IMO, the only thing wrong with the regexp approach used in some responses was forgetting to include the optional \.*\d* in the number being read in for sorting (I also made some style and performance adjustments to the sort algorithm, i.e. the use of a subroutine and an orcish manoeuvre)

perl -w -MData::Dumper -e '
@old = ( "10.5 AA", "9 AC", "2 BB");
%orcish = ();
@new = sort {($orcish{$b} ||= ForceFloat($b)) <=> ($orcish{$a} ||= For
+ceFloat($a))} @old; 
print Dumper \@new;
sub ForceFloat{
    return $_[0] =~ /(\d+\.*\d*)/ ? $1 : 0.0;
}'
[download]