolio --loss=weightedmse++ --seed=1 --instances=400 --testinstances=400 --valfrac=0.5 --stocks=50 --stockalpha=0.1 --lr=0.01 --sampling=random_hessian --samplingstd=0.1 --numsamples=50000 --losslr=0.001 --serial=False

Note: (A) Sampling points can take a long time. (B) In the event that you receive an error along the lines of Thread creation failed: Resource temporarily unavailable, you should run the following command:

export OMP_NUM_THREADS=1

Running Different Approaches

In Table 1 of our paper, we highlight 7 different approaches. To run the domains above with a specific approach, set the --loss parameter to the input corresponding to that approach. (Note: Obtaining the results in the paper requires method-specific hyperparameter tuning.)

Method	Corresponding Input
2-Stage	mse
DFL	dfl
WeightedMSE	weightedmse
DirectedWeightedMSE	weightedmse++
DirectedWeightedMSE	weightedmse++
Quadratic	quad
DirectedQuadratic	quad++
NN	dense

Similarly, to run different sampling strategies, set the --sampling parameter according to the following table:

Method	Corresponding Input
1-perturbed	random_jacobian
2-perturbed	random_hessian
All-perturbed	random