This class provides easy access to build and evaluate surrogate models using Ax.
If an AxModelManager instance is created from an ExplorationDiagnostics object using the build_gp_model method, it will take automatically the varying parameters in ExplorationDiagnostics as the model parameters.
Sof2 is just 2 * f.
However, since in the example we told optimas to find the minimum of f and the maximum f2, the best evaluations fall in opposite regions.
Let's plot the model of 'f' together with the real function:
We can see that, while the oscillations in x1 are well captured by the model, this is not the case for x0 where the model just reproduces the "averaged" trend. This is surely due to the less dense sampling in the x0 dimension.
Draws a slice of the model along x1 with x0 fixed to its middle point.
This class provides easy access to build and evaluate surrogate models using
Ax
. If anAxModelManager
instance is created from anExplorationDiagnostics
object using thebuild_gp_model
method, it will take automatically the varying parameters inExplorationDiagnostics
as the model parameters.This an example for one of the tests in
Optimas
:There are two objectives:
and two varying parameters:
Build (Gaussian Process) model for
f
from Optimas diagnostics data:Evaluate the model over the 10 best scoring
f
evaluations in the exploration diagnostics:Note that the best scoring evaluation in data (index=98) does not coincide with the best scoring according to the model (index=54).
Plot the model for
f
(mean and standard error), mark with crosses the top 10 evaluations and add their trial indices:One can also build a model for other objectives or analyzed parameters, e.g.:
Or a model with multiple metrics. This example use all the objectives present in the Optimas diagnostics:
Evaluate the model for
f2
over the 10 best scoringf2
evaluations in the exploration diagnostics:Plot the mean of the two models (
f
andf2
):The two models for
f
andf2
are very similar. The reason is that the real underlying function forf
andf2
used in the tests is:So
f2
is just 2 *f
. However, since in the example we told optimas to find the minimum off
and the maximumf2
, the best evaluations fall in opposite regions.Let's plot the model of 'f' together with the real function:
We can see that, while the oscillations in
x1
are well captured by the model, this is not the case forx0
where the model just reproduces the "averaged" trend. This is surely due to the less dense sampling in thex0
dimension.Draws a slice of the model along
x1
withx0
fixed to its middle point.