Make Bayesian Optimization (bayesopt) model positive

I am using bayesopt to minimize a function which can only give positive results. However, when bayesopt finishes it gives a best objective function estimate which is negative. Is there a way to enforce objective function model to be positive? I have tried coupled constraints and it didn't seem to work. Thanks

Answer 1

[Alan Weiss]

If I understand you correctly, your objective function is stochastic, but always returns positive values.

If that is true, then I suggest that you optimize the log of the objective function, and take exp(answer) as the value.

If your objective function is not stochastic, then I suggest that you set the 'IsObjectiveDeterministic' name-value pair to true and see if that fixes things.

Alan Weiss

MATLAB mathematical toolbox documentation

Answer 2

[Don Mathis]

Just to add to Alan's answer, bayesopt uses a Gaussian Process model to model the objective function, and Gaussian processes are inherently unbounded: The posterior distribution over Y at a given X is a Gaussian distribution. Often you can use this effectively for bounded functions (e.g., if your function value never gets close to zero), but if you want to be rigorous, you should transform your objective function into one which is unbounded. Optimizing log(y) is one way.