Fix old and new GPR models - Githubissues

cmu-db / ottertune

The automatic DBMS configuration tool

Other

1.21k stars 310 forks source link

Fix old and new GPR models #285

Closed bohanjason closed 4 years ago

bohanjason commented 4 years ago

old GPR model:

Add noise variance (ridge) when computing covariance matrix. set the default value to 1.0
make hyperparameters (lengthscale, magnitude, ridge) trainable. (i.e. use tf.Variable instead of tf.constant)

new GPR model:

use standard deviation instead of variance when computing UCB (upper confidence bound)
change the kernel to Matern12 (exponential kernel), consistent with old GPR models.

dvanaken commented 4 years ago

@bohanjason LGTM thanks!