Robust Losses for Decision-Focused Learning

Authors: Noah Schutte, Krzysztof Postek, Neil Yorke-Smith

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that training two state-of-the-art decision-focused learning approaches using robust regret losses improves test sample empirical regret in general while keeping computational time equivalent relative to the number of training epochs.
Researcher Affiliation Academia 1Delft University of Technology 2Independent Researcher {n.j.schutte, n.yorke-smith}@tudelft.nl, krzysztof.postek@gmail.com
Pseudocode No No section or figure explicitly labeled 'Pseudocode' or 'Algorithm' was found, nor were any structured, code-like algorithmic steps presented.
Open Source Code Yes We use Python-based open-source package Py EPO [Tang and Khalil, 2022] for the data generation of two experimental problems and the training, where the robust losses are implemented on top of the existing code. The k-NN loss is currently available in Py EPO.
Open Datasets Yes Energy-cost aware scheduling. As a third experimental problem we look at energy-cost aware scheduling [Simonis et al., 1999] following precedent in a DFL setting [Mandi et al., 2022]. The dataset consist of 789 days of historical energy price data at 30-minute intervals from 2011 2013 [Ifrim et al., 2012].
Dataset Splits Yes In all cases a validation and test set of size 100 and 1000 are used respectively.
Hardware Specification No No specific hardware details such as GPU models, CPU types, or memory specifications are mentioned for the experimental setup. The paper only mentions 'Gurobi version 10.0.1' which is software.
Software Dependencies Yes We use Python-based open-source package Py EPO [Tang and Khalil, 2022] for the data generation of two experimental problems and the training... We use the Adam optimizer with learning rate 0.01 for the gradient descent and Gurobi version 10.0.1 [Gurobi Optimization, 2023] as the optimization problem solver.
Experiment Setup Yes We compare SPO+ and PFYL (number of samples M = 1, perturbation amplitude σ = 1)... We use the Adam optimizer with learning rate 0.01 for the gradient descent... For the top-k loss we use k = 10; the same for the k-NN loss where w = 0.5. For the RO loss we set ρ = 0.5 and Γ = n/8 , where n = |c|. The batch size is 32.