Recently Marie Boehnstedt reported a bug in the nlm() function for function minimization when both gradient and hessian are provided. She has provided a work-around for some cases and it seems this will get incorporated into the R function eventually.
However, despite the great number of packages on CRAN, there does not appear to be a straightforward Newton approach to function minimization. This may be because providing the code for a hessian (the matrix of second derivatives) is a lot of work and error-prone. (R could also use some good tools for Automatic Differentiation). I have also noted that a number of researchers try to implement textbook methods and run into trouble when maths and computing are not quite in sync. Therefore, I wrote a simple safeguarded Newton and put a small package on R-forge at https://r-forge.r-project.org/R/?group_id=395 Note that Newton's method is used to solve nonlinear equations. In fact, for function minimization, we apply it to solve g(x) = 0 where g is the gradient and x is the vector of parameters. In part, safeguards ensure we reduce the function f(x) at each step to avoid some of the difficulties that may arise from a non-positive-definite hessian. In the package, I have a very simple quadratic test, the Rosenbrock test function and the Wood test function. The method fails on the last function -- the hessian is not positive definite where it stops. Before submitting this package to CRAN, I would like to see its behaviour on other test problems, but am lazy enough to wish to avoid creating the hessian code. If anyone has such code, it would be very welcome. Please contact me off-list. If I get some workable examples that are open for public view, I'll report back here. John Nash ______________________________________________ R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
