Author: Tord Romstad
Date: 09:05:33 02/04/04
Go up one level in this thread
On February 04, 2004 at 11:04:36, Thorsten Czub wrote: >On February 04, 2004 at 06:39:49, Tord Romstad wrote: > >>This human does not find them obvious at all. > > >I guess most newbies in chess would instantly throw the bishop into h6 without >thinking too much. > >This is how CSTal works. > >the static evaluation says: do it. the search now has the problem to make the >best out of the situation. this is a different aproach s8 or other programs >try. they really try to find out about the situation. > >so in fact they do most of the work the opponent should do themselves. > > > >> The game is really cool, and >>the bishop sac on h6 is very interesting. Nevertheless, I am very surprised >>that it worked. > >i was in the same situation. you sit behind the machine. than it is about >TRUSTING. you trust s8 of course - world chess champion, new 18 searches. > >but suddenly the score goes into 0,00. Shredder had problems with >draw-repetition lines. cstal evaluates draw-repetitions different than other >chess programs. >if cstal can draw, the draw is POSITIVE because WE can force whenever we want. > >so we have a draw and MAYBE MORE. this is IMO how humans think too. Phalanx does the same as CSTal. Perpetual check draws are given a small plus score (+0.2, IIRC). It's an interesting idea, but I haven't tried it out myself, and I am not sure how well it would work. Especially when playing against weaker opponents, I am skeptical. If my engine has an advantage of 0.19, I don't want it to sacrifice a handful of pieces for a perpetual check in the hope that there happens to be a mate. But I think some sort of refinement of this idea could be effective. >> After Bxh6 gxh6 Qxh6, white has no attacking >>pieces except the queen, > >but therefore white has to bring pieces to the king. >CSTal knows that this is important. All engines know that this is important. The question is whether it is possible to do it quickly enough. CSTal evidently thinks it is possible, while most other programs don't. Personally I have no idea. >it has special algorithms that tell the >program WHICH constellations of pieces are dangerous and which constellations >are not a problem. > >when it sees a dangerous constellation, it wants to realize it on board. >therefore the normal rook moves. it's like a trajectory for CSTal. I wonder how >Smarthink would play in this position. > >It seems to have similar kind of algos. > >>and needs several moves to bring pieces to the attack. > > >thats called a plan :-)) Thank you for bringing up SmarThink and the idea of trajectories. This topic is very relevant to the discussion. Let's assume that Bxh6 is really correct, and that the attacking plan of Rd3 followed by f4 and Rg3 really is quick enough. This is precisely the kind of maneuvre which tends to be pruned away by null move pruning. Black has a big material advantage, makes a null move, and the reduced depth search is often not deep enough to see an attacking plan containing several non-forcing moves. This is the *really* big problem with recursive null move pruning. It is very surprising that so few engine developers seem to be interested in it. I think it is hard to solve this kind of problem purely by static evaluation. If efficiency was no concern, it would have been nice to include the number of safe moves every black and white piece on the board would need in order to attack/defend the exposed king in the king safety evaluation. But of course, doing such a calculation at every node would be extremely expensive. One of the many items on my list of ideas I want to try some day is to do a "horizontal search" in all nodes more than n plies from the horizon. This horizontal search should try to find the paths each piece can traverse in order to safely reach a square where it participates in the attack, assists a passed pawn, or some other important goal. In the CST-Shredder game, such a search would find the sequence of moves Rd3-f4-Rg3. The move sequences found by the horizontal search would then be used to guide the search by extending the interesting moves and/or avoiding to prune directly after them. This is probably similar to Markoff's trajectories. SmarThink, by the way, is in my opinion the most interesting chess engine which is under active development these days. >>Gothmog, my own attempt to emulate the style and feel of CST, does not like >>Bxh6. It thinks two pawns and an exposed enemy king is not enough to >>compensate for a bishop when there is only one attacking piece. > >but what about the rook. >a queen and a rook ARE dangerous. why is gothmog not capable to see that you get >the rook into the attack ? The question is not only whether it is possible to bring the rook into the attack, but also whether it is possible to do so more quickly than black is able to bring sufficient defensive resources to the kingside. This is not easy to decide without a deep search. Null moves makes it even more difficult, of course, like discussed above. >I remember CSTal played a similar game in paris. also with such a rook movement. >and it saw it STATICALLY. >not by search. but with the evaluations. >like humans do it. I am not sure I agree that that is how homans do it. We always do a small horizontal search as part of our evaluation function (thinking something like "and here I can bring the rook to the attack via d3"). 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.