updating data

[Jan-Ole Koslik]

Hi all,

is there a simple way to update the data when using RTMB without remaking the AD function/ Tape every time?

So if I have

par = list(...)

data = list(...)

f = function(par){

   getAll(par, data)

   ...

}

can I use MakeTape() in combination with DataEval() to only replace the data in the objective?

Cheers,

Ole

[Kasper Kristensen]

RTMB has the capabilities to do this (example below) but there are some side effects to be aware of:

* The tape optimizer can do a better job when knowing the data can't change. It's therefore often worth waiting for the tape to be rebuilt with the new data rather than implementing data updates.

* The tape forward loop is only triggered when parameters change compared to previous evaluation. That is, if you change the data, and re-evaluate the likelihood for the same parameters as previous, nothing will happen (see example)!

* DataEval evaluates an R function and is therefore not thread safe. It therefore won't play well with a parallel version of RTMB.

library(RTMB)
mydat <- rivers
f <- function(p) {
    getDat <- function(x) {
        print("Get data")
        .GlobalEnv$mydat
    }
    ## dummy parameter (DataEval expects an argument)
    empty <- rep(p$mu, 0)
    ## Add R function to the AD tape
    y <- DataEval(getDat, empty)
    -sum(dnorm(y, p$mu, p$sd, log=TRUE))
}
obj <- MakeADFun(f, list(mu=0, sd=1))
obj$fn(obj$par)
obj$fn(obj$par) ## Same parameter as previous - nothing happens!
## Fit with mydat
opt <- nlminb(obj$par, obj$fn, obj$gr)
opt$par
## Refit with new data
mydat <- mydat * .1
opt <- nlminb(obj$par, obj$fn, obj$gr)
opt$par

[Jan-Ole Koslik]

Thanks Kasper,

can DataEval() only get numeric vectors or also lists?

When trying on my real example I get 

Error: Not compatible with requested type: [type=list; target=double].

[Kasper Kristensen]

Yes, it's limited to numeric vectors of fixed length. The first evaluation of DataEval() stores the length and checks that subsequent data lengths are the same.