Author: Tord Romstad
Date: 03:10:21 10/17/03
Go up one level in this thread
On October 16, 2003 at 18:37:50, Ross Boyd wrote: >On October 16, 2003 at 11:01:56, Tord Romstad wrote: > >>On October 16, 2003 at 10:29:25, Ross Boyd wrote: >> >>>Hi Tord, >>>After adding EGTB support to TRACE her rating improvement was barely measurable. >>>I think it was 6 elo after at least 200 games. >>>And she is definitely an engine that lacks endgame knowledge. :-) >> >>I added tablebase support in Gothmog last week, but haven't tested the >>improvement yet. > >Great! I prefer to keep EGTB support even if the improvement is negligible for >TRACE. I hope YMMV. :-) > >> >>>Often I see TRACE swindle a draw in R+PvR endgames and sometimes she swindles a >>>RvB endgame for a win. >> >>In games against Phalanx (my favorite sparring partner), Gothmog often >>gives away a half point in advantageous rook endings by allowing Phalanx >>to simplify to a Philidor-type position. Phalanx apparently knows that >>these positions are drawn, but Gothmog still thinks it has good winning >>chances. I think this happens often enough that adding tablebase support >>should give a small, but measurable improvement. > >These are the type of where the EGTBs do payoff. But even with EGTBs, I think it is a good idea to include knowledge about endgames like KRPKR in the evaluation function. It is clearly advantageous to be able to correctly evaluate these endgames in the qsearch, for instance. The danger of EGTBs is that it is easy to become lazy and not spend any time and effort on teaching the engine to play elementary endgames. >>Of course, it would be much better to actually implement knowledge about >>the third-rank defence in my evaluation function. I will probably do >>so some time in the future. >> > >I agree its best to have the knowledge built in so that the knowledge can be >applied before its too late. Yes. You also cannot expect all users to have the complete 5-men tablebases, of course. >>>But accessing the egtbs is very slow so frequently she loses because her >>>NPS is severely hampered.. >> >>Huh? I haven't noticed any slowdown at all. When do you probe? I probe >>in all positions where the last move was a capture, and the remaining depth >>is at least 5 plies. > >I probe only after captures in all interior nodes... maybe that's my mistake. I >might try limiting this to 5 plies from the frontier as you do. I haven't experimented much with this myself, and I am not sure 5 is the optimal number. The best solution might be to let this be a user-configurable parameter. >>Perhaps I simply have a much faster hard disk than you, or perhaps my NPS >>is so slow to begin with that the disk access doesn't matter much. > >It halves my nps usually... or worse. And my HD is not the latest and >greatest... The widely differing HD speeds is another reason to let the user decide how near the leaves EGTB probing should happen. >BTW, I'm a fan of Gothmog. My preference is for knowledge-based engines like >yours that cut through the tree with lots of smarts. Thanks a lot for the compliments! I would describe it as speculative rather than smart, though. On the few occasions when everything works well it is able to produce some nice kingside attacks or even smooth positional games, but more often it plays embarassingly ugly chess. I share your preference for knowledge-based engines, and Gothmog aspires to be one, but there is still a very long way to go. >Its beaten up TRACE on many occasions. Yes, but I have seen TRACE beat Gothmog several times, too. Our engines are rather close in playing strength, I guess. >She's slow AND stoopid, unfortunately. :-) Like most programmers, you are probably overestimating her stupidity. We are all painfully aware of the shortcomings of our own engines, but don't perceive the weaknesses of our competitors to the same degree. I wish I knew your engine better. There are no Windows computers around here, which means that I cannot play with most of the many nice amateur engines like TRACE. Hopefully I can some day afford an apartment big enough to house a Windows machine in addition to my Mac. Good luck in the last two rounds of the Open Class in the RWBC! It looks like we'll soon meet again in Class C. :-) Tord
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.