Variable number of optimizableVariables in a bayesopt function

As an example, for two variables I am using the bayesopt function, declaring the variables and calling the function like so:

C{1} = optimizableVariable('A',[1 2],'Transform','none');
C{2} = optimizableVariable('B',[1 2],'Transform','none');

fun = @(x)myfun(x.A,x.B);
results = bayesopt(fun,[C{:}],'AcquisitionFunctionName','expected-improvement-plus','IsObjectiveDeterministic',true,'ExplorationRatio',0.1,'NumSeedPoints',10,'MaxObjectiveEvaluations',Steps,'PlotFcn',{@plotObjectiveModel,@plotMinObjective});

I would like the number optimizableVariables to be variable, such that I could define fun as fun = @(x)myfun(x.(variable number of variables)).

The x. notation has been tripping me up here, I have tried using a cell array inside the function definition, which hasn't worked.

How could I achieve this?

Answer 1

[Ayush Aniket]

Hi Iwo Slodczyk,

You can use the cell array to store all the various variables using a custom function( getVariableValues) and then call the function inside the argument in your anonymous function in the following way:

% Define the variable names
variableNames = {'A', 'B', 'C'}; % Add more variable names as needed

% Create the optimizable variables
for i = 1:numel(variableNames)
    C{i} = optimizableVariable(variableNames{i}, [1 2], 'Transform', 'none');
end

% Create the function handle
fun = @(x) myfun(getVariableValues(x, variableNames));

% Call the bayesopt function
Steps = 100; % Number of steps for optimization
results = bayesopt(fun, [C{:}], 'AcquisitionFunctionName', 'expected-improvement-plus', ...
    'IsObjectiveDeterministic', true, 'ExplorationRatio', 0.1, 'NumSeedPoints', 10, ...
    'MaxObjectiveEvaluations', Steps, 'PlotFcn', {@plotObjectiveModel, @plotMinObjective});

% Define the helper function to retrieve variable values
function values = getVariableValues(x, variableNames)
    values = cell(1, numel(variableNames));
    for i = 1:numel(variableNames)
        values{i} = x.(variableNames{i}); %dynamic field reference 
    end
end

The code utilises the concept of dynamic field reference. See this documentation page for more information.

Hope this helps!

Answer 2

[Paolo Manfredi]

Hi @Iwo Slodczyk,

I don't know if this is still relevant, but I had been struggling with the same issue until I realized that bayesopt passes optimizable variables x in the form of a table (see Table class). Hence, when you pass your A and B values as x.A and x.B, you are actually passing the columns of table x. However, you can also access values in table columns as x{:,1}, x{:,2}, etc.

For example, your code could be rewritten as

C(1) = optimizableVariable('A',[1 2],'Transform','none');
C(2) = optimizableVariable('B',[1 2],'Transform','none');

results = bayesopt(@(x) myfun(x{:,1},x{:,2}),C,'AcquisitionFunctionName','expected-improvement-plus', ...
    'IsObjectiveDeterministic',true,'ExplorationRatio',0.1,'NumSeedPoints',10, ...
    'MaxObjectiveEvaluations',Steps,'PlotFcn',{@plotObjectiveModel,@plotMinObjective});

Note that the order of the variables in the columns of x is defined by their order in C.

In this way, you can dynamically allocate a different number of optimizable variable in C and recall them inside myfun using bracket indexing.