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..

FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees

Authors: Hoang Nguyen, Priya Donti

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments across a range of different optimization problems, including both smooth/nonsmooth and convex/nonconvex problems, demonstrate that FSNet can provide feasible solutions with solution quality comparable to (or in some cases better than) traditional solvers, at significantly faster speeds.
Researcher Affiliation Academia Hoang T. Nguyen Massachusetts Institute of Technology Cambridge, MA, USA EMAIL Priya L. Donti Massachusetts Institute of Technology Cambridge, MA, USA EMAIL
Pseudocode Yes Algorithm 1 FSNet Training Algorithm 1: init. NN weights θ, learning rate 2: repeat 3: sample x D 4: predict yθ(x) via NN 5: compute ˆyθ(x) = FS(yθ(x); x) via (4) 6: update θ with θF(yθ(x), ˆyθ(x)) via (6) 7: until convergence
Open Source Code Yes 1The code is available at https://github.com/MOSSLab-MIT/FSNet
Open Datasets Yes We generate a dataset of 10000 samples for each problem following the procedure in [13, 45]... Finally, we consider the problem of AC optimal power flow (ACOPF)... for this problem, we use the formulation and dataset from [46].
Dataset Splits Yes We generate a dataset of 10000 samples for each problem following the procedure in [13, 45], and split it into 7000 training, 1000 validation, and 2000 test samples for all NN-based methods. For the smooth convex settings, we additionally consider larger-scale problems with 500 decision variables, 200 equality constraints, and 200 inequality constraints. We use 17000 training, 1000 validation, and 2000 test samples for this case.
Hardware Specification Yes Traditional solvers were run on 1 Intel(R) Xeon(R) CPU E5-2670 v2 and all NN methods were run on 1 NVIDIA Quadro RTX 8000 GPU.
Software Dependencies No We use OSQP [47] and SCS [48] for convex problems, and use IPOPT [49] for nonconvex problems... the Projection method in the convex QP setting, which orthogonally projects the NN s prediction onto the feasible set using the qpth solver [20].
Experiment Setup Yes All network, training, and solver-related hyperparameters are in Appendix C.