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..
Solver-Free Decision-Focused Learning for Linear Optimization Problems
Authors: Senne Berden, Ali Mahmutoğulları, Dimos Tsouros, Tias Guns
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that our method significantly reduces computational cost while maintaining high decision quality. Our experiments show that our solver-free LAVA loss enables efficient training with minimal degradation in solution quality, particularly on problems that are not highly degenerate. We evaluate our method against state-of-the-art approaches, focusing on the trade-off between training time and decision quality, and how performance scales with the number of variables and constraints. Figure 2 shows the training time and test set regret (log scale) of the various methods (full results with standard errors are given in tabular format in Appendix E). |
| Researcher Affiliation | Academia | Senne Berden, Ali Irfan Mahmuto gulları, Dimos Tsouros, Tias Guns Department of Computer Science, KU Leuven EMAIL |
| Pseudocode | Yes | Algorithm 1 Find adjacent vertices of a basic feasible solution |
| Open Source Code | Yes | All code is implemented in Python, and builds on the Py EPO library [38]. Our code and data are available at https://github.com/ML-KULeuven/Solver-Free-DFL/. |
| Open Datasets | Yes | Multi-dimensional knapsack: The optimization task is a multi-dimensional knapsack problem, as also considered in [25, 28, 38]. ... We use real-world data, taken from a common machine learning benchmark in which the median house prices of districts in California must be predicted from 8 correlated features, including spatial features, features about the districts populations and other aggregate housing statistics [31]. California house prices: This prediction task is taken from a common machine learning benchmark [31], which is offered as a benchmark in the scikit-learn Python package [32]. |
| Dataset Splits | Yes | We assume access to a training set of examples D, sampled from P. ... Reported results are the average taken over 5 independent runs with varying train-validation-test splits. We report the test set performance of the model state that led to the best validation set performance during training. |
| Hardware Specification | Yes | All experiments were conducted on a machine equipped with an Intel(R) Core(TM) i7-1165G7 processor and 16 GB of RAM. |
| Software Dependencies | Yes | All code is implemented in Python, and builds on the Py EPO library [38]. Our code and data are available at https://github.com/ML-KULeuven/Solver-Free-DFL/. We train these models using the Adam optimizer [22]. The scikit-learn Python package [32]. |
| Experiment Setup | Yes | The models are linear regressors, as is common in existing literature [10, 24, 26, 28, 34]. We train these models using the Adam optimizer [22]. All hyperparameters are tuned on independent validation sets prior to training. We train the models until their performance on the validation set has not improved by at least 1% for three checks in a row, or until training time has surpassed 600 seconds (not including evaluation on the validation set), whichever comes first. For LAVA, this training time includes both adjacent vertex precomputation and actual training. We report the test set performance of the model state that led to the best validation set performance during training. When reporting training times, we report the time it took to reach this state. Reported results are the average taken over 5 independent runs with varying train-validation-test splits. ... Considered values for the learning rate were 0.001, 0.01, 0.1 and 1. For all methods, 0.01 performed best. For PFYL, considered values for σ were 0.01, 0.1, 1 and 10. For the random LPs, 0.1 performed best. For the multi-dimensional knapsack problem and the shortest path problem, 1 performed best. A single perturbed cost vector was sampled per backward pass. For NCE, the solve ratio was set to 5%. For Ca VE, both the Ca VE-E and Ca VE+ variants were considered. On random LPs and the shortest path problem, Ca VE-E performed best. On the multi-dimensional knapsack problem, Ca VE+ performed best. For LAVA, ϵ = 0.1 was chosen for all benchmarks. |