CODEWITHSUNDEEP
rmathematical-optimizationnonlinear-optimization
Ray
Ray

Reputation: 123

Optim function returning wrong solution

I'm trying to get a function minimum with optim in R, the problem is that the par solution returned by the method is not correct and change if i introduce a different initial point. I will show you some code and the output in getting. What am I missing?

> funx = function(vx) (vx[1] -2)^4 + (vx[1]-2*vx[2])^2 + -10/(vx[1]^2-vx[2])
> optim(c(0,1), funx)
$par
[1] 0.7080021 1.5315815

$value
[1] 18.03898

$counts
function gradient 
      61       NA 

$convergence
[1] 0

$message
NULL

> optim(c(0,3), funx)
$par
[1] 1.271924 1.617791

$value
[1] -4.5036e+16

$counts
function gradient 
     237       NA 

$convergence
[1] 0

$message
NULL

> funx(c(1.271924, 1.617791))
[1] 29566209

Upvotes: 1

Views: 1466

Answers (2)

Ben Bolker
Ben Bolker

Reputation: 226027

Indeed, this is a rounding problem. If you feed the exact values back to funx() you get a small (i.e. very negative) value: if you round to six digits you get a large value.

oo <-optim(c(0,3), funx)
funx(oo$par)
## [1] -4.5036e+16
 funx(round(oo$par,6))
## [1] 29566209

This is indeed an ill-behaved function, you might need to take this into account when working with it. (I'd also strongly suggest using derivative-based optimization, either calculating the derivatives yourself or using deriv().)

Upvotes: 3

Justin
Justin

Reputation: 374

You may add, for instance, method = "L-BFGS-B" in your optim and define the upper and lower limits of your parameters depending on your definition of the function. It may produce values out of their original intervals.

Upvotes: 0

Related Questions