Backward difference numerical gradient approximation.

Share:

Description

grback computes the backward difference approximation to the gradient of user function userfn.

Usage

1
   grback(par, userfn, fbase=NULL, env=optsp, ...)

Arguments

par

parameters to the user objective function userfn

userfn

User-supplied objective function

fbase

The value of the function at the parameters, else NULL. This is to save recomputing the function at this point.

env

Environment for scratchpad items (like deps for approximation control in this routine). Default optsp.

...

optional arguments passed to the objective function.

Details

Package: grback
Depends: R (>= 2.6.1)
License: GPL Version 2.

Value

grback returns a single vector object df which approximates the gradient of userfn at the parameters par. The approximation is controlled by a global value optderiveps that is set when the package is attached.

Author(s)

John C. Nash

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
cat("Example of use of grback\n")

myfn<-function(xx, shift=100){
    ii<-1:length(xx)
    result<-shift+sum(xx^ii)
}

xx<-c(1,2,3,4)
ii<-1:length(xx)
print(xx)
gn<-grback(xx,myfn, shift=0)
print(gn)
ga<-ii*xx^(ii-1)
cat("compare to analytic gradient:\n")
print(ga)

cat("change the step parameter to 1e-4\n")
optsp$deps <- 1e-4
gn2<-grback(xx,myfn, shift=0)
print(gn2)