suggested modification to the 'mle' documentation?
2007/12/7, Ben Bolker <bolker at ufl.edu>:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Luke Tierney wrote:
On Fri, 7 Dec 2007, Duncan Murdoch wrote:
For working at the general likelihood I think is is better to
encourage the approach of definign likelihood constructor functions.
The problem with using f, data is that you need to mathc the names
used in f and in data, so either you have to explicitly write out f
with the names you have in data or you have to modify data to use the
names f likes -- in the running example think
f <- function(lambda) -sum(dpois(x, lambda, log=T))
d <- data.frame(y=rpois(10000, 12.34))
somebody has to connext up the x in f with the y in d. With a negative
log likelihood constructor defines, for example, as
makePoisonNegLogLikelihood <- function(x)
function(lambda) -sum(dpois(x, lambda, log=T))
this happens naturally with
makePoisonNegLogLikelihood(d$y)
I hate to sound like a jerk, but I do hope that in the end we come up with a solution that will still be accessible to people who don't quite have the concept of writing functions to produce functions. I feel it is "natural" for people who have multiple data sets to have the variables named similarly in different data sets. All of the constructor stuff is still accessible to anyone who wants to use the function that way ... is there any way to do a cheesy default constructor that is just equivalent to taking the likelihood function and arranging for it to be evaluated in an environment containing the data? That way if "nllmaker" below were just a formula or a log-likelihood function it could still work ...
I don't really agree with this.
I found really natural writing functions which builds other functions,
for handling
in a clean way the data-dependency problem, much more than
manipilating function environments.
As a useR, I think that if I'm able to write a likelihood function myself:
data <- whatever
negloglik <- function(theta)
a + very * complicated / function - of %% theta %*% and %o% data
to be used in mle, I'm also good at abstracting it a bit this way:
nllmaker <- function(data)
function(theta)
a + very * complicated / function - of %% theta %*% and %o% data
negloglik <- nllmaker(whatever),
don't you think? I use this kind of tricks routinely for simulations.
In general, I think it should be more emphatized functional style in R coding.
In fact, I like a lot the recent introduction of some higher order
functions in the base package (Reduce, Filter, Map).
Bests,
Antonio, Fabio.
[snip]
Both (simple) bootstrapping and (simple leave-one-out) crossvalidation
require a data structure with a notion of cases, which is much more
restrictive than the conext in which mle can be used. A more ngeneric
aproach to bootstrapping that might fit closer to the level of
generality of mle might be parameterized in terms of a negative log
likelihood constructor, a starting value constructor, and a resampling
function, with a single iteration implemented soemthing like
mleboot1 <- function(nllmaker, start, esample) {
newdata <- resample()
newstart <- do.call(start, newdata)
nllfun <- do.call(nllmaker, newdata)
mle(fnllfun, start = newstart)
}
This would leave decisions on the resampling method and data structure
up to the user. Somehing similar could be done with K-fold CV.
luke
-----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.6 (GNU/Linux) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org iD8DBQFHWcS1c5UpGjwzenMRAig2AJ9iTzhI1p8tBb7Q15jgT4nA+Zds+gCgggc2 sI2que28Hl1M5cVGa+anEL0= =hCiS -----END PGP SIGNATURE-----
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Antonio, Fabio Di Narzo Ph.D. student at Department of Statistical Sciences University of Bologna, Italy