[ExI] An AI program that teaches itself

[John Clark]

Google reports in the current issue
​of​
 the journal Nature that it has a new greatly improved Go program called
 "AlphaGo Zero" that is now the most powerful GO program in the world. And
​the program isn't good because of
 brute force, it needs to make less than one tenth as many calculations as
the previous best GO program "AlphaGo" that defeated the world's top human
GO player in 2015 4 games out of 5
​;​
and yet AlphaGo Zero just defeated AlphaGo in a 100 game tournament 100
games to zero.

Even more interesting is how AlphaGo Zero got so smart. The older program
AlphaGo had to start by analyzing hundreds of thousands of championship
level games made by human players, but AlphaGo Zero started with nothing
but the simple rules of GO and instructions to learn to
​g​
et better. At first the program was terrible but day by day it got better
and after 40 days of thinking about the problem
​became​
the
​best
at it in the world. But of course after 40 days of constant self
modification no human being can say how  AlphaGo Zero works.

https://www.nature.com/nature/journal/v550/n7676/full/nature24270.html

It seems to me the next logical step would be to switch the program's
interest from getting better at the game of GO to improving computer code,
including its own. I wonder where that could lead.

 John K Clark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20171020/bb83c842/attachment.html>