Skip to content

Fix #201: focussearch now handles discrete vector parameters #202

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

Fix #201: focussearch now handles discrete vector parameters #202

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5b32e6e
Fix #201: focussearch now handles discrete vector parameters
Mar 5, 2016
d451c63
More concise handling of discrete vector params in infillOptFocus
Mar 7, 2016
ecdeca8
Cleaning up code.
Mar 8, 2016
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
77 changes: 58 additions & 19 deletions R/infillOptFocus.R
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,14 @@ infillOptFocus = function(infill.crit, models, control, par.set, opt.path, desig
for (restart.iter in seq_len(control$infill.opt.restarts)) {
# copy parset so we can shrink it
ps.local = par.set

# handle discrete vectors:
# The problem is that for discrete vectors, we can't adjust the values dimension-wise.
# Therefore, for discrete vectors, we always drop the last level and instead have a
# mapping that maps, for each discrete vector param and for each dimension, from
# the sampled value (levels 1 to n - #(dropped levels)) to levels with random dropouts.
Copy link
Member

@jakob-r jakob-r Mar 8, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I only maybe understand what you mean here 😃

discreteVectorMapping = lapply(filterParams(par.set, type = c("discretevector", "logicalvector"))$pars,
Copy link
Member

@jakob-r jakob-r Mar 8, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

IMHO you can store filterParams(par.set, type = c("discretevector", "logicalvector")) as ps.disic.log.vec or something, as you use it twice here. Then you could also put everything up to line 39 in if (!isEmpty(ps.disc.log.vec) if i am not mistaken.

function(param) rep(list(setNames(names(param$values), names(param$values))), param$len))
Copy link
Member

@jakob-r jakob-r Mar 8, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe a comment would help "For each discrete/logic vector of length n create n replications of vector, each containing all possible values"?


# do iterations where we focus the region-of-interest around the current best point
for (local.iter in seq_len(control$infill.opt.focussearch.maxit)) {
Expand All @@ -21,13 +29,22 @@ infillOptFocus = function(infill.crit, models, control, par.set, opt.path, desig

# convert to param encoding our model was trained on and can use
newdesign = convertDataFrameCols(newdesign, ints.as.num = TRUE, logicals.as.factor = TRUE)
y = infill.crit(newdesign, models, control, ps.local, design, iter, ...)

# handle discrete vectors
for (dvparam in filterParams(par.set, type = c("discretevector", "logicalvector"))$pars) {
for (dimnum in seq_len(dvparam$len)) {
dfindex = paste0(dvparam$id, dimnum)
mapping = discreteVectorMapping[[dvparam$id]][[dimnum]]
levels(newdesign[[dfindex]]) = mapping[levels(newdesign[[dfindex]])]
}
}
y = infill.crit(newdesign, models, control, par.set, design, iter, ...)
Copy link
Member

@jakob-r jakob-r Mar 7, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is potentially slow and doesn't make use of our helpers. have a look at

par.set.disc.logic.v = filterParams(par.set, type = c("discretevector", "logicalvector"))
dfindex = getParamIds(par.set.disc.logic.v, with.nr = TRUE, repeated = TRUE)


# get current best value
local.index = getMinIndex(y, ties.method = "random")
local.y = y[local.index]
local.x.df = newdesign[local.index, , drop = FALSE]
local.x.list = dfRowToList(recodeTypes(local.x.df, ps.local), ps.local, 1)
local.x.list = dfRowToList(recodeTypes(local.x.df, par.set), par.set, 1)

# if we found a new best value, store it
if (local.y < global.y) {
Expand All @@ -39,24 +56,46 @@ infillOptFocus = function(infill.crit, models, control, par.set, opt.path, desig
ps.local$pars = lapply(ps.local$pars, function(par) {
# only shrink when there is a value
val = local.x.list[[par$id]]
if (!isScalarNA(val)) {
if (isNumeric(par)) {
# shrink to range / 2, centered at val
range = par$upper - par$lower
par$lower = pmax(par$lower, val - (range / 4))
par$upper = pmin(par$upper, val + (range / 4))
if (isInteger(par)) {
par$lower = floor(par$lower)
par$upper = ceiling(par$upper)
if (isScalarNA(val)) {
return(par)
}
if (isNumeric(par)) {
# shrink to range / 2, centered at val
range = par$upper - par$lower
par$lower = pmax(par$lower, val - (range / 4))
par$upper = pmin(par$upper, val + (range / 4))
if (isInteger(par)) {
par$lower = floor(par$lower)
par$upper = ceiling(par$upper)
}
} else if (isDiscrete(par)) {
# randomly drop a level, which is not val
if (length(par$values) <= 1L) {
return(par)
}
# need to do some magic to handle discrete vectors
if (par$type %nin% c("discretevector", "logicalvector")) {
val.names = names(par$values)
# remove current val from delete options, should work also for NA
val.names = val.names[!sapply(par$values, identical, y=val)] # remember, 'val' may not even be a character
to.del = sample(val.names, 1)
par$values[to.del] = NULL
} else {
# we remove the last element of par$values and a random element for
# each dimension in discreteVectorMapping.
par$values = par$values[-length(par$values)]
if (par$type != "logicalvector") {
# for discretevectorparam val would be a list; convert to character vector
val = names(val)
}
} else if (isDiscrete(par)) {
# randomly drop a level, which is not val
if (length(par$values) > 1L) {
val.names = names(par$values)
# remove current val from delete options, should work also for NA
val.names = setdiff(val.names, val)
to.del = sample(seq_along(val.names), 1)
par$values = par$values[-to.del]
for (dimnum in seq_len(par$len)) {
val.names = discreteVectorMapping[[par$id]][[dimnum]]
newmap = val.names
val.names = val.names[val.names != val[dimnum]]
to.del = sample(val.names, 1)
newmap = newmap[newmap != to.del]
names(newmap) = names(par$values)
discreteVectorMapping[[par$id]][[dimnum]] <<- newmap
}
}
}
Expand Down
5 changes: 5 additions & 0 deletions tests/testthat/test_infill_opt_focus.R
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,13 +68,18 @@ test_that("complex param space, dependencies, focusing, restarts", {
if(x$disc2 == 'a') tmp3 = log(x$realA) + x$intA^4 + ifelse(x$discA == 'm', 5, 0)
if(x$disc2 == 'b') tmp3 = exp(x$realB) + ifelse(x$discB == 'R', sin(x$realBR), sin(x$realBNR))
if(x$disc2 == "c") tmp3 = 500
assert(is.list(x$discVec))
assert(x$discVec[[1]] %in% c("a", "b", "c"))
assert(x$discScal %in% c("x", "y", "z"))
tmp1 + tmp2 + tmp3
},
par.set = makeParamSet(
makeNumericParam("real1", lower = 0, upper = 1000),
makeIntegerParam("int1", lower = -100, upper = 100),
makeNumericVectorParam("realVec", len = 10, lower = -50, upper = 50),
makeIntegerVectorParam("intVec", len = 3, lower = 0, upper = 100),
makeDiscreteVectorParam("discVec", len = 3, c(x = "a", y = "b", z = "c")),
makeDiscreteParam("discScal", c(a = "x", b = "y", c = "z")),
makeNumericParam("real2", lower = -1, upper = 1),
makeDiscreteParam("disc1", values = c("foo", "bar"), requires = quote(real2 < 0)),
makeNumericParam("real3", lower = -100, upper = 100, requires = quote(real2 > 0)),
Expand Down