Okay i''m not pretending i know a lot about Genetic Programming, hopefully the knowledge will come in time but that won''t stop me from brainstorming about ideas 8^) Kirk, this one''s for you too (i think it''s good to share ideas so i post this one on the forum) and thanks for the lightningspeed reply! Okay on with the idea! Short said it''s a breeding pit for evolving species. I want to randomly generate a universe in 3D. After that, each planet has it''s own properties and base materials. Then try to create some sort of borg style species. Species with different base material needs... what happens if you put a really valuable material in your universe what''s like gold for humans? Ultimately i want to see different species doing different things; eg trading, aggressive, defending everything tactic/life style that will survive the species in the environment. I''m targeting species intercommunication, eg. one species will attack alot thus another species will ''camouflage'' itself. I hope i presented my idea clear but i''m a little tired and then it''s really difficult if you''re not native english Anyway let''s talk about this idea! // Redmar. "Genetic Programmers started skynet! 8^)"

Study A-life, GA''s, NN and FSM. What I think you''re trying to create is evolving a sort of intelligent decision making. If you really want to tackle this problem, I suggest you test your theory on something simpler first, like 2D tanks battling. This is easier to simulate. Evolving these tanks, you will have some always hiding, some always shooting and some always running.

But if I didn''t get your question, could you be a little more specific?

Edotorpedo

Howdy.

I agree with edtorpedo that you might want to start with something simpler. Not that your idea doesn't have merit, but it would definitely be a tough problem to tackle.

The first thing to consider is how you will model your agents' behavior. To use the old Pac-Man and ghosts example (I know I beat this one to death) you could represent the ghosts directional choices as a decision tree or if-then block.


if PACMAN is EAST then
if MYCURRENT HEADING is NORTH then
TURN RIGHT
else if MYCURRENT HEADING is SOUTH then
TURN LEFT



I realize this is an OVERsimplification, but it gets the point across. Actually, edtorpedo's suggestion of FSM representation is a great idea. You could model an extremely complex behavior in a reasonable space and the actual decision process would be very fast.

After you have your representation you'll need to come up with appropriate genetic operators for crossover and mutation, and consider some type of validity enforcement so that you don't generate invalid decision structures.

As you can probably see from this short list, the problem you suggested would require a huge amount of effort to accomodate all the possible variables. A very noble effort.

Regarding the note by edtorpedo about tank controls, there is a demo out there somewhere which evolved the behavior of tanks in an open environment. The fitness was length of survival, and I forget the details of the controls - essentially I believe it involved controlling the direction of the left and right track and the turrent firing. I ran it a number of times and it always had the same result: The first generation was essentially random movement and firing. After a couple of generations the random movers gave way to those which moved in a straight line and fired thier turret guns constantly. The final winner in every case that I ran (probably 50) was a tank which stayed in its original location, spun in a circle very rapidly, and fired its turret gun at the highest frequency. Essentially it wiped out any body else immediately. Interesting, eh?

-Kirk



[edited by - KirkD on August 2, 2002 9:43:16 AM]

[edited by - KirkD on August 2, 2002 9:43:43 AM]