jacobwilliams / pikaia

Modern Fortran Edition of the Pikaia Genetic Algorithm

Pikaia: Modern Fortran Edition of the PIKAIA Genetic Algorithm

Status

Overview

This is a refactoring of the PIKAIA unconstrained optimization code from the High Altitude Observatory. The original code is public domain and was written by Paul Charbonneau & Barry Knapp. The new code differs from the old code in the following respects:

• The original fixed-form source (FORTRAN 77) was converted to free-form source.
• The code is now object-oriented Fortran 2003/2008. All user interaction is now through the pikaia_class.
• All real variables are now double precision.
• The original random number generator was replaced with MT19937-64 (64-bit Mersenne Twister).
• There are various new options (e.g., a convergence window with a tolerance can be specified as a stopping condition, and the user can specify a subroutine for reporting iterations).
• Mapping the variables to be between 0 and 1 now occurs internally, rather than requiring the user to do it.
• Can now include an initial guess in the initial population.
• Some OpenMP support has been added.

Compiling

The library can be built with the Fortran Package Manager using the provided fpm.toml file like so:

fpm build --release

To use Pikaia within your fpm project, add the following to your fpm.toml file:

[dependencies]
pikaia = { git="https://github.com/jacobwilliams/pikaia.git" }

Or to use a specific version:

[dependencies]
pikaia = { git="https://github.com/jacobwilliams/pikaia.git", tag = "1.0.1" }

The mersenne-twister-fortran library is required to compile Pikaia. FPM will automatically download this library.

Examples

• An example use of Pikaia can be found here.

Documentation

• The API documentation for the current master branch can be found here. This is generated by processing the source files with FORD. Note that the shell script will also generate these files automatically in the doc folder, assuming you have FORD installed.
• The original Pikaia documentation (for v1.2) can be found here.

See also

• PIKAIA description page: http://www.hao.ucar.edu/modeling/pikaia/pikaia.php
• Original source code: http://download.hao.ucar.edu/archive/pikaia/

pikaia's Issues

Use mersenne-twister package

The mersenne-twister module has been split off into a separate FPM package: https://github.com/jacobwilliams/mersenne-twister-fortran

Make that an FPM dependency.

pikaia with openmp?

Dear Jacob,
many thanks for the updated pikaia code.

I am just playing a bit around with it and trying to start it parallel with openmp by adding
some cpu consuming dummy code in the Rosenbrock function (the plan is to apply pikaia to some "real world problem") and

a) in pikaya subroutine after "!Compute initial (random but bounded) phenotypes" comment line

      !$omp parallel do private(ip)
      do ip=istart,me%np
         do k=1,me%n
            oldph(k,ip)=urand()  !from [0,1] ! urand not thread safe!
         end do
         call me%ff(oldph(:,ip),fitns(ip))
      end do
      !$omp end parallel do

(have also set omp_set_nested(.false.) in the example to avoid splitting k loop... )

b) in newpop after "!replace population" comment line

      !$omp parallel do private(i)
      do i=1,me%np

         oldph(:,i)=newph(:,i)

         !get fitness using caller's fitness function
         call me%ff(oldph(:,i),fitns(i))

      end do
      !$omp end parallel do

It seems to bring something (not much but anyway) but I have got also 3 more questions related to the problem:

1. Are there any other good places to do openmp threads in the code? (mb I haven't found them?)
2. Should Fortran built-in random generator and urand() function be parts of private objects for each openmp thread? I.e. do we need to replace this random_number with something more portable and related to OOP ?
3. I cannot ensure reproducibility of calculation in modified version using openmp threads. This seems to be connected with 2? Or there is some kind of limitation at the algorithm level?

Issue with the tests

I just clone and tested this project using fpm (version 0.4.0), as follows:

fpm build
fpm test

and got the following error:

At line 809 of file ././src/pikaia_module.f90
Fortran runtime error: Index '0' of dimension 1 of array 'stackl' below lower bound of 1

Error termination. Backtrace:
#0  0x7f6f65ab57c2 in ???
#1  0x7f6f65ab6289 in ???
#2  0x7f6f65ab689a in ???
#3  0x40a854 in rqsort
	at ././src/pikaia_module.f90:809
#4  0x406665 in __pikaia_module_MOD_rnkpop
	at ././src/pikaia_module.f90:1243
#5  0x40d6a4 in __pikaia_module_MOD_pikaia
	at ././src/pikaia_module.f90:515
#6  0x40bbfe in __pikaia_module_MOD_solve_with_pikaia
	at ././src/pikaia_module.f90:650
#7  0x401a28 in pikaia_test
	at src/tests/pikaia_test.f90:75
#8  0x403836 in main
	at src/tests/pikaia_test.f90:10
<ERROR> Execution failed for object " pikaia_test "
<ERROR>*cmd_run*:stopping due to failed executions
STOP 1

This happens in the subroutine rqsort, and I am not sure how to solve it.

use the Binary/Gray encoding methods replace Decimal

https://github.com/weixing1531/pikaia/tree/Binary
I wrote a subclass named pikaia_binary_class.
In the class, I change the definition of array gn.
integer(IB),dimension(me%n),intent(out) :: gn
operate the array gn in subroutine cross,mutate,encode,decode with using fortran's BIT INTRINSIC PROCEDURES ,then save memory and increase speed.