# feval distorted in ga after using nonlinearconstr

18 views (last 30 days)
Tsuwei Tan on 25 Feb 2021
Commented: Tsuwei Tan on 28 Mar 2021 at 16:59
I am using Genetic Algoritm
[x,fval,exitflag,output,population,scores]=ga(@(x)cost_fun(x),...
nvar,A,b,[],[],lb,ub,@(x)ssp_constraint(x,U,ssp),IntCon,options);
to find the global min with nvar=6. and the options includes a nonlinear constraint function called ssp_constraint. U and ssp are passing parameters.
function [c, ceq]=ssp_constraint(x,U,ssp)
% c is inequality and ceq is equality
U1=U(:,1);U2=U(:,2);U3=U(:,3);U4=U(:,4);U5=U(:,5);U6=U(:,6);
a1=x(1);a2=x(2);a3=x(3);a4=x(4);a5=x(5);a6=x(6);
SSP_r=ssp+a1*U1+a2*U2+a3*U3+a4*U4+a5*U5+a6*U6;
[minin, I]=min(SSP_r);
% c=[I-40;1485-minin];
c=1485-minin;
% True of False
TF1=islocalmin(SSP_restored_1st15day);
ceq=sum(TF1)-1;
end
The issue is if I have two inequality constraints instead of one in this ssp_constraint function, in my case , change
c=1485-minin;
to
c=[I-40;
1485-minin];
then the fval from ga is skewed to a negative value and contiune to search; my objective function cost_fun is written to spit out a positive value, and latter generation should approach from some positve vale to zero.
I am using a High Performace Computer to conduct parallell computation, the options in ga is
options = optimoptions('ga','UseParallel', true, 'UseVectorized', false,...
'MaxTime',3600*12,'MaxGenerations',600,'MaxStallGenerations',600,'PopulationSize',63,...
'PlotFcn',@gaplotbestf,'OutputFcns',@ga_save_each_gen);
I write a fucntion ga_save_each_gen to save each generation
function [state,options,optchanged]=ga_save_each_gen(options,state,flag)
savefolder='result' ; % Save to the subfolder
if isfolder(savefolder)==0; mkdir(savefolder); end
Score_gen=state.Score;
Population_gen=state.Population;
Generation_gen=state.Generation;
optchanged=false;
%ga does not accept changes in options, and ignores optchanged.
save(['./' savefolder '/gen_' num2str(Generation_gen,'%.4d') '.mat'],'Score_gen','Population_gen','Generation_gen')
end
So I should see the saved variable Score_gen which is a vector with the size (63,1) at seach saved gen_xxxx.mat
Tsuwei Tan on 27 Feb 2021
It seems to me that
[mini, Index ]=min(any_input);
that Index output is not allowed inside the non linear constraint function.

Matt J on 27 Feb 2021
The main thing I see that looks out of place is that is the role of SSP_restored_1st15day. If it is a variable, I cannot see where it is defined, but regardless it certainly doesn't depend on the input unknown vector x. If the constraints are not a function of x, that can have crazy effects.
Incidentally also, the constraint
min(SSP_r)>=1485
is linear. ga() would deal with linear constraints more effectively if you used the A,b,Aeq,beq inputs to handle them.
Tsuwei Tan on 28 Mar 2021 at 16:59
When IntCon is nonempty, Aeq and beq must be an empty entry ([]), and nonlcon must return empty for ceq. For more information on integer programming, see Mixed Integer ga Optimization.
Finally figure out why.