How to choose DiffMinChange value?

Dear all,

I am running an optimisation program with the Optimisation settings shown below :

%% Optimisation settings
lb = [];
ub = [];
options = optimoptions(@fmincon,'Algorithm','interior-point','DiffMinChange',0.00089205,'StepTolerance',1.0000e-12,'MaxFunctionEvaluations',100000,'plotFcns',{@optimplotx,@optimplotconstrviolation,@optimplotfval,@optimplotfirstorderopt});
[pos_opt,fval,exitflag,output,lambda,grad,hessian] = fmincon(@(pos_init) objective(pos_init,XX,YY),pos_init,[],[],[],[],lb,ub,@(pos_init) nonlinconst(pos_init),options);

I am ending up getting the first order optimality around 0.138524. In principal the first order optimality should me closer to zero. Is there any optimal value for DiffMinChange I have to take depending on the problem, so that the optimality will reach zero.

The optimisation is going in the proper direction but not reaching zero.

[Walter Roberson]

Remember that fmincon() is not a global optimizer. You probably need better initial conditions.

[Selvaprakash Vijayaraman]

I have chosen a better initial condition but even now I am getting the first order optimality around 0.01 (its a major improvement) which is not close to zero.

I rechecked all my formulations and I reached this point, Choose an even better Initial condition and choose different DiffMinChange value.

Should I go for any global optimiser?

Answer 1

[Alan Weiss]

I suggest that you look at When the Solver Might Have Succeeded. In particular, https://www.mathworks.com/help/optim/ug/when-the-solver-might-have-succeeded.html#br58rj7 and https://www.mathworks.com/help/optim/ug/when-the-solver-might-have-succeeded.html#br544oi-1 seem relevant, but other suggestions might help, too.

Alan Weiss

MATLAB mathematical toolbox documentation