RE: Hardware Progress: $272/GFlop/s

From: Alejandro Dubrovsky (
Date: Tue Aug 27 2002 - 10:07:09 MDT

On Wed, 2002-08-28 at 00:57, Ben Goertzel wrote:
> I'm not so sure there isn't a clear relationship between processor speed and
> narrow-AI efficiency, however.
> AI chess-playing is an example. I think the quality of chess programs'
> gameplay has pretty much increased right along with processor speed.
> Certainly the algorithms involved in these programs (alpha-beta pruning with
> special evaluation heuristics) are simple and are able to make direct use of
> increased processor speed.

I think the increase in performance in chess programs is estimated to be
50 ELO points per processor speed doubling, so, yes, for now, it's a
direct improvement. Improvement of this kind has to drop off at some
point, since there is a diminishing rate of return for search depth.
Note, though, that algorithmic improvements tend to come at least as
quick as hardware. (As a side note, i think you are slightly
oversimplifying the algorithms involved. Noone really uses "standard"
alpha-beta any more, but derivations of it (NegaScout, MTD(f)).
Selective search extensions play a very important part (heuristics for
when to extend can be tricky). Move ordering is also very important.
Forward pruning techniques are also used by pretty much every program.
Also add transposition table replacement heuristics, time management
heuristics, book learning, positional lerning. Still relatively simple,
but alpha-beta plus evaluation heuristics at 20 GHz won't compete with
Fritz at 2GHz)

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT