Jorge Tavares weblog

In my previous post, I wrote an overview of this simple, but important, component which is the generation of pseudo-random numbers, using Common Lisp. However, sometimes you need a little bit more. Naturally, an implementation only offers one generator and no standard facilities exist to change. If you want a different generator the only options are to do it yourself or to use one from a library. Depending on one’s person needs, implementing our own generator might be or not an option. And using a library too. First, there are not many libraries containing different generators (at least, to the best of my knowledge). We can find code of other generators in Lisp but that’s very different from having a good, portable and complete library with everything (or most) we might need. As such, as far as I know, the best option is to use GSLL which allows us to use in Lisp the many things GSL has. GSL provides a unique set of pseudo-random numbers generators with the big advantage that they can be used later with random number distributions.

How can we use this facility that GSL/GSLL gives us? Well, first we need to install both libraries, as said before. As soon as everything is up and running, working with the library is not difficult. So far, I am not aware of a manual/tutorial for the Lisp side of GSL but as the website advises, we first look at the GSL documentation the name of the function that we want, then we use both gsl-lookup and documentation functions to learn about the Lisp equivalent. Looking at the examples might be also very useful. However, here I will focus just on the Lisp functions.

The first thing we need is to create an object that contains our pseudo-random generator. The object is named random-number-generator and we can create it using make-random-number-generator. This function requires an argument indicating the type of the generator. This is one of the big pluses of using GSLL. We can create several objects with different generators and the list provided by GSL is large (there is a function all-random-number-generators but it only returns the pointers, not the Lisp constant symbols). But for the moment, let’s focus on the essentials and assume that we also want a Mersenne Twister generator. This type of generator is represented by the constant +mt19937+. For example, if we want to define *rng* as a global variable that contains our generator from GSLL we only need to do this :

GSL> (defparameter *rng* (make-random-number-generator +mt19937+))
*RNG*

The variable *rng* now holds our generator. Anyway, this just creates our generator with the default seed. The default seed is contained in the constant +default-seed+ and defaults to the value of zero. In the previous example, the call might have been written like (make-random-number-generator +mt19937+ +default-seed+). If we want to seed the generator with a different seed value, we only need to add the seed, e.g., 1024, since it is an optional argument:

GSL> (defparameter *rng* (make-random-number-generator +mt19937+ 1024))
*RNG*

Contrary to the standard of Common Lisp, this allows to seed the generators with known seeds in the same way as in other languages. This could be useful if you want to compare and reproduce results with programs coded in languages other than Lisp. Having said that, we need to keep in mind the randomness issue of a user-specified seed.

We can now create pseudo-random generators but we don’t know how to produce the numbers. The basic function is get-random-number which receives as an argument the generator to be used and returns an integer. Although simple to use, the function returns integers from the interval [min,max] which is generator dependent. If we don’t know these bounds we can use the auxiliary functions rng-min and rng-max to obtain them. Using the MT generator with the default seed:

GSL> (rng-min *rng*)
0
GSL> (rng-max *rng*)
4294967295
GSL> (get-random-number *rng*)
4293858116
GSL> (get-random-number *rng*)
699692587
GSL> (get-random-number *rng*)
1213834231
GSL> (get-random-number *rng*)
4068197670

We can’t specify an upper bound like in the standard random function but it’s easy to make a function using get-random-number where we can set the limits from which we want pseudo-random numbers. Lucky for us, there is another way to produce our numbers by using the function sample. The major difference from get-random-number is that sample can produce floating point numbers and integers. In the first case, sample can generate numbers between 0.0 and 1.0 (exclusive). There is also the possibility to include or not 0.0. In the second case, sample generates integers between 0 and an upper bound given by a parameter (:upperbound). With this in mind, to decide how we want sample to behave the control flag must be one of the following: uniform, uniform>0 and uniform-fixnum. Examples:

Generating numbers between [0.0, 1.0):

GSL> (sample *rng* 'uniform)
0.23165654274635017d0
GSL> (sample *rng* 'uniform)
0.48497361433692276d0

Generating numbers between (0.0, 1.0):

GSL> (sample *rng* 'uniform>0)
0.9574769565369934d0
GSL> (sample *rng* 'uniform>0)
0.7443053431343287d0

Generating integers between 0 (inclusive ) and 100 (exclusive):

GSL> (sample *rng* 'uniform-fixnum :upperbound 100)
54
GSL> (sample *rng* 'uniform-fixnum :upperbound 100)
73

These are the main operations with random number generators using GSLL. As seen, the main advantage is to be able to create different types of generators and having an easier way to control the seed. There are other functions that operate on generators but less interesting. We have two functions that copy the state of a given generator: copy-to-destination and copy-making-destination. In short, these functions make an exact copy of a generator. The first function copies a generator into a given destination with the same type of the source. The second function does not need a destination. Examples:

GSL> (defparameter *rng2* (make-random-number-generator +mt19937+ 1024))
*RNG2*
GSL> (copy-to-destination *rng2* *rng*)
; No value
GSL> (get-random-number *rng*)
2781812667
GSL> (get-random-number *rng2*)
2781812667
GSL> (defparameter *rng3* (copy-making-destination *rng*))
*RNG3*
GSL> (get-random-number *rng2*)
1825252241
GSL> (get-random-number *rng3*)
1825252241

The other remaining functions are name, rng-state, size and get-random-state. The first function returns a string with the name of the generator, rng-state a pointer to the generator state, size the size of the pointer and get-random-state return the complete state of a given generator specified as a vector of bytes. And this is it!

Before concluding this overview of using pseudo-random numbers generators in Lisp using GSLL, just a few thoughts. Using this library seems a good option especially if generating random numbers is important in our programs. The reasons are evident: several generators to use, seed control, and library maturity (GSL). I have no idea of how GSLL affects performance but that is something to be explored and studied.

Recently I’ve just clean up my computer which in short means a format and a new software installation. I also try to clean up my files and keep things in order but that part is harder to accomplish. And as a side effect I guess my MacBookPro is more prepared for the Snow Leopard upgrade (still need to purchase the box). Anyway, as part of this process and my research work, I’ve decided to explore more some options in terms of libraries and related stuff. The idea is not only to use full libraries/frameworks related to Evolutionary Computing but mainly, to find good scientific tools that might help me in developing the things I know that don exist. And so, the number one library I decided to take a look was the GNU Scientific Library (GSL).

GSL has a lot of stuff, some of which I will probably never need, but a quick look at the documentation revealed something that interest me very much: random number generators. Since I develop on Mac installing the library was pretty straightforward using MacPorts: port install gsl and it was done! A quick test showed that it was ok. Using the first example from GSL documentation:

#include <stdio.h
#include <gsl/gsl_sf_bessel.h>
     
int main (void)
{
    double x = 5.0;
    double y = gsl_sf_bessel_J0 (x);
    printf ("J0(%g) = %.18e\n", x, y);
    return 0;
}

Compiling with:

gcc example.c -o example -lgsl -I/opt/local/include -L/opt/local/lib 

and the output was J0(5) = -1.775967713143382642e-01. Since GSL is coded in pure C it’s very easy to include it in Objective-C programs if thats the case.

However, since I am a Lisp guy I needed to find if bindings to Common Lisp were available. And since nowadays you can find almost all you want for Lisp, it was not hard to stumble into GSLL (The GNU Scientific Library for Lisp). The library is very complete and allows an easy manipulation of the library functions, even in a interactive manner.

Installing GSSL on the Mac was a little bit more complicated than expected but thanks to the help of Liam Healy, it is now running fine. The main problem was due to the fact that the headers file of libraries on Mac OS are not always on standard places as other unixes systems and so, things might not run at the first time. And in my case the headers were in a different place than what it was assumed. But now if you install GSLL by using clbuild, most likely there won be problems. In essence, it’s just necessary to follow the instructions on the GSLL website in how to install it using clbuild and then:

CL-USER> (asdf:operate 'asdf:load-op :gsll)

To see a list of examples:

CL-USER> (gsll:examples)

And since I am interested in random number generators:

CL-USER> (gsll:examples 'random-number-generator)
((LET ((RNG (MAKE-RANDOM-NUMBER-GENERATOR +MT19937+ 0)))
   (LOOP FOR I FROM 0 TO 10
         COLLECT (SAMPLE RNG 'UNIFORM-FIXNUM :UPPERBOUND 1000)))
 (LET ((RNG (MAKE-RANDOM-NUMBER-GENERATOR *CMRG* 0)))
   (LOOP FOR I FROM 0 TO 10
         COLLECT (SAMPLE RNG 'UNIFORM))))

A few days ago I’ve started to code my own EC library in Lisp and this will be very useful since I will be able to use and control the random number generation process. To be able to use the Mersenne Twister generator is a big plus for me! Anyway, the exploration of GSL and GSLL has just begun!