Filter löschen
Filter löschen

Bad optimization of an objective nonlinear fitting

4 Ansichten (letzte 30 Tage)
HamzaChem
HamzaChem am 7 Jun. 2023
Bearbeitet: HamzaChem am 7 Jun. 2023
I have a bad understanding why this optimization does not work. I have a raw signal which is the convolution between a signal X (which I would like to find by a fitting) and signal 2 which is known.
I need to optimize a convolution between a three-exponential terms function with signal sig2 (which is a response function that should not change, only the three-exponential fitting that shpould be optimized so that its convolution with sig2 should match the raw signal "sig1").
Data = xlsread('rawdata.xlsx','Sheet1','A2:C119000');
time = Data(:,1) ;
sig1 = Data(:,2) ;
sig2 = Data(:,3) ;
% Define signal 3 (three exponential terms fitting function)
modelFun = @(p,time) - 1.8785e-01*exp(-6E6 *time) - p(2)*exp(-p(3)*time) + (1.8785e-01+p(2))*exp(-p(1)*time);
Now I give very good starting points for the fitting
% Define initial values for the parameters for the three-exponential terms
% fitting to be optimized
p0 = [1.4531e+03 3.5868e-01 1.9509e+05];
No I use lsqcurvefit to construct the final target function which should optimize the final fitting so that the convolution of the fitting with sig2 should reproduce well the raw data (sig1).
[params,resnorm,residual,exitflag,output,lambda,jacobian] = lsqcurvefit(@(p,time) conv(sig2,modelFun(p,time), 'same'),p0,time,sig1);
Local minimum possible. lsqcurvefit stopped because the final change in the sum of squares relative to its initial value is less than the value of the function tolerance.
Now if I display the output of this optimization (I display only the three-exponential term function):
tesfunc1 = - 1.8785e-01*exp(-6E6 *time) - params(2)*exp(-params(3)*time) + (1.8785e-01+params(2))*exp(-params(1)*time);
figure
plot(time,sig1,'DisplayName' ,'raw signal' )
hold on
plot(time, tesfunc1, 'DisplayName' ,'optimized fitting')
xlabel('Time');
ylabel('Intensity');
legend
It is defenitely a bad optmization, if I do convolution with sig2, it does not reproduce theraw data at all:
Objective = conv(sig2,tesfunc1);
figure
plot(time,sig1,'DisplayName' ,'raw signal' )
hold on
plot(time,Objective(1:length(time))*(max(sig2)/max(Objective)),'DisplayName' ,'objective function' ) % the max operation is only to scale as the convolution takes the total amplitude to a high level
xlabel('Time');
ylabel('Intensity');
legend
However, if I use exactly the initial parameters, it should be a very good fitting (apart fron the convolution changes the overall amplitude that I still do not know how to fix).:
tesfunc2 = - 1.8785e-01*exp(-6E6 *time) - p0(2)*exp(-p0(3)*time) + (1.8785e-01+p0(2))*exp(-p0(1)*time);
% objective function if the optimization found parameters like in p0
% (declared initial points)
Objective = conv(sig2,tesfunc2);
figure
plot(time,sig1,'DisplayName' ,'raw signal' )
hold on
plot(time,Objective(1:length(time))*(max(sig2)/max(Objective))*1.1,'DisplayName' ,'objective function' ) % again I do the rescaling manually
xlabel('Time');
ylabel('Intensity');
legend
Of course for this testfunction I knew the parameters, and in reality I would like to find them. for the lsqcurvefit, even if I spoiled it with very good parameters giving the function above, it still missed the fitting completely!
I am wondering what can make this optimization better. And I have seen some use "fminunc" with a cost function that would be in my case (costFunction = @(params) mean((rawData - conv(signal2, fittingfunction(:, params))).^2);)), but I do not really know how to implement it in my case.
Many thanks in advance.
  1 Kommentar
HamzaChem
HamzaChem am 7 Jun. 2023
Bearbeitet: HamzaChem am 7 Jun. 2023
I rewrote the code using fmincon (which is faster) but same fitting outputs:
% Use fmincon to fit signal 1 with the convolution of signals 2 and 3
[params,resnorm,residual,exitflag,output,lambda,jacobian] = fmincon(@(p) mean((sig1-conv(sig2,modelFun(p,time),'same')).^2),p0);
% comparing the optimized fitting with a "should be good" fitting
% optimized three-exponential fitting function
tesfunc1 = - 1.8785e-01*exp(-6E6 *time) - params(2)*exp(-params(3)*time) + (1.8785e-01+params(2))*exp(-params(1)*time);
% three-exponential fitting which should be good
tesfunc2 = - 1.8785e-01*exp(-6E6 *time) - p0(2)*exp(-p0(3)*time) + (1.8785e-01+p0(2))*exp(-p0(1)*time);
figure
plot(time,sig1,'DisplayName' ,'raw signal' )
hold on
plot(time, tesfunc1, 'DisplayName' ,'optimized fitting')
hold on
plot(time, tesfunc2, 'DisplayName' ,'a good fit (known)' )
xlabel('Time');
ylabel('Intensity');
legend

Melden Sie sich an, um zu kommentieren.

Antworten (0)

Produkte


Version

R2022b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by