Reusing ConstraintDerivative from AD in fmincon

Question

Rory Gover 2023-7-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1991428-reusing-constraintderivative-from-ad-in-fmincon

回答： Ayush Aniket 2023-8-25

Hi,

I have a problem of the form:

opts    = optimoptions('fmincon','Algorithm','sqp','Display','iter');
X       = optimvar("X",size(lb),LowerBound=lb,UpperBound=ub);
prob    = optimproblem('Objective',costFunction(X),'Constraints',constraints(X,constantPassthrough,variablePassthrough))

The costFunction() has a quadratic response to the control X, but the constraints() function is non-linear and non-convex and is fed constant passthrough arguments (which are constant within a given run) and variable passthrough arguments (which are iterated through within the run). The constraints function nominally consists of four OptimizationContraints and 1 OptimizationInequality.

The solutions for X are sought for N different conditions defined by variablePassthrough(i).

for i = 1:N
    prob    = optimproblem('Objective',costFunction(X),'Constraints',constraints(X,constantPassthrough,variablePassthrough(i)));
    sol     = solve(prob,x0,'Options',opts,'solver','fmincon','ObjectiveDerivative','auto','ConstraintDerivative','auto');
    Xs(i,:) = sol.X;
end

Currently, the primary issue is that each time the solve() function is called, I believe that the AD is being used to redetermine the ConstraintDerivative. My understanding is that determining the ConstraintDerivative functions are the largest contributor to the computational cost each tme solve() is run.

Is there any way to have the ConstraintDerivative calculated the first time (i.e. i =1), and then reused for subsequent runs of (i = 2:N)?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Ayush Aniket 2023-8-25

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1991428-reusing-constraintderivative-from-ad-in-fmincon#answer_1293792

Hi Rory Gover,

No, there is no way to reuse the ConstraintDerivative as the ‘solve’ function computes it internally and there is no output for it.

A workaround that might help is to pass the gradient information separately for each optimization iteration. The following documentation page discusses that way:

https://in.mathworks.com/help/optim/ug/derivatives-problem-based.html

Hope it helps.