Brent's PRAXIS minimizer is available in FORTRAN 77.        July 1995

"Algorithms for Minimization Without Derivatives"
by Richard P. Brent, Prentice-Hall, 1973
ISBN: 0-13-022335-2

This book by Brent was a groundbreaking effort.
(I believe that it was his Ph.D. thesis at Stanford.)
His algorithms for finding roots and minima in
one dimension have good performance for typical problems
and guaranteed performance in the worst case.
(A later rootfinder by J. Bus and Dekker gave
a much lower bound for the worst case,
but no better performance in typical problems.)
These algorithms were implemented in both ALGOL W
and FORTRAN by Brent, and have been used fairly widely.

Brent also gave a multi-dimensional minimization algorithm,
PRAXIS, but only shows an implementation in ALGOL W.
This routine has not been widely used, at least in the U.S.
The PRAXIS package has been translated into FORTRAN 
by Rosalee Taylor, Sue Pinski, and me, and 
I am making it available via anonymous ftp for use as
freeware (please do not remove our names).

   ftp a.cs.okstate.edu
   anonymous
   [enter your userid as password]
   cd /pub/jpc
   get praxis.f
   quit

 
Brent's method and its performance

Newton's method for minimization can find the minimum of a
quadratic function in one iteration, but is sometimes not
convenient to use.  In the 1960s, several researchers found
iterative methods that solve quadratic problems exactly in a
finite number of steps.  C. S. Smith (1962) and
M. J. D. Powell (1964) devised methods
that had this property and did not require derivatives.
G. W. Stewart modified the Davidon-Fletcher-Powell quasi-Newton
method to use finite difference approximations to approximate
the gradient.  Powell's method, or later versions by Zangwill,
were the most successful of the early direct search methods
having the property of finite convergence on quadratic functions.

Powell's method was programmed at Harwell as subroutine VA04A,
and is available as file va04a.f in the same directory as praxis.f.
VA04A is not extremely robust, and can give underflow, overflow,
or division by zero.  va04a.f has several documented patches in it
where I tried to get around various abnormal terminations.
I do not recommend VA04A very strongly. 

Brent's PRAXIS added orthogonalization and several other features
to Powell's method.  Brent also dealt carefully with roundoff.

William H. Press et al. in their book "Numerical Recipes"
comment that
"Brent has a number of other cute tricks up his sleeve,
and his modification of Powell's method is probably
the best presently known."

Roger Fletcher was less enthusiastic in his review of Brent's book
in The Computer Journal 16 (1973) 314:
"... I am not convinced that the modifications to Powell's
method are the best.  Use of eigenvector directions
is not independent of scale changes to the variables,
and the use of searches in random directions is hardly
appealing.  Nonetheless all the algorithms are demonstrated
to be competitive by numerical examples."

The methods of Powell, Brent, et al. require that the function
for which a local minimum is sought must be smooth;
that is, the function and all of its first partial derivatives
must be continuous.
 
Brent compared his method to the methods of Powell, of Stewart,
and of Davies, Swann, and Campey.  Indirectly, he compared it
also to the Davidon-Fletcher-Powell quasi-Newton method.
He found that his method was about as efficient as the best
of these in most cases, and that it was more robust than others
in some cases.  (Pages 139-155 in Brent's book give fair 
comparisons to other methods.  The results in Table 7.1 on
page 138 are correct, but do not include progress all the way
to convergence, and are therefore not too useful.)

On least squares problems, all of these general minimization
methods are likely to be inefficient compared to least squares
methods such as the Gauss-Newton or Marquardt methods.
  
In addition to the scale dependence that Fletcher deplored,
PRAXIS also had the disadvantage that it required N, the number
of parameters, to be greater than or equal to two.
The failure to handle N=1 is an unnecessary and pointless limitation.


The FORTRAN version

We have followed Brent's PRAXIS rather closely.
I have added a patch to try to handle the case N=1,
and an option to use a simpler pseudorandom number generator,
DRANDM.  The handling of N=1 is not guaranteed.

The user writes a main program and a function subprogram
to compute the function to be minimized.
All communication between the user's main program and PRAXIS
is done via COMMON, except for an EXTERNAL parameter giving
the name of the function subprogram.
The disadvantages of using COMMON are at least two-fold:

   1)  Arrays cannot have adjustable dimensions.

   2)  Because some actual parameters are COMMON variables,
       the FORTRAN version of PRAXIS probably will not pass
       the Bell Labs PFORT package as being 100% standard FORTRAN.
       Nevertheless, this usage will not cause any conflict in
       any commercial FORTRAN compiler ever written.
       (If it does, I will apologize and rewrite PRAXIS.)

The advantage of using COMMON is that it is not necessary to pass
about fifteen more parameters every time the user calls PRAXIS.
At present all arrays are dimensioned (20) or (20,20),
and this can easily be increased using two simple global editing
commands.  (In this case, increase the value of NMAX.)

There are no DATA statements in PRAXIS, and it was not necessary
to use any SAVE statements.

We have used DOUBLE PRECISION for all floating point computations,
as Brent did.  We recommend using DOUBLE PRECISION on all computers
except possibly Cray computers, in which REAL is reasonably precise.
The value of "machine epsilon" is computed in subroutine PRASET
using bisection, and is called EPSMCH.
Brent computes EPSMCH**4 and 1/EPSMCH**4 in PRAXIS,
and uses these quantities later.  
Because EPSMCH in DOUBLE PRECISION is less than 1E-16,
these fourth powers of EPSMCH and 1/EPSMCH will underflow
and overflow on such machines as VAXs and PCs,
which have a range of only about 1E38, grossly insufficient
for scientific computation.  For such machines, Brent recommends
increasing the value of EPSMCH.  
EPSMCH=1E-9 or possibly even 1E-8 might be necessary.
A better solution would be to eliminate the explicit use of
these fourth powers, accomplishing the same result implicitly.

A "bug bounty" of $10 U.S. will be paid by me for the first 
notification of any error in PRAXIS.
The same bounty also applies to any substantive poor design
choice (having no redeeming advantages whatever) in the FORTRAN
package.  (The patch for N=1 is not included, although any
suggested improvements in that will be considered carefully.)

praxis.f includes test software to run any of the test problems
that Brent ran, and is set to run at least one case of each problem.
I have run these on an IBM 3090, essentially the same 
architecture that Brent used, and obtained essentially the same
results that Brent shows on pages 140-155.  The Hilbert problem with
N=12, for which Brent shows no termination results and for which
the results in Table 7.1 are correct but not relevant,
runs a long time; I cut it off at 3000 function evaluations.
I don't particularly like Brent's convergence criterion,
which allows this sort of extremely slow creeping progress,
but have not modified it.

Please notify me of any problems with this software,
or of any suggested modifications.

John Chandler
Computer Science Department
Oklahoma State University
Stillwater, Oklahoma 74078, U.S.A.
(405) 744-5676
jpc@a.cs.okstate.edu

