新浪博客

Matlab优化函数中options选项的修改

2010-11-12 10:43阅读:
通常在使用Matlab优化工具箱中的函数时,需要根据不同要求修改优化选项,例如最大迭代次数、x处的终止容限等等。可通过 optimset 语句来修改优化选项参数:
options=optimset(‘optionName’,’optionValue’);如:
options=optimset('tolx',1e-100);
options=optimset(options,'tolfun',1e-100);

后面还可以依此法添加很多选项信息。
下面举一个完整例子:求取f=4*x(1)^2+5*x(1)*x(2)+2*x(2)^2的最小值
---------------------------------------------------------------------------------------------function [f,g]=ff3(x)
f=4*x(1)^2+5*x(1)*x(2)+2*x(2)^2;
if nargut >1
g(1)=8*x(1)+5*x(2);
g(2)=5*x(1)+4*x(2);
end
通过下面将优化选项结构options.GradObj设置为’on’来得到梯度值。
options=optimset(‘Gradobj’,’
on’);
x0=[1,1];
[x,fval,exitflag]=fminunc(@ff3,x0,options)
---------------------------------------------------------------------------------------------

Options的参数描述(OPTIMSET):

%OPTIMSET PARAMETERS
�rivativeCheck - Compare user supplied derivatives (gradients or Jacobian)
% to finite-differencing derivatives [ on | {off}]

%Diagnostics - Print diagnostic information about the function to be
% minimized or solved [ on | {off}]

%DiffMaxChange - Maximum change in variables for finite difference gradients
% [ positive scalar | {1e-1} ]
%DiffMinChange - Minimum change in variables for finite difference gradients
% [ positive scalar | {1e-8} ]

%Display - Level of display [ off | iter | notify | final ]
%GoalsExactAchieve - Number of goals to achieve exactly (do not over- or
% under-achieve) [ positive scalar integer | {0}]

%GradConstr - Gradients for the nonlinear constraints defined by user
% [ on | {off} ]
%GradObj - Gradient(s) for the objective function(s) defined by user
% [ on | {off}]
%Hessian - Hessian for the objective function defined by user [ on | {off} ]
%HessMult - Hessian multiply function defined by user
% [ function | {[]} ]
%HessPattern - Sparsity pattern of the Hessian for finite-differencing
% [ sparse matrix ]
%HessUpdate - Quasi-Newton updating scheme
% [ {bfgs} | dfp | gillmurray | steepdesc ]
%Jacobian - Jacobian for the objective function defined by user
% [ on | {off}]
%JacobMult - Jacobian multiply function defined by user
% [ function | {[]} ]
%JacobPattern - Sparsity pattern of the Jacobian for finite-differencing
% [ sparse matrix ]
%LargeScale - Use large-scale algorithm if possible [ {on} | off ]
%LevenbergMarquardt - Chooses Levenberg-Marquardt over Gauss-Newton algorithm
% [ on | off]
%LineSearchType - Line search algorithm choice [ cubicpoly | {quadcubic} ]
%MaxFunEvals - Maximum number of function evaluations allowed
% [ positive integer ]
%MaxIter - Maximum number of iterations allowed [ positive integer ]
%MaxPCGIter - Maximum number of PCG iterations allowed [positive integer]
%MeritFunction - Use goal attainment/minimax merit function
% [ {multiobj} | singleobj ]
%MinAbsMax - Number of F(x) to minimize the worst case absolute values
% [ positive scalar integer | {0} ]
%PrecondBandWidth - Upper bandwidth of preconditioner for PCG
% [ positive integer | Inf | {0} ]
%TolCon - Termination tolerance on the constraint violation [ positive scalar ]
%TolFun - Termination tolerance on the function value [ positive scalar ]
%TolPCG - Termination tolerance on the PCG iteration
% [ positive scalar | {0.1} ]
%TolX - Termination tolerance on X [ positive scalar ]
%TypicalX - Typical X values [ vector ]
---------------------------------------------------------------------------------------------
注:
优化问题求解时常常需要对相对误差,使用算法等进行设置.Matlab提供了options向量来对优化函数进行参数设置,options向来由18个元素.目前已经被optimset和optimget代替,详情可查阅函数optimset和optimget
下面逐个说明各个参数的功能和取值.
Options(1)=0(默认值)
功能:控制显示,优化过程中控制输出信息,0表示不显示;1表示显示;-1表隐藏信息.
Options(2)=1e-4
功能:控制x的精度,自变量x的最低精度终止条件.当所有终止条件都满足的时候,优化终止.
Options(3)=1e-4
功能:控制 f 精度,目标函数f的最低精度终止条件.当所有终止条件都满足的时候,优化终止.
options(4)=1e-7
功能:约束g的最低精度终止判别条件.当所有的终止条件都满足的时候,优化终止.
options(5)=0
功能:选择主要优化算法.
options(6)=0
功能:SD算法控制.选择搜索方向算法.
options(7)=0
功能:搜索算法控制.选择线性搜索算法.
options(8)=N/A
功能:函数值,算法结束时极值点的函数值,attgoal和minimax而言,它包含一个到达因子.
options(9)=0
功能:梯度检查控制.当值为1时,在最初的几个迭代周期,梯度将与有限差分计算的结果比较,此时,梯度函数必须存在.
options(10)=N/A
功能:函数计算计数.
options(11)=N/A
功能:梯度计算计数.
options(12)=N/A
功能:限定计数,限定函数梯度计算或差分梯度计算的次数.
options(13)=0
功能:等式约束个数,等式约束必须放在g的前几个元素中.
options(14)=0*n
功能:最大迭代次数,该值缺省时被置为n的100倍,n为自变量x的个数,在fmins中,缺省为n的200倍,在fminu中,为500n.
options(15)=0
功能:目标数,尽可能接近goals的目标数,由函数attgoal使用.
options(16)=1e-8
功能:最小摄动控制.有限差分梯度计算中的最小变化.对函数的梯度计算而言,实际使用的摄动将自动调整以提高精度,它将在最小摄动和最大摄动之间变化.
options(17)=0,1
有限差分梯度计算中变量的最大变化.
options(18)=N/A
功能:步长控制,在第一步迭代被赋值为1或更小

我的更多文章

下载客户端阅读体验更佳

APP专享