Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints
Authors: Jayanta Mandi, Marianne Defresne, Senne Berden, Tias Guns
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experimental Evaluation, Figure 1: Infeasibility ratio and regret of feasible solutions on Test instances, We denote the model as Odece(α) (e.g., Odece(0.1)) to indicate Odece is trained with that specific value of α. We compare the proposed Odece against four competitors: i) MSE: a PFL approach which trains the ML model to minimize the MSE loss over the parameter predictions, ii) Comb Opt Net: the technique proposed by Paulus et al. [26] to compute the gradient of ILP parameters, iii) SFL: the solver-free learning proposed by Nandwani et al. [24] for learning parameters of an ILP, iv) 2s Pt O: the two-stage predict+optimize approach proposed by Hu et al. [15]. We evaluate performance using two metrics: the proportion of infeasible solutions and the normalized regret on the test data. |
| Researcher Affiliation | Academia | Jayanta Mandi Department of Computer Science KU Leuven, Leuven, Belgium EMAIL, Marianne Defresne LAAS-CNRS, Université de Toulouse CNRS, INSA, Toulouse EMAIL, Senne Berden Department of Computer Science KU Leuven, Leuven, Belgium EMAIL, Tias Guns Department of Computer Science KU Leuven, Leuven, Belgium EMAIL |
| Pseudocode | No | The paper describes its methodology using mathematical formulations and textual descriptions in Section 4 "Methodology", but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code for reproducing our results is publicly available at: https://github.com/Jay Man91/ Odece DFLfor Constraints Neurips25. |
| Open Datasets | Yes | Brass Alloy Production. Our third experimental setting involves the Brass alloy production problem. We reproduce this experimental setup from the work by Hu et al. [15]. It is a covering LP problem. We use the publicly available Brass alloy data from their repository.2, 2https://github.com/Elizabethxyhu/Neur IPS_Two_Stage_Predict-Optimize/, Synthetic parameter generation of the MDKP. We adopt the synthetic data generation process described by Elmachtoub and Grigas [12]. |
| Dataset Splits | Yes | For each run, 1500 instances are generated, split into 900 for training, 100 for validation, and 500 for testing. Out of the available 500 instances, we use 350, 50, and 100 for training, validation, and testing, respectively. |
| Hardware Specification | Yes | The experiments were executed on an Intel i7-13800H (20 cores) CPU with 32GB RAM. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers. While it references libraries in the bibliography, it does not state the versions of the software used for its own implementation. |
| Experiment Setup | Yes | α [0, 1] is a hyperparameter we call the infeasibility-aversion coefficient. It reflects how much importance the decision-maker places on avoiding infeasible solutions (through IPL) versus maintaining the feasibility of known feasible ones (through OPL). Finally, in Table 7, we provide the hyperparameter configurations used for each model to reproduce the results reported above. |