bobyqa-java

Derivative-free, bounds-constrained optimizer in pure Java — this is a port of PyBOBYQA with extensions for expensive, stochastic, and multiplicatively-noisy objectives encountered in real-world physical systems and with optimizations for the operational performance.

What's new vs. PyBOBYQA

• Warm-start — initialise the surrogate from prior (x, f(x)) data or a previous ModelSnapshot, skipping the npt interpolation evaluations. Cheap resume from a paused or external run.
• Noise handling — multi-sample averaging plus conservative trust-region defaults activated by noise.objfunHasNoise=true.
• Restart auto-detection — regression on rolling history windows of trust radius, gradient change, and Hessian change.
• GoalType.MAXIMIZE first-class — objective is negated internally; result.f() returns the original (un-negated) value.
• Pure Java — EJML is the sole runtime dependency. No JNI, no native libs.

Installation

For v1.0.1, clone-and-build:

git clone https://github.com/fair-acc/bobyqa-java.git
cd bobyqa-java
mvn clean install

Maven Central coordinates: arriving in v1.1 as io.fair-acc:bobyqa-java.

Quick start

import bobyqa.Bobyqa;
import bobyqa.GoalType;
import bobyqa.OptimResult;
import bobyqa.Params;

double[] x0    = { -1.2,  1.0 };
double[] lower = { -5.0, -5.0 };
double[] upper = {  5.0,  5.0 };

OptimResult result = Bobyqa.solve(
        x -> Math.pow(1.0 - x[0], 2) + 100.0 * Math.pow(x[1] - x[0] * x[0], 2),
        x0, lower, upper,
        2000,
        GoalType.MINIMIZE,
        Params.defaults(2));

System.out.printf("x = %s, f = %.3e, evals = %d, exit = %s%n",
        java.util.Arrays.toString(result.x()),
        result.f(),
        result.nEvals(),
        result.exitFlag());

Warm-start usage

Use prior data when you have observations from an external run (manual measurements, parallel sweeps, etc.):

import bobyqa.WarmStart;
import java.util.List;

List<double[]> priorXValues = List.of(/* sample X-values from previous runs, one double[] per sample */);
List<Double>   priorMeas = List.of(/* corresponding response values from previous runs */);
WarmStart      ws     = WarmStart.fromPriorData(priorXValues, priorMeas);

OptimResult result = Bobyqa.solve(
        objective, x0, lower, upper, maxEvals, GoalType.MINIMIZE,
        Params.defaults(n), ws);

Use a model snapshot when you are resuming a paused optimisation in the same process:

ModelSnapshot modelSnapshot = previousResult.snapshot();   // captured at the end of a previous solve()
WarmStart    ws        = WarmStart.fromSnapshot(modelSnapshot);

OptimResult result = Bobyqa.solve(
        objective, x0, lower, upper, maxEvals, GoalType.MAXIMIZE,
        Params.defaults(n), ws);

Prior data performs a hybrid initialisation (top-k points + fresh axis-aligned fill). Snapshots perform a zero-evaluation restore.

Documentation

The full technical design is in docs/design/bobyqa-java-port-design.md.

Attribution

This is a Java port of PyBOBYQA by the Numerical Algorithms Group (NAG), itself a Python implementation of BOBYQA (Bound Optimization BY Quadratic Approximation) by M. J. D. Powell (2009).

Disclaimer: This software has been authored using AI, namely CLAUDE Code (version(s) 4.6 and following).

License

GPL-3.0-or-later. See LICENSE and NOTICE.

Citation

Machine-readable citation metadata is provided in CITATION.cff — GitHub renders a "Cite this repository" button on the repo homepage from it (also consumed by Zotero, Zenodo, and most citation managers). For manual citation, the BibTeX entries are:

@misc{bobyqa-java,
  author       = {Geithner, Wolfgang},
  title        = {{bobyqa-java}: a Java port of PyBOBYQA},
  year         = {2026},
  publisher    = {GSI Helmholtzzentrum für Schwerionenforschung GmbH},
  howpublished = {\url{https://github.com/fair-acc/bobyqa-java}}
}

@techreport{powell2009bobyqa,
  author      = {Powell, M.J.D.},
  title       = {The {BOBYQA} algorithm for bound constrained optimization without derivatives},
  institution = {DAMTP, University of Cambridge},
  number      = {NA2009/06},
  year        = {2009}
}