It seems that patternsearch does not use my initial guess.

Greetings,

I am working on a optimization problem by patternseach. I have a initial guess vector whose objective function is about 13.5. However, when I used the patterensearch, the objective function is 63.59 far from my initial guess; even after 2200 iterations, it is still far from my inital guess. I checked the x0 of patteren search at iter 0 and found it used [1 0 0...] as x0, instead of my inital guess vector or (my inital guess vector + [1 0 0 ...]).

I wish it can search from my initial guess and the range I provided. Are there any good suggestions?

Answer 1

[Stephane Dauvillier]

Hi,

After checking your code I've seen your write to time lb(222) and the second time you specify lb(222)=0 BUT your initial guess doesn't meet the lower bound (para0(222) = -0.0098). That's why patternsearch was changing your initial guess

[Xin Shen]

Thank you so much!!! I cannot find the bug for several days. I will redo my optimization again!

Answer 2

[Stephane Dauvillier]

Can you provided your coe ?

Is there any constaints for your oproblem and the initakl guess doesn't fit the constraint ?

[Xin Shen]

My decision variable vector is para, which includes w vector and other decision variables. In terms of constraints, I used a equality constratint sum(w) = 1 and lb & ub constraints in my code. And I think initial guess satisfy all constraints.

% w is a decicison vector initialized by value 1/220
% sum(w) = 1 is the only one equality constraint used in optimization
w = 1/220 * ones(220,1);

% other decision variables and initial guess
Vm=2.2*4e-5;dtyr=-0.0098; dile=-0.0344; darg=-0.00418;Km=1.5;dthr=-0.0949;dglu=-3.4032; dasp=-0.2327;
bspec=[1e-3;2e-4;1e-3;1e-3;1e-3;1e-3;1e-4;1e-3;1e-3;1e-3;1e-3;1e-5;1e-4;1;1e-5;1e-3;1e3];

% initial guess vector
para0 = [w;Vm;dtyr; dile; darg;Km;dthr;dglu; dasp; bspec];

% lb and ub constraints used in optimization
lb = zeros(size(para0));
ub = zeros(size(para0));
lb(1:220) = 0;  lb(221) = 0;  lb(222) = -0.9;  lb(223) = -3.4;  lb(224) = -0.42; 
lb(225) = 0;  lb(226) = -9.5;  lb(227) = -340;  lb(228) = -23;  lb(229:245) = 0; 
lb(222) = 0;  lb(245) = 0; 
ub(1:220) = 1; ub(221) = 0.01; ub(222) = 0; ub(223) = 0; ub(224) = 0;
ub(225) = 150; ub(226) = 0; ub(227) = 0; ub(228) = 0; ub(229:245) = 0.1; 
ub(242) = 100;  ub(245) = 1e5; 
options = optimoptions('patternsearch','UseParallel',false,'PlotFcn',{@psplotbestf,@psplotbestx},'Display','iter',...
    'MeshTolerance',1e-6,'OutputFcns',@(optimvalues,options,flag)custom(optimvalues,options,flag,data_exp));

% sum(w) = 1 is the only one equality constraint used in optimization
para_opt = patternsearch(@(para)CostFun(para,data_exp),para0,[],[],...
           [ones(1,size(w,1)) zeros(1,length(para0)-length(w))],1,lb,ub,[],options);