haginile / SwiftOptimizer

A Swift Library for Solving Optimization Problems (i.e., minimization/maximization)

SwiftOptimizer

SwiftOptimizer allows you to solve minimization/maximization problems in Apple's Swift programming language. It is ported from QuantLib and uses the awesome swix library for matrix calculations.

It currently supports the Simplex and BFGS methods, but will be expanded to include least squares, etc.

Example

First things first, subclass CostFunction to create a class representing the function you are trying to minimize. For example, if you are interested in minimizing the Rosenbrock Function, then you need to set up the cost function as follows:

class RosenBrockFunction: CostFunction
{
    override func value(parameters: matrix) -> Double {
        return pow(1.0 - parameters[0], 2) + 100 * pow(parameters[1] - pow(parameters[0], 2), 2.0)
    }
}

The CostFunction, Constraint (if any), and the initial values together define the Problem you are trying to solve. You also need to specify the EndCriteria so that the optimizer knows when to quit:

var costFunction = RosenBrockFunction()
var constraint = NoConstraint()
var initialValue = zeros(2)
var problem = Problem(costFunction: costFunction, constraint: constraint, initialValue: initialValue)

var myEndCriteria = EndCriteria(maxIterations: 1000, 
                                maxStationaryStateIterations: 100, 
                                rootEpsilon: 1.0e-8, 
                                functionEpsilon: 1.0e-9, 
                                gradientNormEpsilon: 1.0e-5)

Finally, this is how you run the Simplex optimizer:

var solver = Simplex(lambda: 0.1)
var solved = solver.minimize(&problem, endCriteria: myEndCriteria)
problem.currentValue    // return matrix([1.000, 1.000])

Other optimization algorithms can be applied analogously. For example, this is how to use the BFGS algorithm:

var bfgsSolver = BFGS()
var bfgsSolved = bfgsSolver.minimize(&problem, endCriteria: myEndCriteria)
problem.currentValue    // return matrix([1.000, 1.000])