Change past observation history while using bayesopt.m

Hi,

I am working on using bayesopt (Bayesian optimization) function in a real-time setting.

My goal is to modify past observation data while the optimization is running.

For example, at n-th iteration, the optimizer determines next input (u(n+1)) based on past histories (inputs: u(1:n-1) and observations: y(1:n-1)) and currnet input and observation (u(n) and y(n)). However, before deciding next input (u(n+1)), I want to modifiy previous observation value (y(n-1)) to different value.

I attemped to use 'OutputFcn' option in bayesopt, but the previous observations that I can access using this options are read-only variables.

Is there any possible way that I can change previous observation while running bayesopt?

Answer 1

[Aneela]

Hi @GilHwan,

“bayesopt” treats past observations as immutable. It is not possible to modify them while running “bayesopt”.

However, to access and modify past observation variables you can consider Custom Bayesian Optimization Loop.

• Use MATLAB's Gaussian Process functions (“fitrgp”, etc.) to model the objective function and predict to determine the next point to evaluate.
• In each iteration, you can manually modify any past observations before updating the model.

Refer to the below reference custom code for Bayesian Optimization loop:

objectiveFcn = @(x) (x - 2).^2 + sin(5 * x); 
maxIterations = 5; 
numInitialPoints = 5; 
bounds = [0, 5]; 
u_history = linspace(bounds(1), bounds(2), numInitialPoints)'; 
y_history = objectiveFcn(u_history); 
% Bayesian optimization loop 
for n = 1:maxIterations 
    gpModel = fitrgp(u_history, y_history, 'KernelFunction', 'matern52', 'Standardize', true);  
    u_next = determineNextInput(gpModel, bounds); 
    y_next = objectiveFcn(u_next); 
    % Modify previous observation if needed 
    if n > 1 
        y_history(end) = modifyObservation(y_history(end)); 
    end 
    u_history = [u_history; u_next]; 
    y_history = [y_history; y_next]; 
    fprintf('Iteration %d: u = %.4f, y = %.4f\n', n, u_next, y_next); 
end 
% determine next input using Expected Improvement 
function u_next = determineNextInput(gpModel, bounds) 
acquisitionFcn = @(x) -expectedImprovement(x, gpModel); 
options = optimoptions('fmincon', 'Display', 'off'); 
u_next = fmincon(acquisitionFcn, mean(bounds), [], [], [], [], bounds(1), bounds(2), [], options); 
end 
% Expected Improvement acquisition function 
function ei = expectedImprovement(x, gpModel) 
[mu, sigma] = predict(gpModel, x); 
y_best = min(gpModel.Y); 
z = (y_best - mu) ./ sigma; 
ei = (y_best - mu) .* normcdf(z) + sigma .* normpdf(z); 
ei(sigma == 0) = 0; 
end 
% Modify the observation 
function y_modified = modifyObservation(y) 
y_modified = y + randn * 0.01; % Example modification 
end 

Hope this helps!

[GilHwan]

Hi Aneela,

Thank you for your help.

I may try to modify fitrgp function to update past observations.