Open JanRotti opened 1 year ago
One thing that is noticeable is that your test example is deterministic – y is just the same of x_1 and x_2, with no added noise. This could explain why the optimizer runs into covariance matrices that are not positive definite, and why you run into trouble with predictions. Two things to try (separately):
y
and y_test
optimize!(gp; noise=false)
)You were totally right! Having noise=false
does solve the problem. Still, I am wondering about the behavior of optimize
beforehand. Shouldn't optimize!(gp)
at least be consistent in its result? (So that it does not give $n-1$ times PosDefException
and then on the $n$-th cycle works.)
Dear all,
This is somewhat related to Issue #221. I was testing around with optimizing hyperparameters using
optimize!(gp)
; I get a lot ofLinearAlgebra.PosDefException
. When rerunning justoptimize!(gp)
, at some point the errors disappear and the optimization converges. But the model performs quite worse on test data, compared to before. After optimization themll
/target
values are actually reduced by the optim.I would assume if the error occurs it should be consistent and not disappear with rerunning. Is this behavior expected?
Here is a minimal example to reproduce:
Greetings, Jan