Re: Question about version 5.05.

[Simon Waters]

"Sternbach, William [IT]" wrote:
> 
> Thank you for your quick reply.
> 
> 1) 5.04 may play a better game of chess than 5.05
>     because 5.04 builds a bigger opening book with more depth.
>     Why does 5.05 reduce the book size and depth level.
>     As PC's are getting faster and faster, shouldn't we be increasing
>     the book.dat size and increasing the depth?

Because it was modified to behave as documented using depth in
ply not moves.

If people want the deeper book they can rebuild, we'll probably
make it runtime configurable, when the dreaded ".chessrc" file
appears.

I don't think the change in book depth has much impact on chess
strength, the book, and book selection algorithmns aren't good
enough for that yet. The bigger problem seems to be cluelessness
on leaving book, this should be better when it can ponder.

My PC hasn't got any faster since I started work on GNU Chess 5
;-) My PC does in 6 hours what Deeper Blue did in one second :-)
Deep Fritz would take about 4 minutes to analyse the same number
of positions from what I can gather.

Actually it is kind of useful as I can compare benchmarks
easily, I might save this one for just that when I move my
desktop to something faster.

> 2) In your installation instructions, perhaps you may want to recommend
>     that everyone use gcc -O3 for maximum optimization so their
>     program will decide on its moves as quickly as possible.
>     This can obviously be done by typing the following line:
>     CFLAGS="-O3 -foptimize-sibling-calls" ./configure

We already suggest trying compiler options "-O3", thanks to
Lukas I believe, see INSTALL.

'optimize-sibling-calls' is in my list of things to check.

GCC v3 was the last big jump in compiler performance, but these
optimisations make surprisingly little difference to playing
strength. Shaving 10 or 15% of run time will give you maybe 10
ELO points, this is hardly noticable except maybe at lightning
chess. However if someone has the resources to test different
options across a lot of architectures, we and the gcc team will
be interested in the results I'm sure.


The single biggest headache in performance and skill, comes from
the hashtable size. However for next release we will stop
clearing this as often, this will free us to use a larger
default table as clearing it will be less problematic for
blitz/lightning games.

These changes will also introduce pondering, so we need a bigger
hashtable to usefully utilise the output of pondering. Even at
at modest time controls on modern PC's you may end up wanting a
serious sized hashtable to get best performance, something like
30,000,000 slots as a minimum. I don't think everyone in the PC
world is quite ready for that yet - I don't have a decent test
PC for those kind of hashtable sizes. 32 bits architectures are
on the way out!

Anyone got a cross platform how much "real" memory can I grab
routine? Alternatively we could drop a "nag" message into the
engine interface, so it nags if the hashsize is still the
(small) default.

Chess playing strength isn't the number one request from users.
The number one question is "can I make it easier to beat", and
the leading feature request seems to be "can we have
personalities" (which I think may be related to making it easier
to beat!).

However we aren't Microsoft, and we can ruthlessly ignore user
requests if we like. Besides users can play it at assymetric
time controls and select different options if they want to beat
it.

As a more serious chess hobbyist and free software promoter I
want to make it provide adequate feature to satisfy users of the
common free software interfaces i.e. Winboard/Xboard/SCID. So
"analyze" is high on the target list! We have gained "hint" and
"book" in recent versions.

 Simon