Dodoni.MathLibrary.Native.NLopt

1. Overview

Provides wrapper for the NLopt library, a free/open-source library for nonlinear optimization, established by Steven G. Johnson. You need to download at least a precompiled release of the library. Copy the file(s) to the binary folder of the Dodoni.net project and rename it to libNLopt.dll.

2. Dependencies

This assembly depends on

3. Main concepts and helpful code snippets

NLoptPtr

The class NLoptPtr serves as wrapper for the internal opaque pointer of the algorithm of the NLopt library. Therefore you will find for many function of the API of the NLopt library a counterpart in this class.

 NLoptPtr ptr = new NLoptPtr(NLoptAlgorithm.LN_PRAXIS, 2);

 ptr.TrySetAbsoluteFValueTolerance(1E-6);
 ptr.TrySetRelativeFValueTolerance(1E-6);
 ptr.TrySetAbsoluteXTolerance(1E-6);
 ptr.TrySetRelativeXTolerance(1E-6);

 ptr.SetFunction((n, x, gradient, data) => { return x[0] * x[0] + x[1] * x[1] + 1.123; });

 var argMin = new double[2]{ 1.0, 4.8 };
 var errorCode = ptr.FindMinimum(argMin, out double minimum); // expected minimum: 1.123

NLoptMultiDimOptimizer

In contrast to NLoptPtr, the class NLoptMultiDimOptimizer implements the more common class (Ordinary)MultiDimOptimizer of Dodoni.BasicMathLibrary which serves as infastructure for multi-dimensional optimization in general. Therefore one first create a specific (NLopt)MultiDimOptimizer object. Constraints and the representation of the objective function are specific for the NLopt library. Therefore one has to apply the specific factory for it, for example shown in the following code snippet:

 var cobyla = new NLoptMultiDimOptimizer(NLoptAlgorithm.LN_COBYLA);
 var nloptBoxConstraint = cobyla.Constraint.Create(
                           MultiDimRegion.Interval.Create(dimension: 2, 
                           lowerBounds: new[]{ 1.0, 5.0 }, 
                           upperBounds: new[]{ 12.4, 34.2 }));

 var alg = cobyla.Create(nloptBoxConstraint);
 alg.Function = cobyla.Function.Create(2,x => x[0]*x[0] + x[1]*x[1] + 1.123);
// alternative: alg.SetFunction(x => x[0]*x[0] + x[1]*x[1] + 1.123);

 var argMin = new []{ 1.4, 5.8 };
 var errorCode = alg.FindMinimum(argMin, out double minimum);

Moreover a logging can be used via an optional argument of NLoptMultiDimOptimizer.

You need more freedom in the adjustment of the specific algorithm? The constructors of NLoptMultiDimOptimizer provides an optional argument that is a delegate (function pointer) which will be applied to the internal NLoptPtr object. Therefore one can apply lambda calculus, for example to establish a local optimizer, set initial step size etc.

 var cobyla = new NLoptMultiDimOptimizer(NLoptAlgorithm.LN_COBYLA,
      nloptPtr => {nloptPtr.SetInitialStepSize(new []{1.0, 2.0}} );

Keep in mind that the lambda expression will be evaluated in the Create method of NLoptMultiDimOptimizer, in particular the dimension of the feasible set is unknown in the constructor of the NLoptMultiDimOptimizer class, but the nloptPtr object in the lambda expresion above contains a property with the specific dimension.

NLoptConfiguration If you need additional information of a specific NLopt algorithm, i.e. whether it is a local/global approach, required gradient etc. you can use the Configuration property of NLoptMultiDimOptimizer. Moreover you can create such a configuration separately, for example via

 var config = NLoptConfiguration.Create(NLoptAlgorithm.LN_COBYLA);

Tip: One may have a look in the unit test project of Dodoni.MathLibrary.Native.NLopt; see API documentation.

4. Important Remarks

It has been tested with the 32-bit Windows dll, release 2.4.1.