NLPOpt-Net

NonLinear Parametric Optimization Network (NLPOpt-Net) is a GPU supported python package for learning parameters to solution mappings for nonlinear parametric optimization problem.

Installation

Currently, NLPOpt-Net requires Python 3.9 or later and supports installation for both CPU-only and GPU-accelerated environments.

• CPU-only (Linux/macOS/Windows)

pip install nlpoptnet

• GPU (NVIDIA, CUDA 12)

pip install "nlpoptnet[cuda12]"

Quick Overview

To give a quick understanding of what we are doing, we are trying to predict a solution ($y$) of the parameterized optimization problem, $$ \begin{aligned} \quad \min \quad & f(x,y) \ \text{s.t.} \quad & h(x,y) = 0, \ & g(x,y) \le 0, \ & l(x) \le y \le u(x), \end{aligned} $$

for a given set of parameter ($x$). The core idea is to train a machine learning model in an unsupervised way and predict new solutions. For feasibility guarantee a projection layer is employed. Particularly, we approximate the original problem into a specific structure of quadratic program and solve that to retain feasibility. Graphically, the projection as training progresses would look like this:

For a detailed idea, we refer readers to this article. Complete user documentation can be found here.

Core API

• NLPOptNet(config=..., type=...) is the user-facing entry point.
• model.extract(problem.npz) loads structured matrices and exposes them as model attributes like model.Q, model.A, model.b, and so on.
• model.dataset(...), model.simplex(...), and model.box(...) define the parameter region.
• model.constraints.box.add(...) is separate from general inequalities so the bound constraints stay on the dedicated projection path.
• model.optimize() trains and returns a result dictionary with output_dir, metadata_path, summary, and history.
• model.load(metadata_path) restores a trained model, and model.predict(x_value) returns projected variable predictions.

General Workflow

The core ideas can be generalized in a workflow for training and inference tasks.

Model Creation and Training

Following is a general pseudocode for building and running a problem with NLPOpt-Net.

from nlpoptnet import NLPOptNet 

CONFIG = {...}

model = NLPOptNet(config=CONFIG, name="problem_name")

x = model.add_parameter([...])
y = model.add_variable([...])

model.extract("path/to/problem_data")
model.objective(...)

model.constraints.equality.add(...)
model.constraints.inequality.add(...)
model.constraints.box.add(...)

model.dataset(parameters="path/to/parameters.csv")

model.build()
result = model.optimize()

run_dir = result["output_dir"]

Load and Use a Trained Model

After training, a saved model can be reloaded to inspect the run summary, visualize the training history, and make new predictions.

reloaded = NLPOptNet().load(metadata_path)

reloaded.summary()
reloaded.plot_history()

sample_pred = reloaded.predict(sample_x, projection_backend="backend/type/for/prediction")

Example

Following is an example of the summary and history plot for a trained model.

model.summary()

📊 NLPOptNet Summary
------------------------------------------------------------
Model Name                        : example_model
No. of Parameters                 : 50
No. of Variables                  : 100
No. of Equalities                 : 50
No. of Inequalities               : 50
No. of Train Samples              : 1000
No. of Validation Samples         : 1000
Maximum Constraint Violation      : 7.5433e-11
Training Time                     : 1518.25 s
Est. JAX Single Inference Time    : 91.49 ms
Est. Native Single Inference Time : 1.99 ms
Est. JAX Batch Inference Time     : 140.02 ms
------------------------------------------------------------

model.plot_history()

Reporting a Bug or Error

We understand that running into issues can be frustrating. If you experience any errors while using NLPOpt-Net, please let us know so we can address them in future updates.

When reporting a bug, it is helpful to include:

• A brief description of the issue
• Relevant code snippets
• Error messages or logs
• Steps to reproduce the problem

Your feedback is greatly appreciated and helps us improve the package.