Dual Lagrangian Learning for Conic Optimization

Authors: Mathieu Tanneau, Pascal Van Hentenryck

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

Reproducibility Variable Result LLM Response
Research Type Experimental This paper presents Dual Lagrangian Learning (DLL), a principled learning methodology for dual conic optimization proxies. [...] The effectiveness of DLL is demonstrated on linear and nonlinear conic optimization problems. The proposed methodology significantly outperforms a state-of-the-art learning-based method, and achieves 1000x speedups over commercial interior-point solvers with optimality gaps under 0.5% on average. [...] Section 5 reports numerical results.
Researcher Affiliation Academia Mathieu Tanneau, Pascal Van Hentenryck H. Milton Steward School of Industrial and Systems Engineering NSF AI Institute for Advances in Optimization Georgia Institute of Technology {mathieu.tanneau,pascal.vanhentenryck}@isye.gatech.edu
Pseudocode No No explicit pseudocode or algorithm block found.
Open Source Code Yes The code used for experiments is available under an open-source license.1 1https://github.com/AI4OPT/Dual Lagrangian Learning
Open Datasets Yes For each number of items n {100, 200, 500} and number of resources m {5, 10, 30}, a total of 16384 instances are generated using the same procedure as the MIPLearn library [SXQG+23]. [...] This dataset is split in training, validation and testing sets, which contain 8192, 4096 and 4096 instances, respectively.
Dataset Splits Yes This dataset is split in training, validation and testing sets, which contain 8192, 4096 and 4096 instances, respectively.
Hardware Specification Yes All experiments are conducted on the Phoenix cluster [PAC17] with Intel Xeon Gold 6226@2.70GHz + Tesla V100 GPU nodes; each job was allocated 1 GPU, 12 CPU cores and 64GB of RAM.
Software Dependencies Yes All ML models are formulated and trained using Flux [ISF+18]; unless specified otherwise, all (sub)gradients are computed using the auto-differentiation backend Zygote [Inn18]. All linear problems are solved with Gurobi v10 [GO18]. All nonlinear conic problems are solved with Mosek [MOS23b].
Experiment Setup Yes All ML models are trained in a self-supervised fashion following the training scheme outlined in Section 4.3, and training is performed using the Adam optimizer [KB15]. The training scheme uses a patience mechanism where the learning rate η is decreased by a factor 2 if the validation loss does not improve for more than Np epochs. The initial learning rate is η = 10 4. Training is stopped if either the learning rate reaches ηmin = 10 7, or a maximum Nmax epochs is reached. [...] For the output layer, a negated softplus activation ensures y 0. The dual completion procedure follows Example (1). Hyperparameters The patience parameter is Np = 32, and the maximum number of training epochs is Nmax = 1024.