Posted: 2/25/2007 6:16 AM
Post subject: Genetic algorithms for TASing
Quote
Ramzi
Joined: 3/25/2004
Posts: 459
I do not know much about this, aside from what I read on the Wikipedia article. So maybe some of you guys can tell me what you think. Suppose we take a game like Mario 1, and develop a bunch of random strategies like: run forward only, jump only, run and jump, do actions randomly. Then we define a cost function based upon a) the distance Mario makes it in the game, and b) the time it takes Mario to get there. Then we measure all of the strategies by our cost function, and select the most fit to go and make a mutated offspring. In enough generations Mario should be running forwards and not dying. I don't think it will be able to match what we humans have created, what with our knowledge of glitches and all, but maybe this technique can be used somewhere? Maybe?
Posted: 2/25/2007 6:30 AM
Quote
Bag_of_Magic_Food
He/Him
Active player (278)
Joined: 5/29/2004
Posts: 5712
That's actually been discussed before: http://tasvideos.org/forum/viewtopic.php?t=2885 (towards the end there)
put yourself in my rocketpack if that poochie is one outrageous dude
Posted: 2/26/2007 10:08 AM
Quote
qFox
Joined: 4/25/2004
Posts: 615
Location: The Netherlands
Also, there's more to it then just running. You need to jump at very precise times to actually make it run-worthy. GA's aren't that precise (although you could code in all the obstacles and... no I still believe it's very inefficient).
qfox.nl