Online parameter optimization using genetic algorithm - Experimental setup

I'm currently working on a project that intents to reduce the amplitude of vibration on a cylinder caused by transverse water flow ( VIV phenomen).

The actuator is basically composed by 4 dc motors around the main cylinder that produces a disturb on the flow and, over specific range of rotation, can reduce the response amplitude of the main cylinder.

The idea is to use a MOEA ( Multi Objective Evolutionary Algorithm) to find optimal rotation of each cylinder that reduce the vibration while optmize power consumption.

My simulink model to manual run DC motor is already woking properly, runing on external mode, using a S-Function to control the motors conected to a Arduino boad.

Unfortunatelly, when I try to test a simple optimization, it fails. As I don't have experimental setup built yet, tested a simple objective function optimization while running he DC motor on external mode. A known simple cost function (y= x^2 -6x) to have the value 'x' minimized between lower and upper boundary [0 ; 6] should be found as "x = 3, y -9" after some iterations ( using fminsearch", "ga"or any other optimization tool). The idea is to use each iteration value to run DC motor at a specific rotation, add a delay for a few seconds, and release to run next iteraton until optimization converge. I tried simulate a simple using simulink tool "Analysis - Parameter Optimization" (1) or by code usinga wrapper function using "coder-extrinsic" comand (2). Important to remember that this is easily achieved on simulnk "normal mode", the problem is when I run on "external mode".

Below ate the errors found:

(1) Changing property 'ReturnWorkspaceOutputsName' is not allowed

(2)The function 'fmincon' is not supported for standalone code generation. See the documentation for coder.extrinsic to learn how you can use this function in simulation.

So, basically my question is: How can I run parameter optimization over online experimental data, simulink external mode, where the cost function is unknown ( I only have the response value measured) and the parameters tested at each step are used to generate next response.