[ExI] DeepMind wins Game1 in Go championship Match

Anders Sandberg anders at aleph.se
Thu Mar 10 09:15:38 UTC 2016

On 2016-03-09 20:39, spike wrote:
>  I am seeing a replay of what computers were doing in chess in the 
> 1990s. We were persistently astonished at how good those things were 
> getting. They didn't exhibit the usual human pattern of steady growth 
> with a fairly sudden (and persistent) leveling of ability at some point.

If you look at the growth of the maximal computer chess Elo scores in 
this era, it was essentially a straight line, starting way below the 
master human level - however, since it is the difference that determines 
the probability of winning, the linear increase implied a quick (about a 
decade) shift from humans nearly certain of winning to computers nearly 
certain of winning.

You can see plots on slide 28 of this presentation:

William's question about learning curves is partially answered by the 
previous slide, 27, which shows some plots from DeepMind's IMHO even 
more impressive Atari game player. Over time scores increase n a convex 
curve not too dissimilar to a human learning curve. The important issue 
here is rather the generalisation ability: the same system can learn 
many different games. However, what it is lacking is the ability to 
learn them all at the same time or transfer skills from one game to the 

On page 29 I have some plots from Katja Grace's excellent review, which 
points out that algorithmic improvements typically produce pretty 
drastic jumps in capability, unlike the individual learning curve or the 
gradual collective improvement in chess.

Anders Sandberg
Future of Humanity Institute
Oxford Martin School
Oxford University

More information about the extropy-chat mailing list