Easy way to handle positive objective functions #1465
-
|
I am trying to minimise wall-time for a simulation. I know that the value would always be positive and I must somehow enforce it on my GP model. As a novice in coding, I couldn't find an easy enough solution for this that I could implement. Some tutorials do this but I wasn't able to use those for my problem. My first question is, what is the easiest way to constrain the objective to be above a fixed known value? For a certain X input, If I could model the Y output wall-time as a Poission distribution or a Weibull distribution, all my problems relating to unrealistic wall-functions will be solved. Based on what I have been reading, too many things are made simple with using GP models. So just to confirm, my second question is, is it possible to easily switch the GP model to a Poisson Process model? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
|
One way to model a known positive value with GP is to model the logarithm of the value with a GP instead. So, the
We do not have a Poisson Process model that is available out of box in BoTorch. You would have to implement one yourself. |
Beta Was this translation helpful? Give feedback.
-
|
Great idea! I forgot about this trick. |
Beta Was this translation helpful? Give feedback.
-
|
I am trying to do |
Beta Was this translation helpful? Give feedback.
-
|
Also, Log transform for noisy observations doesn't seem to be implemented yet. I have the mean and variance data for each of the observations. |
Beta Was this translation helpful? Give feedback.
-
|
I believe that's because there isn't an analytic formula for log-transforming the variance observations. If we assume that the original observations have normally distributed observation noise, then the log-transformed observations will not have a normally distributed observation noise (which is assumed by GP). |
Beta Was this translation helpful? Give feedback.
One way to model a known positive value with GP is to model the logarithm of the value with a GP instead. So, the
log Ycan take any value from(-inf, inf)but theY = exp ( log Y )will be restricted to(0, inf). You can use similar transforms to encode greater than a known constant type constraints.We do not have a Poisson Process model that is available out of box in BoTorch. You would have to implement one yourself.