[R] Noisy objective functions

[Enrico Schumann]

On Sun, 13 Aug 2023, Hans W writes:

> While working on 'random walk' applications, I got interested in
> optimizing noisy objective functions. As an (artificial) example, the
> following is the Rosenbrock function, where Gaussian noise of standard
> deviation `sd = 0.01` is added to the function value.
>
>     fn <- function(x)
>               (1+rnorm(1, sd=0.01)) * adagio::fnRosenbrock(x)
>
> To smooth out the noise, define another function `fnk(x, k = 1)` that
> calls `fn` k times and returns the mean value of those k function
> applications.
>
>     fnk <- function(x, k = 1) {     # fnk(x) same as fn(x)
>         rv = 0.0
>         for (i in 1:k) rv <- rv + fn(x)
>         return(rv/n)
>     }
>
> When we apply several optimization solvers to this noisy and smoothed
> noise functions we get for instance the following results:
> (Starting point is always `rep(0.1, 5)`, maximal number of iterations 5000,
>  relative tolerance 1e-12, and the optimization is successful if the
> function value at the minimum is below 1e-06.)
>
>       k   nmk       anms neldermead     ucminf optim_BFGS
>      ---------------------------------------------------
>       1  0.21       0.32       0.13       0.00       0.00
>       3  0.52       0.63       0.50       0.00       0.00
>      10  0.81       0.91       0.87       0.00       0.00
>
> Solvers: nmk = dfoptim::nmk, anms = pracma::anms [both Nelder-Mead codes]
>          neldermead = nloptr::neldermead,
>          ucminf = ucminf::ucminf, optim_BFGS = optim with method "BFGS"
>
> Read the table as follows: `nmk` will be successful in 21% of the
> trials, while for example `optim` will never come close to the true
> minimum.
>
> I think it is reasonable to assume that gradient-based methods do
> poorly with noisy objectives, though I did not expect to see them fail
> so clearly. On the other hand, Nelder-Mead implementations do quite
> well if there is not too much noise.
>
> In real-world applications, it will often not be possible to do the
> same measurement several times. That is, we will then have to live
> with `k = 1`. In my applications with long 'random walks', doing the
> calculations several times in a row will really need some time.
>
> QUESTION: What could be other approaches to minimize noisy functions?
>
> I looked through some "Stochastic Programming" tutorials and did not
> find them very helpful in this situation. Of course, I might have
> looked into these works too superficially.
>

Since Nelder--Mead worked /relatively/ well: in my
experience, its results can sometimes be improved by
restarting it, i.e. re-initializing the simplex.  See this
note here: http://enricoschumann.net/R/restartnm.htm ,
which incidentally also uses the Rosenbrock function.

Just for the record, I think Differential Evolution (DE)
can handle such problems well, though it would usually
need more iterations.  As computational "proof" ;-) , I
run DE 50 times for the k == 1 case, and each time store
whether the resulting objective-function value is below
1e-6.  I let DE take way more function evaluations
(population of 100 times 500 generations = 50000 function
evaluations); but it gets a value below 1e-6 in all 50
cases.


    library("NMOF")
    ofv.below.threshold <- logical(50)
    for (i in seq_along(ofv.below.threshold)) {
        sol <- DEopt(fnk,
                     algo = list(
                         nP = 100, nG = 500,
                         min = rep(  5, 5), max = rep(   10, 5)))
        
        ofv.below.threshold[i] <- sol$OFvalue < 1e-6
    }
    sum(ofv.below.threshold)/length(ofv.below.threshold)


(These 50 runs take less than half a minute on my
machine.) Note that I have on purpose initialized the
population in the range 5 to 10, i.e. way off the optimum.




-- 
Enrico Schumann
Lucerne, Switzerland
http://enricoschumann.net