You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, I’m a student and learning about BayesianOptimization rencently. I’m trying to make fabolas compactible to George 0.3.1 and I think I did it. And I hope I can give you some suggestions:
I suggest that using stationary kernel (E.g. SE kernel) instead of non-stationary kernel (LinearKernel in Fabolas.py), because when you run get_incumbent() (in Fabolas.py), you will project the environment variables to 1, and then change to 0 because of _quadratic_bf(). Then you will run predict() in get_incumbent(), and the parameters of predict() will be matrix with env=0 (E.g. (a1,b1,0) ,(a2,b2,0) ,(a3,b3,0)...)
If you use LinearKernel, the var of predict() will be a zeroes, and mean will of predict() will be a vector with same elements.As a result, the epmgp.py cannot work
The parameter of EnvPrior() “n_lr=degree+1” can change to “n_lr= len(env_kernel) “
The text was updated successfully, but these errors were encountered:
Hi, I’m a student and learning about BayesianOptimization rencently. I’m trying to make fabolas compactible to George 0.3.1 and I think I did it. And I hope I can give you some suggestions:
I suggest that using stationary kernel (E.g. SE kernel) instead of non-stationary kernel (LinearKernel in Fabolas.py), because when you run get_incumbent() (in Fabolas.py), you will project the environment variables to 1, and then change to 0 because of _quadratic_bf(). Then you will run predict() in get_incumbent(), and the parameters of predict() will be matrix with env=0 (E.g. (a1,b1,0) ,(a2,b2,0) ,(a3,b3,0)...)
If you use LinearKernel, the var of predict() will be a zeroes, and mean will of predict() will be a vector with same elements.As a result, the epmgp.py cannot work
The parameter of EnvPrior() “n_lr=degree+1” can change to “n_lr= len(env_kernel) “
The text was updated successfully, but these errors were encountered: