Bayesian optimization for nonlinear programming

Hi there,

I was wondering whether bayesian optimization is suitable for control ( nonlinear programming problem). In many cases, the bayesian optimization is used in hyperparameter optimization.

Here is my problem:

min

s.t. Ypred =f(U)

where Yr is setpints, U is the manipulated variables, f() represents a neural network model, Q and R are matrices. I am trying to find the optimal U which enables Ypred tracking setpoints Yr.

I have tried many other solver for this problem, e.g. fmincon, Yalmip. All of them worked well.

However, the bayesian optimization gives a fluctuating control action U

(black lines are setpoints, red lines are manipulated variables and outputs)

Here is bayesopt solver setting:

results=bayesopt(@my_optfun, dUdata_M, ...
    'IsObjectiveDeterministic',true,'plotFcn', [], ...
    'NumSeedPoints',6, 'ExplorationRatio', 0.2, 'MaxObjectiveEvaluations',55, ...
    'XConstraintFcn',@(Input) xconstraint(Input, u_k_1, ...
    'InitialX', Initial_du);

Could you please tell how to modify the solver in order to improve the control performances?

Thank you in advance.