In CollinErickson/UGP: Interchangeable Gaussian Process Models

Trouble with laGP

I'm having trouble using laGP, specifically when updating the data with new data points. If I just use updateGPsep it looks bad, but if I follow it with jmleGPsep it often gives an error, and if it doesn't then the result doesn't look good.

Example, fit with mlegp

I've had some trouble getting good results from laGP.

Here's my function. Note that the values are all between -2 (blue) and +2 (pink).

library(contourfilled)
f1 <- function(x) {sin(8*pi*x[1]) + sin(8*pi*x[2])}
contourfilled.func(f1)

I'll create two datasets, the first with sample size 10 and the second with sample size 100

set.seed(1)
n <- 10
d <- 2
n2 <- 100
X1 <- matrix(runif(n*d),n,d)
Z1 <- apply(X1,1,f1)
X2 <- matrix(runif(n2*d),n2,d)
Z2 <- apply(X2,1,f1)
Xall <- rbind(X1,X2)
Zall <- c(Z1,Z2)

Here's what mlegp does with the first data set (understandably bad) and then all the data (looks good).

contourfilled.data(X1,Z1, xcontlim = 0:1, ycontlim = 0:1)
contourfilled.data(Xall,Zall, xcontlim = 0:1, ycontlim = 0:1)

Using laGP

Here I create a new GP giving it only the first data points.

set.seed(2)
library(laGP)
da <- darg(list(mle=TRUE), X=X1)
ga <- garg(list(mle=TRUE), y=Z1)

## initialize new separable GP
gpi <- newGPsep(X=X1, Z=Z1, d=da$start, g=ga$start, dK=TRUE)

## joint inference for lengthscale and nugget under prior calculated above
mle <- jmleGPsep(gpi, drange=c(da$min, da$max), grange=c(ga$min, ga$max), dab=da$ab, gab=ga$ab, verb=0)

Okay result for n=10 as shown below.

contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(X1, pch=19)

Now I want to update it adding in the second data set. It clearly updates the parameters in some way, but it is terrible (note the min/max values are not near -2/+2). I guess that means I have to tell it to jmleGPsep. (Sometimes I get an error here since all predictions are NaN or NA, I forget which.)

updateGPsep(gpsepi = gpi, X = Xall, Z = Zall)
contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

Here I do jmleGPsep, using updated values of da and ga. I put it in a try since a lot of the time it gives the error

Error in mleGPsep(gpi, tmin = da$min, tmax = da$max, ab = da$ab, verb = 0) :

L-BFGS-B needs finite values of 'fn'

Even with seeds set sometimes I get the error when I run it, sometimes I don't. If it works, it doesn't do well as the plot shows below. Otherwise it's the same thing as before.

da <- darg(list(mle=TRUE), X=Xall)
ga <- garg(list(mle=TRUE), y=Zall)
mle <- try(jmleGPsep(gpi, drange=c(da$min, da$max), grange=c(ga$min, ga$max), dab=da$ab, gab=ga$ab, verb=0))
print(mle)

contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

## clean up
deleteGPsep(gpi)

Let me try again, setting the nugget to 1e-3 so it doesn't have to estimate it

set.seed(2)
library(laGP)
da <- darg(list(mle=TRUE), X=X1)

## initialize new separable GP
gpi <- newGPsep(X=X1, Z=Z1, d=da$start, g=1e-3, dK=TRUE)

## joint inference for lengthscale and nugget under prior calculated above
mle <- mleGPsep(gpi, param="d", tmin=da$min, tmax=da$max, ab=da$ab, verb=0)

Okay result for n=10 again.

contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(X1, pch=19)

Update it with the new points, it updates something but is awful. Min/max values are bad again.

updateGPsep(gpsepi = gpi, X = Xall, Z = Zall)
contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

This time when I do mleGPsep it usually gives the error. The plot below will probably be the same as above since there was an error.

da <- darg(list(mle=TRUE), X=Xall)
#ga <- garg(list(mle=TRUE), y=Zall)
mle <- try(mleGPsep(gpi, tmin=da$min, tmax=da$max, ab=da$ab, verb=0))
print(mle)

contourfilled.func(function(xx){predGPsep(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

## clean up
deleteGPsep(gpi)

Let me try a third time, setting the nugget to 1e-3 and not using sep, also removing tmin, tmax, and ab

set.seed(2)
library(laGP)
da <- darg(list(mle=TRUE), X=X1)

## initialize new separable GP
gpi <- newGP(X=X1, Z=Z1, d=da$start, g=1e-3, dK=TRUE)

## joint inference for lengthscale and nugget under prior calculated above
#mle <- mleGP(gpi, param="d", tmin=da$min, tmax=da$max, ab=da$ab, verb=0)
mle <- mleGP(gpi, verb=0)

Okay result for n=10 again.

contourfilled.func(function(xx){predGP(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(X1, pch=19)

Update it with the new points, it updates something but is awful (bad min/max and shape).

updateGP(gpi = gpi, X = Xall, Z = Zall)
contourfilled.func(function(xx){predGP(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

This time when I do mleGPsep it usually gives the error. The plot below will probably be the same as above since there was likely an error.

da <- darg(list(mle=TRUE), X=Xall)
#ga <- garg(list(mle=TRUE), y=Zall)
#mle <- try(mleGP(gpi, tmin=da$min, tmax=da$max, ab=da$ab, verb=0))
mle <- try(mleGP(gpi,  verb=0))
print(mle)

contourfilled.func(function(xx){predGP(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

For some reason I get an error below when I try to deleteGP, not really a big issue.

## clean up
message(gpi)
try(deleteGP(gpi))

Can it fit well if I start it with all the points?

Now it has all 110 points to begin with.

set.seed(2)
library(laGP)
da <- darg(list(mle=TRUE), X=Xall)

## initialize new separable GP
gpi <- newGP(X=Xall, Z=Zall, d=da$start, g=1e-3, dK=TRUE)

## joint inference for lengthscale and nugget under prior calculated above
#mle <- mleGP(gpi, param="d", tmin=da$min, tmax=da$max, ab=da$ab, verb=0)
mle <- mleGP(gpi, verb=0)

Now it can fit the surface fine.

contourfilled.func(function(xx){predGP(gpi,XX=xx,lite=T)$mean},batchmax = 100)
points(Xall, pch=19)

## clean up
message(gpi)
try(deleteGP(gpi))

CollinErickson/UGP documentation built on Jan. 31, 2023, 11:26 a.m.