SheffieldML / GPyOpt

Gaussian Process Optimization using GPy
BSD 3-Clause "New" or "Revised" License
927 stars 261 forks source link

Not feasible points - constraints not respected #156

Open LEleonora opened 6 years ago

LEleonora commented 6 years ago

Hi,

currently a batch solution can contain infeasible points in case of batch generation and local penalization (LP). The constrains are checked when the acquisition function is evaluated, but then the acquisition function value is penalized for the local penalization strategy. I would suggest to consider the case of not feasible points as a separate case so that these points will not be consider as possible minima in the next steps. For example, if the constraints are not respected, the penalized acquisition function value is set to infinite. Would this be a proper solution?

Thank you.

oliversheridanmethven commented 6 years ago

I think I have encountered a similar solution, although I had constructed (naively?) functions which throw assertions then infeasible points are evaluated.

cf Suggested point is infeasible #94