The basic situation is this: in my fitness function, I use the genes of the AI::Genetic::Individual object to sort another list. On the second and subsequent times through my fitness function, it appears that the magic variables $a and $b inside the block passed to sort are being set to undef.

My current wild guess is that perl has problems when a sort block inside one package calls something that results in a sort block being called in a different package, but I'm not at all sure about that. In any case, I'm seeing this both in my cygwin 5.8.5 perl and my 5.6.1 ActiveState perl. I've simplified the code down as far as I can; it's behind the readmore tag:
Update: It's not AI::Genetic, at least not directly. It appears that I can duplicate this without depending on any external modules (see reply)

#!/usr/bin/perl
use strict;
use warnings;
use AI::Genetic;
use Data::Dumper;

$|=1;
my $ga = AI::Genetic->new
  (
   -fitness => \&my_fitness,
   -type => 'rangevector'
  );

$ga->init([ [0,10], [0,10], [0,10], [0,10] ]);
$ga->sortIndividuals($ga->people());
# Note that using this instead of sortPopulation does NOT
# cause the same warnings:
# my @populace = sort {$a->score <=> $b->score} @{$ga->people()};

sub my_fitness {
  my $genes = shift;

  my @order = (0,1,2,3);
  my @sorted = sort {
    local $SIG{__WARN__} = sub {print Dumper($a,$b); print @_; sleep 1
+;};
    $genes->[$a] <=> $genes->[$b]
  } @order;

  my $score = $sorted[1] + 2*$sorted[2] + 3*$sorted[3];

  print "DEBUG: score was $score\n";
  return $score;
}
[download]

--
@/=map{[/./g]}qw/.h_nJ Xapou cets krht ele_ r_ra/;
map{y/X_/\n /;print}map{pop@$_}@/for@/
[download]