Combine parallel and Rapid Accelerator Mode for an optimization task with ga

I have an optimization task where the cost function includes a Simulink Model which is set to run in Rapid Accelerator Mode.

Now I like to use parallel optimization for my genetic algorithm (ga).

How do I set this up?

What I tried is the following:

model = 'Model14';
open_system(model);
options = optimoptions('ga', 'UseParallel', true);
nvars = 4;
A = [];
b = [];
Aeq = [];  %no constraint
beq = []; 
lb = [1e5 10 2 20];
ub = [8.5e5 4e6 50 40];
nonlcon = [];
[x, fval, output] = ga(@costfct_optim, nvars, A, b, Aeq, beq, lb, ub, nonlcon, options);

The cost function (costfct) looks as follows:

function ctot = costfct_optim(x)
plim = x(1);
ebat = x(2);
h = x(3);
Tthres = x(4);

model = 'Model14';
in = Simulink.SimulationInput(model);
in = in.setModelParameter('SimulationMode', 'rapid');
in = in.setVariable('limgain', plim);
in = in.setVariable('Ebatgain', ebat);
in = in.setVariable('hgain', h);
in = in.setVariable('Tthres', Tthres);

out = sim(in);

T_eol = out.yout{1}.Values.Time(end)/3600/24/365;
Etot = out.yout{1}.Values.Data(end)/3.6e6/T_eol;
[ctot, ~, ~, ~] = costfct(plim, ebat, T_eol, Etot);

This does not work. I guess this is because there is no setup function supplied, which I use when combining parallel simulations with Rapid Accelerator outside optimization tasks.

The line would look as follows:

out = parsim(in, 'ShowProgress', 'on', ...
    'SetupFcn', @() sldemo_parallel_rapid_accel_sims_script_setup(model));

Including this line in the cost function for the optimization task does not work. I guess this is because in that case every single cost function evaluation would try to set up a separate parallel simulation.

For completeness, the error I am getting with the first two pieces of code:

>> opt_ga
### Building the rapid accelerator target for model: Model14
### Successfully built the rapid accelerator target for model: Model14
Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 8).
### Build procedure for Model14 aborted due to an error.
Error using costfct_optim (line 21)
Unable to build a standalone executable to simulate the model 'Model14' in rapid accelerator mode.

Error in createAnonymousFcn>@(x)fcn(x,FcnArgs{:}) (line 11)
fcn_handle = @(x) fcn(x,FcnArgs{:});

Error in fcnvectorizer (line 16)
        parfor (i = 1:popSize)

Error in makeState (line 63)
        Score = fcnvectorizer(state.Population(initScoreProvided+2:end,:),FitnessFcn,1,options.SerialUserFcn);

Error in galincon (line 17)
state = makeState(GenomeLength,FitnessFcn,Iterate,output.problemtype,options);

Error in ga (line 401)
            [x,fval,exitFlag,output,population,scores] = galincon(FitnessFcn,nvars, ...

Error in opt_ga (line 41)
[x, fval, output] = ga(@costfct_optim, nvars, A, b, Aeq, beq, lb, ub, nonlcon, options);

Caused by:
    Error using Simulink.Simulation.internal.DesktopSimHelper.sim
    Dot indexing is not supported for variables of this type.
    Failure in user-supplied fitness function evaluation. GA cannot continue.

Answer 1

[Abdolkarim Mohammadi]

You should set ga to pass all of the population to the objective function at once, and there you implement the parallelization inside the objective function. Change the following in your code.

optimoptions('ga', 'UseVectorized', true);

and

out = parsim(in);

In addition, you should change your objective function to create SimulationInput object to set variables for the entire population. And the objective function's outputs to be calculated for the entire population using the SimulationOutput object which has the same dimensions as the SimulationInput object.

By the way, I think you should not open the model at the start of the optimization since it slows down Simulink. Instead, use close_system to close the model if it is open. Beware that you should not halt parsim with Ctrl+C, otherwise, the parallel pool becomes unusuable most of the time and you need to restart MATLAB.

[Florian Müller]

Exactly what I was looking for, thank you a lot!

Also thanks for the additional hints, I did exactly as suggested.

[Abdolkarim Mohammadi]

Another note: In simheurstics (simulation-based optimization via metaheuristics), the simulation model is run for many times, which requires a lot of time budget. In addition to paralellization, you can use the FastRestart option of the parsim to speed up even more.

out = parsim(in, 'UseFastRestart','on'); 

[Florian Müller]

Thanks also for this additional hint.

But somehow, FastRestart does not work. It leads to all workers containing errors.

I had this error before and could not figure the reason. Do you have an idea?

[Abdolkarim Mohammadi]

Read the limitations. Fast restart does not support Rapid Accelerator. You should switch to Accelerator mode.

https://www.mathworks.com/help/simulink/ug/how-fast-restart-improves-iterative-simulations.html

[Florian Müller]

Thanks again. Well, I will just keep it off. Rapid Accelerator is about factor 10 faster than Accelerator, so fast restart likely won't make up for that.

[Abdolkarim Mohammadi]

Time is the most important thing here. If it doesn't improve time, you shouldn't use fast restart.