public void ConstructorTest2()
{
var function = new NonlinearObjectiveFunction(2, x => x[0] * x[1]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(function, x => 1.0 - x[0] * x[0] - x[1] * x[1])
};
Cobyla cobyla = new Cobyla(function, constraints);
for (int i = 0; i < cobyla.Solution.Length; i++)
cobyla.Solution[i] = 1;
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
double sqrthalf = Math.Sqrt(0.5);
Assert.AreEqual(-0.5, minimum, 1e-10);
Assert.AreEqual(sqrthalf, solution[0], 1e-5);
Assert.AreEqual(-sqrthalf, solution[1], 1e-5);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
public void QuadraticConstraintConstructorTest()
{
IObjectiveFunction objective = null;
double[,] quadraticTerms =
{
{ 1, 2, 3 },
{ 4, 5, 6 },
{ 7, 8, 9 },
};
double[] linearTerms = { 1, 2, 3 };
objective = new NonlinearObjectiveFunction(3, f => f[0] + f[1] + f[2]);
QuadraticConstraint target = new QuadraticConstraint(objective,
quadraticTerms, linearTerms,
ConstraintType.LesserThanOrEqualTo, 0);
var function = target.Function;
var gradient = target.Gradient;
FiniteDifferences fd = new FiniteDifferences(3, function);
double[][] x =
{
new double[] { 1, 2, 3 },
new double[] { 3, 1, 4 },
new double[] { -6 , 5, 9 },
new double[] { 31, 25, 246 },
new double[] { -0.102, 0, 10 },
};
{ // Function test
for (int i = 0; i < x.Length; i++)
{
double expected =
(x[i].Multiply(quadraticTerms)).InnerProduct(x[i])
+ linearTerms.InnerProduct(x[i]);
double actual = function(x[i]);
Assert.AreEqual(expected, actual, 1e-8);
}
}
{ // Gradient test
for (int i = 0; i < x.Length; i++)
{
double[] expected = fd.Compute(x[i]);
double[] actual = gradient(x[i]);
for (int j = 0; j < actual.Length; j++)
Assert.AreEqual(expected[j], actual[j], 1e-8);
}
}
}
/// <summary>
/// Initializes a new instance of the <see cref="BaseOptimizationMethod"/> class.
/// </summary>
///
/// <param name="function">The objective function whose optimum values should be found.</param>
///
protected BaseOptimizationMethod(NonlinearObjectiveFunction function)
{
if (function == null)
throw new ArgumentNullException("function");
init(function.NumberOfVariables);
this.Function = function.Function;
}
public void AugmentedLagrangianSolverConstructorTest1()
{
Accord.Math.Tools.SetupGenerator(0);
// min 100(y-x*x)²+(1-x)²
//
// s.t. x <= 0
// y <= 0
//
var f = new NonlinearObjectiveFunction(2,
function: (x) => 100 * Math.Pow(x[1] - x[0] * x[0], 2) + Math.Pow(1 - x[0], 2),
gradient: (x) => new[]
{
2.0 * (200.0 * x[0]*x[0]*x[0] - 200.0 * x[0] * x[1] + x[0] - 1), // df/dx
200 * (x[1] - x[0]*x[0]) // df/dy
}
);
var constraints = new List<NonlinearConstraint>();
constraints.Add(new NonlinearConstraint(f,
function: (x) => x[0],
gradient: (x) => new[] { 1.0, 0.0 },
shouldBe: ConstraintType.LesserThanOrEqualTo, value: 0
));
constraints.Add(new NonlinearConstraint(f,
function: (x) => x[1],
gradient: (x) => new[] { 0.0, 1.0 },
shouldBe: ConstraintType.LesserThanOrEqualTo, value: 0
));
var solver = new AugmentedLagrangian(f, constraints);
Assert.IsTrue(solver.Minimize());
double minValue = solver.Value;
Assert.IsFalse(Double.IsNaN(minValue));
Assert.AreEqual(1, minValue, 1e-5);
Assert.AreEqual(0, solver.Solution[0], 1e-5);
Assert.AreEqual(0, solver.Solution[1], 1e-5);
}
public void ConstructorTest4()
{
var function = new NonlinearObjectiveFunction(2, x =>
Math.Pow(x[0] * x[0] - x[1], 2.0) + Math.Pow(1.0 + x[0], 2.0));
NelderMead solver = new NelderMead(function);
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
Assert.AreEqual(0, minimum, 1e-10);
Assert.AreEqual(-1, solution[0], 1e-5);
Assert.AreEqual(1, solution[1], 1e-4);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
public void ConstructorTest4()
{
// Weak version of Rosenbrock's problem.
var function = new NonlinearObjectiveFunction(2, x =>
Math.Pow(x[0] * x[0] - x[1], 2.0) + Math.Pow(1.0 + x[0], 2.0));
Subplex solver = new Subplex(function);
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
Assert.AreEqual(2, solution.Length);
Assert.AreEqual(0, minimum, 1e-10);
Assert.AreEqual(-1, solution[0], 1e-5);
Assert.AreEqual(1, solution[1], 1e-4);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
// We don't use this at the moment
static List<NonlinearConstraint> CreateConstraints (Parameter[] x, NonlinearObjectiveFunction f)
{
// Now we can start stating the constraints
var nlConstraints = x.SelectMany ((p, i) => {
Func<double[], double> cfn = args => x [i].Value;
return new[] {
new NonlinearConstraint
( f
, function: cfn
, shouldBe: ConstraintType.GreaterThanOrEqualTo
, value: p.Min
, gradient: Grad (x.Length, cfn)),
new NonlinearConstraint
( f
, function: cfn
, shouldBe: ConstraintType.LesserThanOrEqualTo
, value: p.Max
, gradient: Grad (x.Length, cfn)),
};
}).ToList ();
return nlConstraints;
}
public void ConstructorTest2()
{
Accord.Math.Tools.SetupGenerator(0);
var function = new NonlinearObjectiveFunction(2,
function: x => x[0] * x[1],
gradient: x => new[] { x[1], x[0] });
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(function,
function: x => 1.0 - x[0] * x[0] - x[1] * x[1],
gradient: x => new [] { -2 * x[0], -2 * x[1]}),
new NonlinearConstraint(function,
function: x => x[0],
gradient: x => new [] { 1.0, 0.0}),
};
var target = new ConjugateGradient(2);
AugmentedLagrangian solver = new AugmentedLagrangian(target, function, constraints);
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
double sqrthalf = Math.Sqrt(0.5);
Assert.AreEqual(-0.5, minimum, 1e-5);
Assert.AreEqual(sqrthalf, solution[0], 1e-5);
Assert.AreEqual(-sqrthalf, solution[1], 1e-5);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
private static void test2(IGradientOptimizationMethod inner)
{
// maximize 2x + 3y, s.t. 2x² + 2y² <= 50
//
// http://www.wolframalpha.com/input/?i=max+2x+%2B+3y%2C+s.t.+2x%C2%B2+%2B+2y%C2%B2+%3C%3D+50
// Max x' * c
// x
// s.t. x' * A * x <= k
// x' * i = 1
// lower_bound < x < upper_bound
double[] c = { 2, 3 };
double[,] A = { { 2, 0 }, { 0, 2 } };
double k = 50;
// Create the objective function
var objective = new NonlinearObjectiveFunction(2,
function: (x) => x.InnerProduct(c),
gradient: (x) => c
);
// Test objective
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i * 2 + j * 3;
double actual = objective.Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
// Create the optimization constraints
var constraints = new List<NonlinearConstraint>();
constraints.Add(new QuadraticConstraint(objective,
quadraticTerms: A,
shouldBe: ConstraintType.LesserThanOrEqualTo, value: k
));
// Test first constraint
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
var input = new double[] { i, j };
double expected = i * (2 * i + 0 * j) + j * (0 * i + 2 * j);
double actual = constraints[0].Function(input);
Assert.AreEqual(expected, actual);
}
}
// Create the solver algorithm
AugmentedLagrangian solver =
new AugmentedLagrangian(inner, objective, constraints);
Assert.AreEqual(inner, solver.Optimizer);
Assert.IsTrue(solver.Maximize());
double maxValue = solver.Value;
Assert.AreEqual(18.02, maxValue, 1e-2);
Assert.AreEqual(2.77, solver.Solution[0], 1e-2);
Assert.AreEqual(4.16, solver.Solution[1], 1e-2);
}
public void ConstructorTest6_3()
{
bool thrown = false;
try
{
var function = new NonlinearObjectiveFunction(2, x => -x[0] - x[1]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(2, x => x[1] - x[0] * x[0]),
new NonlinearConstraint(4, x => 1.0 - x[0] * x[0] - x[1] * x[1]),
};
Cobyla cobyla = new Cobyla(function, constraints);
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
}
catch (Exception)
{
thrown = true;
}
Assert.IsTrue(thrown);
}
public void ConstructorTest7()
{
/// This problem is taken from Fletcher's book Practical Methods of
/// Optimization and has the equation number (14.4.2).
var function = new NonlinearObjectiveFunction(3, x => x[2]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(3, x=> 5.0 * x[0] - x[1] + x[2]),
new NonlinearConstraint(3, x => x[2] - x[0] * x[0] - x[1] * x[1] - 4.0 * x[1]),
new NonlinearConstraint(3, x => x[2] - 5.0 * x[0] - x[1]),
};
Cobyla cobyla = new Cobyla(function, constraints);
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
Assert.AreEqual(-3, minimum, 1e-5);
Assert.AreEqual(0.0, solution[0], 1e-5);
Assert.AreEqual(-3.0, solution[1], 1e-5);
Assert.AreEqual(-3.0, solution[2], 1e-5);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
/// <summary>
/// Creates a new <see cref="ResilientBackpropagation"/> function optimizer.
/// </summary>
///
/// <param name="function">The function to be optimized.</param>
///
public ResilientBackpropagation(NonlinearObjectiveFunction function)
: this(function.NumberOfVariables, function.Function, function.Gradient)
{
}
/// <summary>
/// Creates a new instance of the L-BFGS optimization algorithm.
/// </summary>
///
/// <param name="function">The function to be optimized.</param>
///
public BroydenFletcherGoldfarbShanno(NonlinearObjectiveFunction function)
: this(function.NumberOfVariables, function.Function, function.Gradient)
{
}
public void SubspaceTest1()
{
var function = new NonlinearObjectiveFunction(5, x =>
10.0 * Math.Pow(x[0] * x[0] - x[1], 2.0) + Math.Pow(1.0 + x[0], 2.0));
NelderMead solver = new NelderMead(function);
solver.NumberOfVariables = 2;
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
Assert.AreEqual(5, solution.Length);
Assert.AreEqual(-0, minimum, 1e-6);
Assert.AreEqual(-1, solution[0], 1e-3);
Assert.AreEqual(+1, solution[1], 1e-3);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
/// <summary>
/// Creates a new instance of the Augmented Lagrangian algorithm.
/// </summary>
///
/// <param name="function">The objective function to be optimized.</param>
/// <param name="constraints">
/// The <see cref="NonlinearConstraint"/>s to which the solution must be subjected.</param>
///
public AugmentedLagrangian(NonlinearObjectiveFunction function, IEnumerable<NonlinearConstraint> constraints)
: base(function.NumberOfVariables)
{
init(function, constraints, null);
}
/// <summary>
/// Creates a new instance of the Augmented Lagrangian algorithm.
/// </summary>
///
/// <param name="function">The objective function to be optimized.</param>
/// <param name="constraints">
/// The <see cref="NonlinearConstraint"/>s to which the solution must be subjected.</param>
///
public AugmentedLagrangian(NonlinearObjectiveFunction function, IEnumerable <NonlinearConstraint> constraints)
: base(function.NumberOfVariables)
{
init(function, constraints, null);
}
public void ConstructorTest5()
{
var function = new NonlinearObjectiveFunction(2, x =>
10.0 * Math.Pow(x[0] * x[0] - x[1], 2.0) + Math.Pow(1.0 + x[0], 2.0));
Subplex solver = new Subplex(function);
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
Assert.AreEqual(-0, minimum, 1e-6);
Assert.AreEqual(-1, solution[0], 1e-3);
Assert.AreEqual(+1, solution[1], 1e-3);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
public void ConstructorTest8()
{
/// This problem is taken from page 66 of Hock and Schittkowski's book Test
/// Examples for Nonlinear Programming Codes. It is their test problem Number
/// 43, and has the name Rosen-Suzuki.
var function = new NonlinearObjectiveFunction(4, x => x[0] * x[0]
+ x[1] * x[1] + 2.0 * x[2] * x[2]
+ x[3] * x[3] - 5.0 * x[0] - 5.0 * x[1]
- 21.0 * x[2] + 7.0 * x[3]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(4, x=> 8.0 - x[0] * x[0]
- x[1] * x[1] - x[2] * x[2] - x[3] * x[3] - x[0] + x[1] - x[2] + x[3]),
new NonlinearConstraint(4, x => 10.0 - x[0] * x[0]
- 2.0 * x[1] * x[1] - x[2] * x[2] - 2.0 * x[3] * x[3] + x[0] + x[3]),
new NonlinearConstraint(4, x => 5.0 - 2.0 * x[0] * x[0]
- x[1] * x[1] - x[2] * x[2] - 2.0 * x[0] + x[1] + x[3])
};
Cobyla cobyla = new Cobyla(function, constraints);
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
double[] expected =
{
0.0, 1.0, 2.0, -1.0
};
for (int i = 0; i < expected.Length; i++)
Assert.AreEqual(expected[i], cobyla.Solution[i], 1e-4);
Assert.AreEqual(-44, minimum, 1e-10);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
/// <summary>
/// Creates a new instance of the L-BFGS optimization algorithm.
/// </summary>
///
/// <param name="function">The function to be optimized.</param>
///
public ResilientBackpropagation(NonlinearObjectiveFunction function)
: this(function.NumberOfVariables, function.Function, function.Gradient)
{
}
public void ConstructorTest9()
{
/// This problem is taken from page 111 of Hock and Schittkowski's
/// book Test Examples for Nonlinear Programming Codes. It is their
/// test problem Number 100.
///
var function = new NonlinearObjectiveFunction(7, x =>
Math.Pow(x[0] - 10.0, 2.0) + 5.0 * Math.Pow(x[1] - 12.0, 2.0) + Math.Pow(x[2], 4.0) +
3.0 * Math.Pow(x[3] - 11.0, 2.0) + 10.0 * Math.Pow(x[4], 6.0) + 7.0 * x[5] * x[5] + Math.Pow(x[6], 4.0) -
4.0 * x[5] * x[6] - 10.0 * x[5] - 8.0 * x[6]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(7, x => 127.0 - 2.0 * x[0] * x[0] - 3.0 * Math.Pow(x[1], 4.0)
- x[2] - 4.0 * x[3] * x[3] - 5.0 * x[4]),
new NonlinearConstraint(7, x => 282.0 - 7.0 * x[0] - 3.0 * x[1] - 10.0 * x[2] * x[2] - x[3] + x[4]),
new NonlinearConstraint(7, x => 196.0 - 23.0 * x[0] - x[1] * x[1] - 6.0 * x[5] * x[5] + 8.0 * x[6]),
new NonlinearConstraint(7, x => -4.0 * x[0] * x[0] - x[1] * x[1] + 3.0 * x[0] * x[1]
- 2.0 * x[2] * x[2] - 5.0 * x[5] + 11.0 * x[6])
};
Cobyla cobyla = new Cobyla(function, constraints);
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
double[] expected =
{
2.330499, 1.951372, -0.4775414, 4.365726, -0.624487, 1.038131, 1.594227
};
for (int i = 0; i < expected.Length; i++)
Assert.AreEqual(expected[i], cobyla.Solution[i], 1e-4);
Assert.AreEqual(680.63005737443393, minimum, 1e-6);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
public void ConstructorTest3()
{
// minimize f(x) = x*y*z,
// s.t.
//
// 1 - x² - 2y² - 3z² > 0
// x > 0,
// y > 0
//
// Easy three dimensional minimization in ellipsoid.
var function = new NonlinearObjectiveFunction(3,
function: x => x[0] * x[1] * x[2],
gradient: x => new[] { x[1] * x[2], x[0] * x[2], x[0] * x[1] });
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(3,
function: x => 1.0 - x[0] * x[0] - 2.0 * x[1] * x[1] - 3.0 * x[2] * x[2],
gradient: x => new[] { -2.0 * x[0], -4.0 * x[1], -6.0 * x[2] }),
new NonlinearConstraint(3,
function: x => x[0],
gradient: x => new[] { 1.0, 0, 0 }),
new NonlinearConstraint(3,
function: x => x[1],
gradient: x => new[] { 0, 1.0, 0 }),
new NonlinearConstraint(3,
function: x => -x[2],
gradient: x => new[] { 0, 0, -1.0 }),
};
for (int i = 0; i < constraints.Length; i++)
{
Assert.AreEqual(ConstraintType.GreaterThanOrEqualTo, constraints[i].ShouldBe);
Assert.AreEqual(0, constraints[i].Value);
}
var inner = new BroydenFletcherGoldfarbShanno(3);
inner.LineSearch = LineSearch.BacktrackingArmijo;
inner.Corrections = 10;
var solver = new AugmentedLagrangian(inner, function, constraints);
Assert.AreEqual(inner, solver.Optimizer);
Assert.IsTrue(solver.Minimize());
double minimum = solver.Value;
double[] solution = solver.Solution;
double[] expected =
{
1.0 / Math.Sqrt(3.0), 1.0 / Math.Sqrt(6.0), -1.0 / 3.0
};
for (int i = 0; i < expected.Length; i++)
Assert.AreEqual(expected[i], solver.Solution[i], 1e-3);
Assert.AreEqual(-0.078567420132031968, minimum, 1e-4);
double expectedMinimum = function.Function(solver.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
private void init(NonlinearObjectiveFunction function,
IEnumerable<NonlinearConstraint> constraints, IGradientOptimizationMethod innerSolver)
{
if (function != null)
{
if (function.NumberOfVariables != NumberOfVariables)
{
throw new ArgumentOutOfRangeException("function",
"Incorrect number of variables in the objective function. " +
"The number of variables must match the number of variables set in the solver.");
}
this.Function = function.Function;
this.Gradient = function.Gradient;
}
if (innerSolver == null)
{
innerSolver = new BroydenFletcherGoldfarbShanno(NumberOfVariables)
{
LineSearch = Optimization.LineSearch.BacktrackingArmijo,
Corrections = 10
};
}
List<NonlinearConstraint> equality = new List<NonlinearConstraint>();
List<NonlinearConstraint> lesserThan = new List<NonlinearConstraint>();
List<NonlinearConstraint> greaterThan = new List<NonlinearConstraint>();
foreach (var c in constraints)
{
switch (c.ShouldBe)
{
case ConstraintType.EqualTo:
equality.Add(c); break;
case ConstraintType.GreaterThanOrEqualTo:
greaterThan.Add(c); break;
case ConstraintType.LesserThanOrEqualTo:
lesserThan.Add(c); break;
default:
throw new ArgumentException("Unknown constraint type.", "constraints");
}
}
this.lesserThanConstraints = lesserThan.ToArray();
this.greaterThanConstraints = greaterThan.ToArray();
this.equalityConstraints = equality.ToArray();
this.lambda = new double[equalityConstraints.Length];
this.mu = new double[lesserThanConstraints.Length];
this.nu = new double[greaterThanConstraints.Length];
this.dualSolver = innerSolver;
dualSolver.Function = objectiveFunction;
dualSolver.Gradient = objectiveGradient;
}
public void ConstructorTest10()
{
/// This problem is taken from page 415 of Luenberger's book Applied
/// Nonlinear Programming. It is to maximize the area of a hexagon of
/// unit diameter.
///
var function = new NonlinearObjectiveFunction(9, x =>
-0.5 * (x[0] * x[3] - x[1] * x[2] + x[2] * x[8]
- x[4] * x[8] + x[4] * x[7] - x[5] * x[6]));
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(9, x => 1.0 - x[2] * x[2] - x[3] * x[3]),
new NonlinearConstraint(9, x => 1.0 - x[8] * x[8]),
new NonlinearConstraint(9, x => 1.0 - x[4] * x[4] - x[5] * x[5]),
new NonlinearConstraint(9, x => 1.0 - x[0] * x[0] - Math.Pow(x[1] - x[8], 2.0)),
new NonlinearConstraint(9, x => 1.0 - Math.Pow(x[0] - x[4], 2.0) - Math.Pow(x[1] - x[5], 2.0)),
new NonlinearConstraint(9, x => 1.0 - Math.Pow(x[0] - x[6], 2.0) - Math.Pow(x[1] - x[7], 2.0)),
new NonlinearConstraint(9, x => 1.0 - Math.Pow(x[2] - x[4], 2.0) - Math.Pow(x[3] - x[5], 2.0)),
new NonlinearConstraint(9, x => 1.0 - Math.Pow(x[2] - x[6], 2.0) - Math.Pow(x[3] - x[7], 2.0)),
new NonlinearConstraint(9, x => 1.0 - x[6] * x[6] - Math.Pow(x[7] - x[8], 2.0)),
new NonlinearConstraint(9, x => x[0] * x[3] - x[1] * x[2]),
new NonlinearConstraint(9, x => x[2] * x[8]),
new NonlinearConstraint(9, x => -x[4] * x[8]),
new NonlinearConstraint(9, x => x[4] * x[7] - x[5] * x[6]),
new NonlinearConstraint(9, x => x[8]),
};
Cobyla cobyla = new Cobyla(function, constraints);
for (int i = 0; i < cobyla.Solution.Length; i++)
cobyla.Solution[i] = 1;
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
double[] expected =
{
0.688341, 0.725387, -0.284033, 0.958814, 0.688341, 0.725387, -0.284033, 0.958814, 0.0
};
for (int i = 0; i < expected.Length; i++)
Assert.AreEqual(expected[i], cobyla.Solution[i], 1e-2);
Assert.AreEqual(-0.86602540378486847, minimum, 1e-10);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}
/// <summary>
/// Maximizes the given function.
/// </summary>
///
/// <param name="function">The function to be maximized.</param>
///
/// <returns>The maximum value found at the <see cref="Solution"/>.</returns>
///
public double Maximize(NonlinearObjectiveFunction function)
{
if (function.NumberOfVariables != numberOfVariables)
throw new ArgumentOutOfRangeException("function",
"Incorrect number of variables in the objective function. " +
"The number of variables must match the number of variables set in the solver.");
this.Function = x => -function.Function(x);
this.Gradient = x => function.Gradient(x).Multiply(-1);
minimize();
return -Function(Solution);
}
public void ConstructorTest4()
{
// Example code from
// https://groups.google.com/forum/#!topic/accord-net/an0sJGGrOuU
int nVariablesTest = 4; // number of variables
int nConstraintsTest = 2; // number of constraints
double constraintsTolerance = 1e-100;
double[,] ATest = new double[,] { { 1, 2, 3, 4 }, { 0, 4, 3, 1 } }; // arbitary A matrix. A*X = b
double[,] bTest = new double[,] { { 0 }, { 2 } }; // arbitary A matrix. A*X = b
double[,] XSolve = ATest.Solve(bTest); // uses the pseudoinverse to minimise norm(X) subject to A*X = b
// recreate Solve function using AugmentedLagrangian
var fTest = new NonlinearObjectiveFunction(nVariablesTest, ds => ds.InnerProduct(ds), ds => ds.Multiply(2.0)); // minimise norm(ds)
var nonlinearConstraintsTest = new List<NonlinearConstraint>(nConstraintsTest); // linear constraints A*X = b
for (int i = 0; i < nConstraintsTest; i++)
{
int j = i; // http://blogs.msdn.com/b/ericlippert/archive/2009/11/12/closing-over-the-loop-variable-considered-harmful.aspx
nonlinearConstraintsTest.Add(new NonlinearConstraint(fTest, ds => ATest.GetRow(j).InnerProduct(ds) - (double)bTest.GetValue(j, 0), ConstraintType.EqualTo, 0.0, ds => ATest.GetRow(j), constraintsTolerance));
}
var innerSolverTest = new ResilientBackpropagation(nVariablesTest);
innerSolverTest.Tolerance = constraintsTolerance;
innerSolverTest.Iterations = 1000;
var solverTest = new Accord.Math.Optimization.AugmentedLagrangian(innerSolverTest, fTest, nonlinearConstraintsTest);
solverTest.MaxEvaluations = 0;
bool didMinimise = solverTest.Minimize();
var errorConstraintRelative = XSolve.Subtract(solverTest.Solution, 1).ElementwiseDivide(XSolve); // relative error between .Solve and .Minimize
var errorConstraintAbsolute = XSolve.Subtract(solverTest.Solution, 1); // absolute error between .Solve and .Minimize
double[] errorConstraintsTest = new double[nConstraintsTest];
for (int i = 0; i < nConstraintsTest; i++)
{
errorConstraintsTest[i] = nonlinearConstraintsTest[i].Function(solverTest.Solution);
}
}
private void init(NonlinearObjectiveFunction function,
IEnumerable <NonlinearConstraint> constraints, IGradientOptimizationMethod innerSolver)
{
if (function != null)
{
if (function.NumberOfVariables != NumberOfVariables)
{
throw new ArgumentOutOfRangeException("function",
"Incorrect number of variables in the objective function. " +
"The number of variables must match the number of variables set in the solver.");
}
this.Function = function.Function;
this.Gradient = function.Gradient;
}
if (innerSolver == null)
{
innerSolver = new BroydenFletcherGoldfarbShanno(NumberOfVariables)
{
LineSearch = Optimization.LineSearch.BacktrackingArmijo,
Corrections = 10,
Epsilon = 1e-10,
MaxIterations = 100000
};
}
var equality = new List <NonlinearConstraint>();
var lesserThan = new List <NonlinearConstraint>();
var greaterThan = new List <NonlinearConstraint>();
foreach (var c in constraints)
{
switch (c.ShouldBe)
{
case ConstraintType.EqualTo:
equality.Add(c); break;
case ConstraintType.GreaterThanOrEqualTo:
greaterThan.Add(c); break;
case ConstraintType.LesserThanOrEqualTo:
lesserThan.Add(c); break;
default:
throw new ArgumentException("Unknown constraint type.", "constraints");
}
}
this.lesserThanConstraints = lesserThan.ToArray();
this.greaterThanConstraints = greaterThan.ToArray();
this.equalityConstraints = equality.ToArray();
this.lambda = new double[equalityConstraints.Length];
this.mu = new double[lesserThanConstraints.Length];
this.nu = new double[greaterThanConstraints.Length];
this.g = new double[NumberOfVariables];
this.dualSolver = innerSolver;
dualSolver.Function = objectiveFunction;
dualSolver.Gradient = objectiveGradient;
}
public void AugmentedLagrangianSolverConstructorTest2()
{
// min 100(y-x*x)²+(1-x)²
//
// s.t. x >= 0
// y >= 0
//
var f = new NonlinearObjectiveFunction(2,
function: (x) => 100 * Math.Pow(x[1] - x[0] * x[0], 2) + Math.Pow(1 - x[0], 2),
gradient: (x) => new[]
{
2.0 * (200.0 * Math.Pow(x[0], 3) - 200.0 * x[0] * x[1] + x[0] - 1), // df/dx
200 * (x[1] - x[0]*x[0]) // df/dy
}
);
var constraints = new List<NonlinearConstraint>();
constraints.Add(new NonlinearConstraint(f,
function: (x) => x[0],
gradient: (x) => new[] { 1.0, 0.0 },
shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 0
));
constraints.Add(new NonlinearConstraint(f,
function: (x) => x[1],
gradient: (x) => new[] { 0.0, 1.0 },
shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 0
));
var solver = new AugmentedLagrangianSolver(2, constraints);
double minValue = solver.Minimize(f);
Assert.AreEqual(0, minValue, 1e-10);
Assert.AreEqual(1, solver.Solution[0], 1e-10);
Assert.AreEqual(1, solver.Solution[1], 1e-10);
Assert.IsFalse(Double.IsNaN(minValue));
Assert.IsFalse(Double.IsNaN(solver.Solution[0]));
Assert.IsFalse(Double.IsNaN(solver.Solution[1]));
}
/// <summary>
/// Creates a new instance of the Augmented Lagrangian algorithm.
/// </summary>
///
/// <param name="innerSolver">The <see cref="IGradientOptimizationMethod">unconstrained
/// optimization method</see> used internally to solve the dual of this optimization
/// problem.</param>
/// <param name="function">The objective function to be optimized.</param>
/// <param name="constraints">
/// The <see cref="NonlinearConstraint"/>s to which the solution must be subjected.</param>
///
public AugmentedLagrangian(IGradientOptimizationMethod innerSolver,
NonlinearObjectiveFunction function, IEnumerable<NonlinearConstraint> constraints)
: base(innerSolver.NumberOfVariables)
{
if (innerSolver.NumberOfVariables != function.NumberOfVariables)
throw new ArgumentException("The inner unconstrained optimization algorithm and the "
+ "objective function should have the same number of variables.", "function");
init(function, constraints, innerSolver);
}
public void AugmentedLagrangianSolverConstructorTest3()
{
// min x*y+ y*z
//
// s.t. x^2 - y^2 + z^2 - 2 >= 0
// x^2 + y^2 + z^2 - 10 <= 0
//
double x = 0, y = 0, z = 0;
var f = new NonlinearObjectiveFunction(
function: () => x * y + y * z,
gradient: () => new[]
{
y, // df/dx
x + z, // df/dy
y, // df/dz
}
);
var constraints = new List<NonlinearConstraint>();
constraints.Add(new NonlinearConstraint(f,
function: () => x * x - y * y + z * z,
gradient: () => new[] { 2 * x, -2 * y, 2 * z },
shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 2
));
constraints.Add(new NonlinearConstraint(f,
function: () => x * x + y * y + z * z,
gradient: () => new[] { 2 * x, 2 * y, 2 * z },
shouldBe: ConstraintType.LesserThanOrEqualTo, value: 10
));
var solver = new AugmentedLagrangianSolver(3, constraints);
solver.Solution[0] = 1;
solver.Solution[1] = 1;
solver.Solution[2] = 1;
double minValue = solver.Minimize(f);
Assert.AreEqual(-6.9, minValue, 1e-1);
Assert.AreEqual(+1.73, solver.Solution[0], 1e-2);
Assert.AreEqual(-2.00, solver.Solution[1], 1e-2);
Assert.AreEqual(+1.73, solver.Solution[2], 1e-2);
Assert.IsFalse(Double.IsNaN(minValue));
Assert.IsFalse(Double.IsNaN(solver.Solution[0]));
Assert.IsFalse(Double.IsNaN(solver.Solution[1]));
Assert.IsFalse(Double.IsNaN(solver.Solution[2]));
}
/// <summary>
/// Creates a new <see cref="Subplex"/> optimization algorithm.
/// </summary>
///
/// <param name="function">The objective function whose optimum values should be found.</param>
///
public Subplex(NonlinearObjectiveFunction function)
: base(function)
{
init(function.NumberOfVariables);
}
public void AugmentedLagrangianSolverConstructorTest7()
{
// maximize 2x + 3y, s.t. 2x² + 2y² <= 50
// Max x' * c
// x
// s.t. x' * A * x <= k
// x' * i = 1
// lower_bound < x < upper_bound
double[] c = { 2, 3 };
double[,] A = { { 2, 0 }, { 0, 2 } };
double k = 50;
// Create the objective function
var objective = new NonlinearObjectiveFunction(2,
function: (x) => x.InnerProduct(c),
gradient: (x) => c
);
// Test objective
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i * 2 + j * 3;
double actual = objective.Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
// Create the optimization constraints
var constraints = new List<NonlinearConstraint>();
constraints.Add(new QuadraticConstraint(objective,
quadraticTerms: A,
shouldBe: ConstraintType.LesserThanOrEqualTo, value: k
));
// Test first constraint
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i * (2 * i + 0 * j) + j * (0 * i + 2 * j);
double actual = constraints[0].Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
// Create the solver algorithm
AugmentedLagrangianSolver solver =
new AugmentedLagrangianSolver(2, constraints);
double maxValue = solver.Maximize(objective);
Assert.AreEqual(18.02, maxValue, 0.01);
Assert.AreEqual(2.77, solver.Solution[0], 1e-2);
Assert.AreEqual(4.16, solver.Solution[1], 1e-2);
}
/// <summary>
/// Implements the actual optimization algorithm. This
/// method should try to minimize the objective function.
/// </summary>
protected override bool Optimize()
{
if (Function == null)
{
throw new InvalidOperationException("function");
}
if (Gradient == null)
{
throw new InvalidOperationException("gradient");
}
NonlinearObjectiveFunction.CheckGradient(Gradient, Solution);
int n = NumberOfVariables;
int m = corrections;
String task = "";
String csave = "";
bool[] lsave = new bool[4];
int iprint = 101;
int[] nbd = new int[n];
int[] iwa = new int[3 * n];
int[] isave = new int[44];
double f = 0.0d;
double[] x = new double[n];
double[] l = new double[n];
double[] u = new double[n];
double[] g = new double[n];
double[] dsave = new double[29];
int totalSize = 2 * m * n + 11 * m * m + 5 * n + 8 * m;
if (work == null || work.Length < totalSize)
{
work = new double[totalSize];
}
int i = 0;
{
for (i = 0; i < UpperBounds.Length; i++)
{
bool hasUpper = !Double.IsInfinity(UpperBounds[i]);
bool hasLower = !Double.IsInfinity(LowerBounds[i]);
if (hasUpper && hasLower)
{
nbd[i] = 2;
}
else if (hasUpper)
{
nbd[i] = 3;
}
else if (hasLower)
{
nbd[i] = 1;
}
else
{
nbd[i] = 0; // unbounded
}
if (hasLower)
{
l[i] = LowerBounds[i];
}
if (hasUpper)
{
u[i] = UpperBounds[i];
}
}
}
// We now define the starting point.
{
for (i = 0; i < n; i++)
{
x[i] = Solution[i];
}
}
double newF = 0;
double[] newG = null;
// We start the iteration by initializing task.
task = "START";
iterations = 0;
//
// c ------- the beginning of the loop ----------
//
L111:
if (Token.IsCancellationRequested)
{
return(false);
}
iterations++;
//
// c This is the call to the L-BFGS-B code.
//
setulb(n, m, x, 0, l, 0, u, 0, nbd, 0, ref f, g, 0,
factr, pgtol, work, 0, iwa, 0, ref task, iprint, ref csave,
lsave, 0, isave, 0, dsave, 0);
//
if ((task.StartsWith("FG", StringComparison.OrdinalIgnoreCase)))
{
newF = Function(x);
newG = Gradient(x);
evaluations++;
f = newF;
for (int j = 0; j < newG.Length; j++)
{
g[j] = newG[j];
}
}
// c
else if ((task.StartsWith("NEW_X", StringComparison.OrdinalIgnoreCase)))
{
}
else
{
if (task == "ABNORMAL_TERMINATION_IN_LNSRCH")
{
Status = BoundedBroydenFletcherGoldfarbShannoStatus.LineSearchFailed;
}
else if (task == "CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH")
{
Status = BoundedBroydenFletcherGoldfarbShannoStatus.FunctionConvergence;
}
else if (task == "CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL")
{
Status = BoundedBroydenFletcherGoldfarbShannoStatus.GradientConvergence;
}
else
{
throw OperationException(task, task);
}
for (int j = 0; j < Solution.Length; j++)
{
Solution[j] = x[j];
}
newF = Function(x);
newG = Gradient(x);
evaluations++;
if (Progress != null)
{
Progress(this, new OptimizationProgressEventArgs(iterations, 0, newG, 0, null, 0, newF, 0, true)
{
Tag = new BoundedBroydenFletcherGoldfarbShannoInnerStatus(
isave, dsave, lsave, csave, work)
});
}
return(true);
}
if (Progress != null)
{
Progress(this, new OptimizationProgressEventArgs(iterations, 0, newG, 0, null, 0, f, 0, false)
{
Tag = new BoundedBroydenFletcherGoldfarbShannoInnerStatus(
isave, dsave, lsave, csave, work)
});
}
goto L111;
}
public void AugmentedLagrangianSolverConstructorTest4()
{
// min x*y+ y*z
//
// s.t. x^2 - y^2 + z^2 - 2 >= 0
// x^2 + y^2 + z^2 - 10 <= 0
// x + y = 1
//
double x = 0, y = 0, z = 0;
var f = new NonlinearObjectiveFunction(
function: () => x * y + y * z,
gradient: () => new[]
{
y, // df/dx
x + z, // df/dy
y, // df/dz
}
);
var constraints = new List<NonlinearConstraint>();
constraints.Add(new NonlinearConstraint(f,
function: () => x * x - y * y + z * z,
gradient: () => new[] { 2 * x, -2 * y, 2 * z },
shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 2
));
constraints.Add(new NonlinearConstraint(f,
function: () => x * x + y * y + z * z,
gradient: () => new[] { 2 * x, 2 * y, 2 * z },
shouldBe: ConstraintType.LesserThanOrEqualTo, value: 10
));
constraints.Add(new NonlinearConstraint(f,
function: () => x + y,
gradient: () => new[] { 1.0, 1.0, 0.0 },
shouldBe: ConstraintType.EqualTo, value: 1
)
{
Tolerance = 1e-5
});
var solver = new AugmentedLagrangian(f, constraints);
solver.Solution[0] = 1;
solver.Solution[1] = 1;
solver.Solution[2] = 1;
Assert.IsTrue(solver.Minimize());
double minValue = solver.Value;
Assert.AreEqual(1, solver.Solution[0] + solver.Solution[1], 1e-4);
Assert.IsFalse(Double.IsNaN(minValue));
Assert.IsFalse(Double.IsNaN(solver.Solution[0]));
Assert.IsFalse(Double.IsNaN(solver.Solution[1]));
Assert.IsFalse(Double.IsNaN(solver.Solution[2]));
}
/// <summary>
/// Creates a new <see cref="NelderMead"/> non-linear optimization algorithm.
/// </summary>
///
/// <param name="function">The objective function whose optimum values should be found.</param>
///
public NelderMead(NonlinearObjectiveFunction function)
: base(function)
{
init(function.NumberOfVariables);
}
public void AugmentedLagrangianSolverConstructorTest5()
{
// Suppose we would like to minimize the following function:
//
// f(x,y) = min 100(y-x²)²+(1-x)²
//
// Subject to the constraints
//
// x >= 0 (x must be positive)
// y >= 0 (y must be positive)
//
double x = 0, y = 0;
// First, we create our objective function
var f = new NonlinearObjectiveFunction(
// This is the objective function: f(x,y) = min 100(y-x²)²+(1-x)²
function: () => 100 * Math.Pow(y - x * x, 2) + Math.Pow(1 - x, 2),
// The gradient vector:
gradient: () => new[]
{
2 * (200 * Math.Pow(x, 3) - 200 * x * y + x - 1), // df/dx = 2(200x³-200xy+x-1)
200 * (y - x*x) // df/dy = 200(y-x²)
}
);
// Now we can start stating the constraints
var constraints = new List<NonlinearConstraint>();
// Add the non-negativity constraint for x
constraints.Add(new NonlinearConstraint(f,
// 1st constraint: x should be greater than or equal to 0
function: () => x, shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 0,
gradient: () => new[] { 1.0, 0.0 }
));
// Add the non-negativity constraint for y
constraints.Add(new NonlinearConstraint(f,
// 2nd constraint: y should be greater than or equal to 0
function: () => y, shouldBe: ConstraintType.GreaterThanOrEqualTo, value: 0,
gradient: () => new[] { 0.0, 1.0 }
));
// Finally, we create the non-linear programming solver
var solver = new AugmentedLagrangian(f, constraints);
// And attempt to solve the problem
Assert.IsTrue(solver.Minimize());
double minValue = solver.Value;
Assert.AreEqual(0, minValue, 1e-10);
Assert.AreEqual(1, solver.Solution[0], 1e-6);
Assert.AreEqual(1, solver.Solution[1], 1e-6);
Assert.IsFalse(Double.IsNaN(minValue));
Assert.IsFalse(Double.IsNaN(solver.Solution[0]));
Assert.IsFalse(Double.IsNaN(solver.Solution[1]));
}
private static void test1(IGradientOptimizationMethod inner, double tol)
{
// maximize 2x + 3y, s.t. 2x² + 2y² <= 50 and x+y = 1
// Max x' * c
// x
// s.t. x' * A * x <= k
// x' * i = 1
// lower_bound < x < upper_bound
double[] c = { 2, 3 };
double[,] A = { { 2, 0 }, { 0, 2 } };
double k = 50;
// Create the objective function
var objective = new NonlinearObjectiveFunction(2,
function: (x) => x.InnerProduct(c),
gradient: (x) => c
);
// Test objective
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i * 2 + j * 3;
double actual = objective.Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
// Create the optimization constraints
var constraints = new List<NonlinearConstraint>();
constraints.Add(new QuadraticConstraint(objective,
quadraticTerms: A,
shouldBe: ConstraintType.LesserThanOrEqualTo, value: k
));
constraints.Add(new NonlinearConstraint(objective,
function: (x) => x.Sum(),
gradient: (x) => new[] { 1.0, 1.0 },
shouldBe: ConstraintType.EqualTo, value: 1,
withinTolerance: 1e-10
));
// Test first constraint
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i * (2 * i + 0 * j) + j * (0 * i + 2 * j);
double actual = constraints[0].Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
// Test second constraint
for (int i = 0; i < 10; i++)
{
for (int j = 0; j < 10; j++)
{
double expected = i + j;
double actual = constraints[1].Function(new double[] { i, j });
Assert.AreEqual(expected, actual);
}
}
AugmentedLagrangian solver =
new AugmentedLagrangian(inner, objective, constraints);
Assert.AreEqual(inner, solver.Optimizer);
Assert.IsTrue(solver.Maximize());
double maxValue = solver.Value;
Assert.AreEqual(6, maxValue, tol);
Assert.AreEqual(-3, solver.Solution[0], tol);
Assert.AreEqual(4, solver.Solution[1], tol);
}
public void ConstructorTest6_2()
{
/// This problem is taken from Fletcher's book Practical Methods of
/// Optimization and has the equation number (9.1.15).
var function = new NonlinearObjectiveFunction(2, x => -x[0] - x[1]);
NonlinearConstraint[] constraints =
{
new NonlinearConstraint(2, x => -(x[1] - x[0] * x[0]) <= 0),
new NonlinearConstraint(2, x => -(-x[0] * x[0] - x[1] * x[1]) <= 1.0),
};
Cobyla cobyla = new Cobyla(function, constraints);
Assert.IsTrue(cobyla.Minimize());
double minimum = cobyla.Value;
double[] solution = cobyla.Solution;
double sqrthalf = Math.Sqrt(0.5);
Assert.AreEqual(-sqrthalf * 2, minimum, 1e-10);
Assert.AreEqual(sqrthalf, solution[0], 1e-5);
Assert.AreEqual(sqrthalf, solution[1], 1e-5);
double expectedMinimum = function.Function(cobyla.Solution);
Assert.AreEqual(expectedMinimum, minimum);
}