"raylopez99" wrote in message
oups.com...
On Apr 26, 6:24 pm, "David Kane" wrote:
"raylopez99" wrote in message

oups.com...

So, understanding how chess works, and how chess playing computers
work, and having seen Crafty evaluate pretty good myself, I have to
side with the original article.

I would not go so far as to say that I side with the original argument,
only that Riis' objections were groundless.

I agree Riis' objections were groundless.

In fact, the original authors
have done some groundbreaking work on developing a
methodology to rate chess players. It is, at the very least,
very interesting, and a refreshing change from the pseudo-science
historical ELO/chessmetrics stuff.

If you have a moment kindly articulate what you mean by pseudo-
science. I like Sonas' work, which has been used by the PGA (defunct
GM association). What I am aware of is that (from memory) it seems
Sonas rates players from different rating pools (which Sonas
acknowledges is faulty), that is to say from different time periods,
when in fact as Arpad Elo pointed out, as time progresses the rating
pool as a whole gets stronger. Also Sonas apparently (from reading
some of these threads) makes assumptions such as a player who plays
infrequently should gain or lose a different number of Elo points than
a player who plays regularly. What other 'pseudo-science' (which I
take to be what you feel are defects) are you referring to?

ELOs methodology was designed for a certain set of conditions
(playing actively in a pool) but is applied under conditions way outside
of those conditions. The issue is whether those extrapolations
are supported by any evidence.


The problem with the work is
that it applies a new method to a very hard problem (ranking
world champions) when they haven't even shown the method's
worth when applied to easy problems (ranking everybody else).


Ranking everybody else meaning what? If the players compete regularly
against one another, the Elo Gaussian distribution seems a good way of
ranking to me.

That is a big if. Look at actual ratings lists and you will find: 1. That they
contain a very small number of players compared to the
total number of chess players 2. Many of the players in the list have
very few games 3. Many are based on old results.

The proposed method should be confirmed by applying
it to problems where we know what the answer should be. I.e.
first determine a general relationship between moves and wins,
and *then* apply it to a hard problem (like ranking world
champions).

BTW I've seen Arpad Elo's scheme even applied to rank
world football (soccer) teams, and it's surprising how well the system
seems to work (Brazil was #1, as expected, and the other familiar
winners were in the top 10, even Greece, which won the Euro
Championships in 2004 and were considered a 'surprise team', in fact
was ranked at the time in the top 10 by this system, so the 2004
championship wasn't quite that big an upset, not unlike Euwe's victory
over Alekhine was not that big an upset because in fact Euwe was quite
a good player, albeit obscure).

I'm not familiar with the soccer ratings which you speak of, but many ratings
of this form consider factors *other* than the result. Famously,
statisticians have analyzed in great detail what wins games in major
league baseball. While certainly results tell us something, it is absurd to
think
that W/L results are the *only* things with predictive value.


I have previously expressed belief in the theory that "move rating" will
eventually surpass "result rating" as the gold standard measurement of
chess skill. This is a small first step, but there is much work left to
do.

Agreed, though I doubt move rating will eventually surpass 'result
rating' anytime soon, unless we develop a very powerful PC (which is
possible) and/or a quantum computer that can instantly solve the
entire chess tree to give a 'perfect' verdict on every move (i.e., the
move, with perfect play will end in victory, defeat or a draw).

This is not at all necessary. Most people conveniently overlook
that, in practice, the conditions required for ELO rating are rarely
met. Move rating becomes useful when its errors are smaller than
those already present in result rating. The key advantage of move
rating is that you get much more information per game, so in theory
get better ratings faster.