search

NIEHS / PrestoGP

Penalized Regression on Spatiotemporal Outcomes using Gaussian Processes a.k.a. PrestoGP
https://niehs.github.io/PrestoGP/
0 stars 0 forks source link

Evaluate the effect of initial theta estimates #22

Open ericbair-sciome opened 7 months ago

ericbair-sciome commented 7 months ago

The current initial theta estimates are arbitrary and probably suboptimal. I have noticed that the estimates for the range parameters are often really bad in spatiotemporal simulations. I wonder if this is due to the fact that the initial estimates are bad and the algorithm never corrects itself. As a first step, it would be worthwhile to feed the procedure more accurate initial estimates and see how it affects the accuracy of the model. If it results in significant improvement, then we should try to find a more intelligent way to choose the initial theta's.

kyle-messier commented 7 months ago

@ericbair-sciome Is this a current problem? I think understanding the effects of [default] initial values for both the thetas and betas is critical. @sciome-bot Switching to version 1.0

ericbair-sciome commented 7 months ago

We did do some research on the initial beta estimates. We found that a better choice of initial beta's did speed things up significantly, but it made almost no difference to the final parameter estimates. My suspicion is that the initial theta estimates are going to make even less difference, but I do want to do some testing to confirm that.