private IVector <T> MinimizeInternal(ILeastSquaresFunctional <T> objective, IParametricFunction <T> function, IVector <T> initialParameters, IVector <T> minimumParameters = default, IVector <T> maximumParameters = default) { var la = LinearAlgebra.Value; var x = initialParameters.Clone() as IVector <T>; var bindF = function.Bind((x)); var iter = 0; var residual = objective.Residual(bindF); var error = la.Sqrt(la.Dot(residual.ToArray(), residual.ToArray())); var oldError = la.Cast(1000); while (la.Compare(la.Sub(error, oldError), Eps) == 1 || iter++ < MaxIteration) { var jacobi = objective.Jacobian(bindF); var jacobiT = jacobi.Transpose(); //JT var jTj = jacobiT.Mult(jacobi); //jTj var jTj_1 = jTj.Inverse(); var jTj_1jT = jTj_1.Mult(jacobiT); var temp = jTj_1jT.Mult(residual); var gamma = GoldenRatioMethod <T> .FindMin(objective, function, x, temp, Eps); x.Sub(temp.Mult(gamma)); this.ApplyMinimumAndMaximumValues(minimumParameters, maximumParameters, x, la); oldError = error; bindF = function.Bind(x); residual = objective.Residual(bindF); error = la.Sqrt(la.Dot(residual.ToArray(), residual.ToArray())); } return(x); }
private IVector <T> MinimizeInternal(IDifferentiableFunctional <T> objective, IParametricFunction <T> function, IVector <T> initialParameters, IVector <T> minimumParameters = default, IVector <T> maximumParameters = default) { var la = LinearAlgebra.Value; var xOld = initialParameters.Clone() as IVector <T>; //Calculate initial gradient var bindF = function.Bind(xOld); var fiNew = objective.Gradient(bindF); var xNew = initialParameters.Clone() as IVector <T>; var k = 0; var normOld = la.Sqrt(la.Dot(fiNew.ToArray(), fiNew.ToArray())); var normNew = la.Cast(1000); while (k++ < MaxIteration && (la.Compare(la.Sqrt(la.Dot(fiNew.ToArray(), fiNew.ToArray())), Eps) == 1) && (la.Compare(la.Sub(normNew, normOld), Eps) == 1)) { bindF = function.Bind(xNew); fiNew = objective.Gradient(bindF); normOld = normNew; normNew = la.Sqrt(la.Dot(fiNew.ToArray(), fiNew.ToArray())); var gamma = GoldenRatioMethod <T> .FindMin(objective, function, xOld, fiNew, Eps); xOld = xNew.Clone() as IVector <T>; xNew = xOld.Sub(fiNew.MultWithCloning(gamma)); this.ApplyMinimumAndMaximumValues(minimumParameters, maximumParameters, xNew, la); } return(xNew); }
private IVector <T> MinimizeInternal(IDifferentiableFunctional <T> objective, IParametricFunction <T> function, IVector <T> initialParameters, IVector <T> minimumParameters = default, IVector <T> maximumParameters = default) { var la = LinearAlgebra.Value; var xNew = initialParameters.Clone() as IVector <T>; var bindF = function.Bind(xNew); var curVal = objective.Value(bindF); var prevVal = curVal; var gradient = objective.Gradient(bindF); var p = objective.Gradient(bindF).Clone() as IVector <T>; gradient.Mult(la.Cast(-1)); var gradSquare = la.Dot(p.ToArray(), p.ToArray()); int numIter = 0; do { T alpha, beta, newGradSquare; IVector <T> newGrad; //Ищем минимум F(x + alpha * p) с помощью метода одномерной оптимизации alpha = GoldenRatioMethod <T> .FindMin(objective, function, xNew, p, Eps); xNew = xNew.Add(p.MultWithCloning(la.Mult(la.Cast(-1), alpha))); this.ApplyMinimumAndMaximumValues(minimumParameters, maximumParameters, xNew, la); bindF = function.Bind(xNew); newGrad = objective.Gradient(bindF).Mult(la.Cast(-1)); newGradSquare = la.Dot(newGrad.ToArray(), newGrad.ToArray()); beta = numIter % (5 * SpaceSize) == 0 ? la.GetZeroValue() : la.Div(la.Mult(la.Cast(-1), la.Sub(newGradSquare, la.Dot(newGrad.ToArray(), gradient.ToArray()))), gradSquare); p.Mult(beta).Add(newGrad); prevVal = curVal; curVal = objective.Value(bindF); gradient = newGrad; gradSquare = newGradSquare; } while (la.Compare(gradSquare, Eps) == 1 && MaxIteration > numIter++); return(xNew); }