Main
Date: 05 Dec 2006 00:52:57
From: palmchessfan
Subject: Palm HIARCS set ELO levels - too strong ?
Hi,

I have a question related to the set ELO levels on Palm HIARCS. Has
anyone who has an established ELO level (either USCF or European)
played any of the set ELO levels on Palm HIARCS ? Do you feel they are
reasonably accurate ?

I suspect that even when setting the ELO levels to e.g. 1500 ELO
(around my level I estimate) it is playing a fair bit stronger than a
human rated (European) at roughly the same level. I remember someone
else (1 year ago?) posting similar comments a while back.
I really like the program a lot, great playing style, but the most
important aspect for me is setting the playing strength to my level (or
slightly above) and I feel that palm HIARCS has trouble constraining
its ELO levels. I would appreciate anyone else's comments and
experiance.

If k Uniacke (of PALM) is reading this - how did you estimate the
(set) ELO levels ? Thanks anyway for a great program, looking forward
to further upates !

(I have a palm T5 with HIARCS 9.6 installed).

Cheers,
k.





 
Date: 27 Dec 2006 12:44:09
From: Bernard Hill
Subject: Re: Palm HIARCS set ELO levels - too strong ?
In message <D1z*[email protected] >, David Richerby
<[email protected] > writes
>palmchessfan <[email protected]> wrote:
>> Note the MMV (although a fairly old model) has a solid 1874 SSDF
>> rating (i.e. a solid USCF Expert)
>
>No! An 1874 SSDF rating means no such thing! The SSDF ratings are
>generated solely from computer-vs-computer matches and tell you
>nothing about how the computer would perform against a human being.
>All you can tell from the SSDF rating list is that the MMV would lose
>the same proportion of games to a computer rated 2074 on that list as
>a 1874USCF human would lose to a 2074USCF human.

Bear in mind also that the USCF ratings are about 100 points higher than
the SSDF ratings. It's a country thing.


--
Bernard Hill
Selkirk, Scotland



 
Date: 05 Dec 2006 14:04:57
From: palmchessfan
Subject: Re: Palm HIARCS set ELO levels - too strong ?
A follow up to my previous message out of curiosity, I have just played
Palm HIARCS v9.6(playing black) set to 1450 ELO against a Mephisto MM V
averaging 60 - 90 secs per move. The result was a draw (~80 moves) !!
Judging by the game, both sides appeared very even throughout (a good
game).

Although one game in itself is not conclusive, and the MMV was at
approx. half speed of its valid rating, this does seem to indicate a
big problem with Palm HIARCS set ELO levels, at least in this range. I
am convinced that Human ELO 1450 players don't play this good !

Note the MMV (although a fairly old model) has a solid 1874 SSDF rating
(i.e. a solid USCF Expert), and plenty of past good games against
Expert rated human players.

To me, and I assume all none-master rated players (i.e. most people
here), the ability to lower the rating of programs, while maintaining
quality play, is easily the most useful feature (together with
analysing). This issue has been raised in the past concerning Palm
HIARCS on an older thread, and the problem still appears unfixed.

Can the HIARCS team please investigate ?

The nodes per second (nps) throughout the game was between 10000 and
100000. Surely this should be lower at ELO 1450 ??
I have a PGN of the game, which contains the following metrics:
Plycount =162, Time Control = 3600, H9.5: 2048Kb, %lev 5.4, %hopt
538573967, %rmv 163, %tns 0,29535, %ctv 0. A first user time control
(probably ignored) was set to 60 mins.
Note that the H9.5 figure looks a bit suspicious ?

Regards,
k



  
Date: 06 Dec 2006 09:50:37
From: David Richerby
Subject: Re: Palm HIARCS set ELO levels - too strong ?
palmchessfan <[email protected] > wrote:
> Note the MMV (although a fairly old model) has a solid 1874 SSDF
> rating (i.e. a solid USCF Expert)

No! An 1874 SSDF rating means no such thing! The SSDF ratings are
generated solely from computer-vs-computer matches and tell you
nothing about how the computer would perform against a human being.
All you can tell from the SSDF rating list is that the MMV would lose
the same proportion of games to a computer rated 2074 on that list as
a 1874USCF human would lose to a 2074USCF human.


> and plenty of past good games against Expert rated human players.

That, on the other hand, tells you a lot about the strength of this
computer.


> The nodes per second (nps) throughout the game was between 10000 and
> 100000. Surely this should be lower at ELO 1450 ??

Not necessarily. Perhaps HIARCS dumbs itself down by weakening the
evaluation function rather than by slowing the search.


Dave.

--
David Richerby Mouldy Old-Fashioned Dictator (TM):
www.chiark.greenend.org.uk/~davidr/ it's like a totalitarian leader but
it's perfect for your grandparents
and starting to grow mushrooms!