CODEWITHSUNDEEP
roptimizationnonlinear-optimization
Adam
Adam

Reputation: 65

Setting up non-linear optimization in R

I have a function that calculates the 'value' of a model based on the inputs.

ModelValue <- function(x, y, z) {
#long, most likely non-linear stuff
return(Value)}

x takes is a number between 0-1, and is fixed based on the model I'm testing. Ie, x will just happen to be 0.56.

y and z are groups of 4 thresholds that I want to optimize to maximize Value.

y=c(pa, pb, pc, pd) and z=c(ta, tb, tc, td), subject to:
all are => 0.001,
all are =<0.997,
pa + pb + pc + pd == 1; and
ta + tb + tc + td == 1

I've tried looking at optimx and spg; (this is as far as I could get) but no matter what I do, I can't seem to create an optimization function that doesn't freak out as soon as I try and mention pa, pb, pc etc. I don't understand how to tell the model what variables its meant to be optimizing...

OptimizeModel <- function(x) {
p0=1 #initial guess
fn = ModelValue(x, y, z)

lo <- c(0.001, 0.001, 0.001, 0.001,
          0.001, 0.001, 0.001, 0.001) # lower limits for parameters
hi <- c(0.997, 0.997, 0.997, 0.997,
          0.997, 0.997, 0.997, 0.997)  # upper limits for parameters

y<- c(pa, pb, pc, pd)
z<- c(ta, tb, tc, td)
pa + pb + pc + pd = 1
ta + tb + tc + td = 1
# pa > 0.001
# pb > 0.001 
# pc > 0.001 
# pd > 0.001 
# ta > 0.001 
# tb > 0.001 
# tc > 0.001 
# td > 0.001 

ans1 <- spg(par=p0,
            fn=fn,
            lower=lo,
            upper=hi,
            control=list(maximize=TRUE, trace=FALSE))

return (ans1)  
}

When I try it:

OptimizeModel(0.56)
#Error in OptimizeModel(0.56) : object 'pa' not found

Upvotes: 1

Views: 89

Answers (1)

Erwin Kalvelagen
Erwin Kalvelagen

Reputation: 16724

Just pass on pa,pb,pc to spg and use them in the function evaluation. The documentation of spg is showing an example where this is done:

valley.f <- function(x, cons) {
...}

Here x are the decision variables and cons is extra data (constants).

The call to spg looks like:

ans.spg2 <- spg(par=p0, fn=valley.f, cons=k, method=2)

You see cons=k is specified. This argument is not for spg itself but is passed on to the function evaluation (and gradient evaluation) functions. You need to do the same for pa,pb,pc.

See: https://www.rdocumentation.org/packages/BB/versions/2019.10-1/topics/spg

Upvotes: 1

Related Questions