[computer-go] Re: Chess Programmer at the Olympiad

["chrilly" <chrilly@xxxxxxxx>]

 Dear Dave,
>
>Don't think that go programmers don't understand the state of the art in
>search algorithms.  You can
>take a look at arimaa.com and see that my arimaa program is much stronger
>than others written
>by chess programers.  This program uses alpha-beta PVS search with a big
>hash table, null move,
>selective search extensions, killer and history heuristic for move sorting,
>etc. and searches about 300K NPS.
>
I have since a few months some interest in computer-Go. My impression is:
The computer-Go and computer-Chess are relative seperated communities. There
is certainly a different culture in both communities. E.g. this forum has a
very positive, friendly atmosphere. One can not say this about similar
computer-chess forums.
Like in all other seperated cultures, there are probably misconceptions
about the other culture. The misconception of the chess-community is
probably, that Go-programms do not search (at least not in the chess sense).
But the difference is indeed not so clear cut. Brutus searches typically 18
Plies in the middle-game. But this does not mean at all, that this is an
18-Plies uniform search. The range of the actual variations length is 7 to
60 Plies.

The misconception of the Go community is certainly, that chess programs have
a simple evaluation function. The Top-4 programms Brutus, Fritz, Junior,
Shredder have a very complex evaluation function. They play a very dynamic
chess. With other words. There are a lot of mutual threats on the board. The
consequences of these threats can not be handled totally by the search. One
can not play this style of play without an evaluation functions which has at
least some idea about the - far reaching - consequences of these threats.
One can - of course - also not play this style without a deep search. But
the relative simple minded programms of the past Iike e.g. the first
versions of Deep-Blue (Chip-Test and Deep-Thought) would not have the
slightest chances against the Top-4. Deep-Blue II had also a complex
evaluation. To a certain extend it was already to complex. It is a very
serious problem to tune such an evaluation. I have the same problem in
Brutus.

Another example of the different perspectives is the node-count. Chess
programms use also auxiliary searches, which are not a direct part of the
Alpha-Beta search (e.g. to determine if a position/move is forced). But
these nodes spent in the forced-detection are counted too. The same holds
for the Quiesence search. One could - and this seems to be the Go-approach -
consider the quiesence search as a part of the evaluation.
Brutus is - concerning node-counts - the most Go-like program. The
node-count does not consider the FPGA-part. Calling the FPGA 3-Ply+Quiesence
search is simply 1 node. From the point of the software part of the
programm, the FPGA-search is an evaluation function. The Top-FAQ in Graz
was: "Why is Brutus so slow".

Best Regards
Chrilly Donninger


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go