Joined: Oct. 2009
|Quote (midwifetoad @ Oct. 08 2012,09:47)|
|To my bemusement, the "winning" patterns are very simple and regular. As if you could have an oracle that sets "weasel" as the target without knowing the target string.|
What remains of the ID argument is the unevidenced assertion that that are no fitness gradients or plateaus. Gpuccio in particular, treats fitness as entirely one dimensional, and selection always striving for a particular optimum sequence.
He seems unable conceptually to grasp the idea that natural selection sees all aspects of fitness simultaneously.
As I recall, there was a machine learning system that learned how to play checkers. The programmers didn't tell it any rules or even what a winning condition was.
All they did was, after 5 runs, gave the machine a win/loss ratio. That's it.
The computer learned the rules of checkers and then learned optimal play. When the microsoft gaming servers were online, they played the machine against live opponents there and found that it was nearly master level... something like an ELO score around 1480 (with 1500 being master).
So it's possible.
Ignored by those who can't provide evidence for their claims.