RE: neural Network/optimization/perl (fwd)

[heberf]

Here is the simulated annealing code.  It's a dynamically loaded octave
function (a .oct file).  The C++ code is attached.  It's a bit messy
because he wrote it for a specific purpose but I think it would be pretty
easy to clean up and make completely general.  Below I've attached his
comments about it.  More than one person has expressed interest and
asked to be able to change it.  I'll be happy to be the focal point of
development efforts on this as far as it is needed.


---------- Forwarded message ----------
Date: Wed, 6 Oct 1999 15:12:19 -0500 
From: Paskalis Glabadanidis <address@hidden>
To: "'address@hidden'" <address@hidden>
Subject: RE: neural Network/optimization/perl (fwd)

Sure - I've attached the file SA2.CC below. Unfortunately, my comments are
rather sketchy so
a few references and explanations are in order.

First, this algorithm is based on a freeware GAUSS code written by Tsionas
which I optimized for efficiency (can't find the site but it can be found by
a Web search - that's how I found it) which in turn was based on a FORTRAN
code by Bill Goffe (also available somewhere on the Web). Both Tsionas's and
Goffe's code follow closely the algorithm outlined in Corana, Marchesi,
Martini and Ridella, ACM Transactions on Mathematical Software (1987), 13,
262-280 and Goffe, Ferrier and Rogers, Journal of Econometrics (1994), 60,
65-99.

Second, a few words about my code. It is tailored to a particular objective
function which required some data as well as parameters in order to be
evaluated; so feel free to delete the extra variables Z1, Z2, UNVAR1, UNVAR2
and UNCOV12 (DUMMY is also redundant)
The absolutely necessary arguments of SA2 for it to work are the following:
        X       parameter N-vector (N is X's dimension)
        LB      an N-vector of lower bounds on X
        UB      an N-vector of upper bounds on X
        C       an N-vector which controls step length adujstments (see
articles)
                CMMR's suggested value is 2*ones(N,1)
        VM      an N-vector of initial step lengths (usually equal to at
least UB-LB)
        RT      controls the speed of temperature reduction (suggested value
0.85)
        EPS     convergence parameter (suggested value 1e-8 to 1e-6)
        T       initial temperature (see articles for suggested values)
        NS      suggested value NS=20
        NT      suggested value NT=max(100,5*N) (CMMR) although GFR
suggested lower values
        NEPS    convergence parameter (suggested values NEPS=4 to 10)
        MAXEVL  limit on function evaluations
        FUNC_NAME       a string for the objective function name; I use the
octave value list ARG to pass X as well as additional parameters if
necessary to the FUNC_NAME function.
The use of the parameters MAX and IPRINT are commented out to increase
efficiency and it is straightforward to uncomment them. The default
objective of SA2.CC is to maximize FUNC_NAME (MAX=1).

Hope this is sufficient but feel free to shoot me an email if you have
further questions.
Paskalis

-----Original Message-----
From: address@hidden [mailto:address@hidden
Sent: Tuesday, October 05, 1999 10:27 PM
To: Paskalis Glabadanidis
Subject: Re: neural Network/optimization/perl (fwd)


I mentioned your simulated annealing code on the Octave list and now lots
of people want it.  If you send me the files I'll give them to those who
want them.  A few instructions would be handy as well.

____________________________________________________________________________
___
Heber Farnsworth
Assistant Professor of Finance
John M. Olin School of Business
Washington University
Campus Box 1133                                 phone: (314) 935-4221   
One Brookings Drive                             FAX: (314) 935-6359
St. Louis, MO 63130-4899                


---------- Forwarded message ----------
Date: Wed, 6 Oct 1999 09:51:46 +1000 (EST)
From: Pierre Scotney <address@hidden>
To: address@hidden
Subject: Re: neural Network/optimization/perl

On Tue, 5 Oct 1999 address@hidden wrote:

> A PhD student here has written some octave code that does simulated
> annealing if anyone is interested.

I am interested, is there a web site?

Regards

Pierre

-- 
Dr Pierre Scotney
St. Vincent's Institute of Medical Research
Melbourne, Australia
http://foo.medstv.unimelb.edu.au/~pierre

sa2.cc
Description: Binary data