Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / opponent modeling
- By Vasik Rajlich (Silver) [hu] Date 2007-01-09 14:25
Nelson Hernandez wrote:

-----------------

What a supremely confident statement! This reminds me of Vas's remark after winning the Freestyle tournament that went something to the effect "why mess with programs that are 150 ELO weaker?"

Well, here's why you should reconsider. If you collect a large number of varied engine games you will notice that certain engines invariably favor certain moves in certain positions. The result is that whole branches in the move-tree are dominated by specific engines. You find cases where two engines dominate a particular tree-branch because they work together idiosyncratically to achieve a given position. Position 'x' may be a common Fritz-Spike position which would never be attained if the game were Spike-Fritz, let alone Fritz-Rybka.

The result of Rybka-incest is that your opening book gets narrow-gauged and overly anticipates Rybka moves which are certainly not optimal in 100% of all cases.

------------------

Alan wrote:

-------------------

Nelson,

Good points.

My understanding is that Noonen is in charge of books for engine tournaments and Krasnakow is in charge of openings for freestyle tournaments so I'm not sure that Vas spends a lot of time thinking about openings.

The concept is definitely important, but Team Rybka won't need to worry about this until their engine tournament opponents wise up and start doing opening prep for the games. Minus any effort in the opening, Rybka will do well based on its greater strength and Vas's penchent for turning up with the fastest hardware.

Regards,
Alan

-------------------

Exactly, Jeroen Noomen handles our tournament openings (and Krasenkow the freestyle openings - whatever happened to those tournaments, anyway?). In opening theory, the opponent definitely does matter and Jeroen plays tons of test games against other engines.

There are probably some evaluation features which should in theory be opponent-dependent, but would make things quite a bit more complicated and the benefit would probably be pretty small.

Alan - your comment reminds me a bit of something I heard about Kasparov - that he doesn't just prepare novelties, he anticipates novelties for his opponents and analyzes those.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / opponent modeling

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill