Different number of function evaluations and iterations while using MEX of objective function in patternsearch

I generated a MEX function of my objective function to speed up the optimization process (pattern search) using Matlab coder (Matlab R2023b). The optimization with pattern search using the MEX function results in identical solutions. However, the number of function evaluations used by patternsearch is different while using the MEX of objective function. Similarly, the number of iterations employed by patternsearch is also different when I use the MEX function. The rest of the settings are identical.

I am not able to figure out the cause behind this discrepancy since MEX is essentially the same function.

Thanks in advance for any help in understanding this!

[Steven Lord]

I am not able to figure out the cause behind this discrepancy since MEX is essentially the same function.

What does "essentially" mean here?

[aksri]

I meant MEX of my Objective function should be identical to the Objective function as far as Pattern search is concerned in terms of function inputs and outputs. What could be the reason for different number of function evaluations. Thanks.

[Steven Lord]

How did you generate the MEX-file?

Are you generating your x0 consistently across all your runs, or are you using some function that involves random numbers like rand?

[aksri]

@Steven Lord Thanks for your response. I used Matlab coder to generate the MEX of the objective function.

The Objective function I am using is deterministic and I am not using any function which involves random numbers. As I mentioned in my question, The solutions and function values are identical for both the cases i.e. when using MEX of objective function 'f' and, when using 'f'. The difference is in the number of iterations and function evaluations patternsearch is taking in optimisation. This is what I am not able to understand.

[Torsten]

Can you save the input parameters with which the objective function is called in both cases and compare ?

[Steven Lord]

How are you calling patternsearch? Can you show us the exact call you make (and how the variables that call uses are created)?

[aksri]

Thanks @Torsten and @Steven Lord. Here is how I am calling the function:

%ObjectiveFunction = @(R) MyFun(R,par1,par2,…);

ObjectiveFunction = @(R) MyFun_mex(R,par1,par2,…); % When using Mex of MyFun

options = optimoptions('patternsearch','Display','iter', ...

'CacheSize',inf,'UseParallel',true,..);

%start optimization

[R,Fval,exitflag,output] = patternsearch(ObjectiveFunction,R0,...

A,b,[],[],LB,UB,[],options);

Feval= output.funccount;

In a single simulation, I am calling patternsearch 4 times for 4 different usecases. It is the last case i.e. the 4th call to patternsearch which gives me different number of function evaluations. The first 3 cases are identical for both MyFun and MyFun_mex.

[Torsten]

What if you don't use the "UseParallel" option ?