Author: Inmann Werner
Date: 05:32:37 03/22/00
Go up one level in this thread
On March 22, 2000 at 08:05:13, Jan Pernicka wrote: >On March 22, 2000 at 07:12:59, Inmann Werner wrote: > >>On March 22, 2000 at 06:30:58, Jan Pernicka wrote: >> >> >>I use up to 9 ext in one branch. If you have deep mate threats with much check, >>it is helpful. >> > > OK - but I think it slows speed (i.e: completing searching) too much... > - Do you distinct "good" and "bad" checks ?! > That number - 9 extens. - do you use it for all (which?) types of extensions? > this questions(better say-answers) are crucial in making "good" program. > Also interesting is (in average) ratio: r = a/b , where: > b - #nodes visited when no extension is used > a - #nodes visited (in the same position as in a) ) with extensions > Note: in this case I include quiescence search also into b because > I think (is it true?) that other extensions have nothing common with > quiescence search - i.e.:when reaching the horizon (depth==0), they > in case a) use the same quiescence search as in case b)... (?) > > I suppose that ratio r is normally around 3 or 4 - is it true or not? > (But in your case it can sometimes be much, much more - seems...?!) > All the following is only, what I do. (InmiChess) It sure could be done better, but also more bad...(IMHO) - All Extensions together in one branch are not allowed to exceed 9 ext. (seems good with actiual search depths...) - I make check extensions always, if distance to quietsearch is less than 7 . - taking back a piece without changing material balance is a ext, if it is at the last two plies before quiet search. - if a pawn gets to line 7 at the horizon, I extend (and only then!) - promotions are in the quiet search - if in check, also in quiet I allow all moves to get out - if in check, and only one move leads out I extend. - null move extension, i only allow once in a branch (it blows up the tree too much) - I only extend once in one depth, never twice! But i remember the extensions for deeper down. - my quiet search is no own routine. it is within my normal search, own movegenerator and some ifs (although ifs are bad at the new computers, but small code is good for cache hits..) My ratio from one ply to another is between 2.7 and 3.8 normally (player-opponent effect) I do not know the exact "explosion rate" of my extensions. Don´t ask me, why I do this in particular. I changed it one year long, and at last, i was happy with it. I am not really happy with the pawn extension (there should be more), but till now found nothing better working, not exploding the tree. >> >>Reasons of win: >>3)the engine makes a king side attack, not seeing anything tactically, but >>"liking" positionally the attack. It gives pressure until it sees something >>tactically.(this is tricky, cause often, if the attack fails, you are lost) >>(Comet? Hossa) > > This seems to me not too safe... It is not safe, and therefore I always again got rid of it. But some "king safety" would be fine :-) Example: 3 pieces attacking around near the king, and the king has no defenders, but no piece win in horizon.... > >>4)Both engines do not know really what they do, but at one time, one stands >>clearly better and wins. (Both engines give 0.00 or so and move around) > > It would be interesting to know how ofter this (point 4) ) happens... more often, then anybody thinks. But you can also the,"Oh what wonderful positional understanding", although the eval says something different. (It sure more often happens with the non-top-engines.) > >> >>This also is a ranking IMHO > > Excuse me, what IMHO means ( i am not too strong in abbr. :) ) I like your question :) I posted the same sometimes ago, and learned, that it means "in my humble opinion" and now I try to show, how a good "learner" I am :-) > > So - let's summarize the problem: > how to make extension "good-and-safe-working"? > (i.e: better than exhaustive searching with relatively low time costs) really good question, no real answer. Werner
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.