Argh - forget about my ramblings on the Step function - it is indeed
implemented correctly. I made the mistake of specifying a range
outside of -5.12, 5.12 - which works fine for the other EC functions,
but not that one. In trying to figure out what was wrong, I looked up
a definition of the Step function which was illustrated in a very
grainy image - and thus mistook the floor symbol in that image for an
Mea culpa! Sorry about that.
Still - would be great to get some feedback on the best GA
implementations to use.
On 25/01/2008, at 9:33 PM, Michael Hart wrote:
> Hi All,
> I'm really enjoying using ECJ - it's a wonderfully sophisticated
> First off - I think the Step function in the ECSuite is implemented
> incorrectly and leads to a min that's actually below 0 (ie, fitness
> becomes > 0) - it should be using Math.abs instead of Math.floor.
> Secondly, I was wondering whether there's a best practices guide
> when it comes to using GA with ECJ. The basic GA set is useful, but
> it looks like there's many different ways to implement it (Simple,
> Steady State, DE, Coevolution, Mu-Lambda, etc). Are any of these
> techniques on average better than others? If that question is a
> little too broad, then which of the aforementioned methods (or any
> of the others included in ECJ) performs best on the ECSuite set of
> functions? When I say "best" I guess I'm talking on-average best
> result for a given number of evolutions.
> Are there any papers, websites or any other documentation which
> compare all of these techniques?