Rybka vs. Hiarcs

Date: 15 Jun 2007 02:34:58
From: help bot
Subject: Rybka vs. Hiarcs

One of the top challengers to the top-rated engine, Rybka, is
Hiarcs.
So I go to the Web site for Hiarcs and a set of positions pops up,
which are supposed to show how great this program is relative to the
competition:

http://www.hiarcs.com/Games/H10pos.htm

The very first problem position is from a game between GMs Tal and
Uhlman, Moscow 1971. (If you recall, this was a time when the former
world champ was setting the record for games in-a-row without a loss.)
It is claimed that Hiarcs 10 finds the solution in less than one
second,
whereas other programs cannot find it even at tournament time
controls.

But when I set this up in Arena and, from a dead start and with no
hashtables, the free version of Rybka finds GM Tal's N-f5 at ply 14,
just
within tournament time controls. Oddly enough, it at first scores
this
move as just a teensy tad better than others. I didn't bother to let
this
run for a long time, since the boisterous claim had already been de-
bunked by a crippled (I don't know why it wasn't using any
hashtables)
old Rybka 1.0 engine, running in the background.

Some of the other problems on this site looked interesting, but my
computer is very sluggish right now and the claim included a long
list of engines I don't have.

-- help bot

Date: 15 Jun 2007 18:09:25
From: help bot
Subject: Re: Rybka vs. Hiarcs

On Jun 15, 4:59 pm, tin Brown <

Date: 15 Jun 2007 18:05:02
From: help bot
Subject: Re: Rybka vs. Hiarcs

On Jun 15, 12:28 pm, tin Brown <

Date:
From: Martin Brown
Subject: Re: Rybka vs. Hiarcs

Date: 15 Jun 2007 05:01:35
From: help bot
Subject: Re: Rybka vs. Hiarcs

Date: 15 Jun 2007 04:07:33
From: help bot
Subject: Re: Rybka vs. Hiarcs

On Jun 15, 6:20 am, David Richerby <[email protected] >
wrote:

> You haven't debunked anything. It's possible the uncrippled version
> of Rybka, by analyzing more positions, would see something that it
> mistakenly thought refuted Tal's move. Evaluations can go down as
> well as up. Consult a financial adviser before investing.
>
> Try it again with Rybka operating under reasonable conditions and see
> what it does. If it still thinks Nf5 is good, then you have a point.

I think you have missed the point. Not only was that
article written about an older version of Hiarcs, but it
obviously could not have tested these positions against
the latest, greatest version of Rybka either.

My point was simply that on my meager hardware,
even the old demo version of Rybka -- which is hardly
as strong as the current version -- found the move the
author claimed no other program could find "at
tournament time controls". Obviously, by simply
running the free demo Rybka on a speedy desktop,
his claim would easily be debunked.

My test started by entering the position from scratch,
whereas in a tournament, the hashtables would not
only be utilized (an improvement in itself), but pre-
loaded with relevant data to speed up the search a
bit. It goes without saying that they would also use
a much faster, desktop computer (not an old
notebook). In sum, had the Rybka free demo engine
and a modern desktop computer been facing GM
Uhlman, it too would have played GM Tal's N-f5!,
winning. This *by itself* refutes the author's claim.

Now, once you run through the moves Nf5 exf5,
Re1+ Be6, Qd6, even my crippled machine will
evaluate the position as a clearcut win.

But as for the idea that Rybka might think it sees
a refutation to GM Tal's move, you need to look at
the screen. When an engine "overlooks" or rather,
mis-evaluates a move, this can happen. But here,
what we have is a case where the deeper it looks,
the more it "realizes" that GM Tal's move is the best.
(Of course, in rare cases it could prefer such a move
for several plies and then revert back to its initial,
mistaken preference for a capture on c6.)

I wanted to check the claims as to the other
positions, to see if in fact "no other program" could
find the solutions given. I suspect that in some
cases, this is because there is an alternate or
superior solution. For example, in the position
from one of GM Alekhine's games it is not really
necessary to sac the Bishop in order to win, and
in the Rook and pawn ending it is anything but
clear that "g4!" is the only, or even the fastest win.

Despite the exaggerated claims, it is interesting
that Hiarcs can (allegedly) solve tactical problems
with lightning speed. I especially liked the last
problem, where ...B-d5 (capturable many different
ways) forces the issue. But to be fair, I wonder if
a purely random sample of chess problems would
give such results as the ones on that site. For
example, in discussing the Whitaker endgame it
seems that Rybka far out-shined other programs
tested by others -- and that is without using the
tablebases!

-- help bot

Date: 15 Jun 2007 11:20:45
From: David Richerby
Subject: Re: Rybka vs. Hiarcs

help bot <[email protected] > wrote:
> But when I set this up in Arena and, from a dead start and with no
> hashtables, the free version of Rybka finds GM Tal's N-f5 at ply 14,
> just within tournament time controls. Oddly enough, it at first
> scores this move as just a teensy tad better than others. I didn't
> bother to let this run for a long time, since the boisterous claim
> had already been de- bunked by a crippled (I don't know why it
> wasn't using any hashtables) old Rybka 1.0 engine, running in the
> background.

You haven't debunked anything. It's possible the uncrippled version
of Rybka, by analyzing more positions, would see something that it
mistakenly thought refuted Tal's move. Evaluations can go down as
well as up. Consult a financial adviser before investing.

Try it again with Rybka operating under reasonable conditions and see
what it does. If it still thinks Nf5 is good, then you have a point.

Dave.

--
David Richerby Gigantic Nuclear Beer (TM): it's like
www.chiark.greenend.org.uk/~davidr/ a refreshing lager that's made of
atoms but it's huge!