Author: Uri Blass
Date: 01:41:36 02/16/06
Go up one level in this thread
On February 15, 2006 at 18:07:33, Ryan B. wrote: >On February 15, 2006 at 17:52:35, Nathan Thom wrote: > >>My engine is still very young (aptly named LittleThought) and I'm having some >>issues explaining its behaviour. Features so far: >>- Iterative Deepening >>- AlphaBeta with limited move ordering (only captures are scored so far, with >>SEE) >>- Eval is simply material+piece sq >>- Q search non-losing captures+checks+proms >>- No hashing, no extensions, no pruning >> >>I just put in the Q search and a test run of 20 secs from the opening gives: >> >>With Q Search >> >>00:00:00.00 20n 1/1 0.58 1. e4 >>00:00:00.01 512n 2/2 0.08 1. e4 d6 >>00:00:00.01 2364n 3/4 0.00 1. h3 h6 2. g3 >>00:00:00.03 9224n 4/6 0.00 1. h3 h6 2. g3 h5 >>00:00:00.18 88Kn 5/9 0.00 1. h3 h6 2. g3 h5 3. f3 >>00:00:00.65 361Kn 6/17 0.00 1. h3 h6 2. g3 h5 3. f3 h4 >>00:00:12.97 8628Kn 7/37 0.00 1. h3 h6 2. g3 h5 3. f3 f5 4. e3 >>00:00:20.00 13Mn 8/38 0.00 1. h3 h6 2. g3 h5 3. f3 f5 4. e3 >>Total Nodes = 994854 >>Total Q Nodes = 866141 >>Total Beta Cuts = 99773 >>Total Q Beta Cuts = 459231 >> >>Then I wanted to see how deep it could get without Q search, assuming it should >>get further due to less Q nodes searched: >> >>Without Q Search >> >>00:00:00.02 20n 1/1 0.58 1. e4 >>00:00:00.05 512n 2/2 0.08 1. e4 d6 >>00:00:00.08 10Kn 3/3 0.56 1. e4 d6 2. d4 >>00:00:00.31 148Kn 4/4 -0.05 1. Nf3 d6 2. e4 e5 >>00:00:02.23 1468Kn 5/5 0.93 1. e4 e6 2. Bc4 d6 3. Bxe6 >>00:00:20.00 16Mn 6/6 -0.42 1. d4 Nf6 2. Nf3 Nc6 3. e4 Nxe4 >>Total Nodes = 2709329 >>Total Q Nodes = 0 >>Total Beta Cuts = 184327 >>Total Q Beta Cuts = 0 >> >>This really surprised me and I still cant explain it properly. It searched more >>nodes overall but to a lesser depth. My feeling is that its to do with beta >>cutoffs and the Q search was somehow causing more of them, but after counting >>them I see that the beta cuts within the normal search tree is actually less >>when Q search is turned on. >> >>I'm sure its somehow related with the fact that the Q search seems to return >>scores of 0.0 all the time (which sounds right as the opening is very stable) >>and without the Q search, the scores seem to alternate +/- depending on who made >>the last move. Also without the Q search, the moves seem to be smarter and make >>use of the piece square tables more. >> >>Does it sound like a bug, or is this expected behaviour? > > >Not a bug, because of the qsearch you are getting more reliable scores causing >quicker cutoffs and a more stable search. Also people may be surprised that if >the knowledge is good some times added chess knowledge helps search depth for >the same reasons. > >Ryan still it seems to me a bug. With qsearch he gets 2364 nodes at depth 3 and without qsearch he gets 10Knodes. considering the fact that there are almost no captures at the first 3 plies from the opening position the big difference suggests that there is a bug. The fact that he has pv of h3 with qsearch also suggests a bug. 512 nodes at depth 2 also suggest a bug considering the fact that from the opening position perft(2)=400 and alphabeta should give smaller node count at depth 2. Note that I am not surprised if good knowledge can help search depth espacially when I think that good knowledge is not only knowledge in the evaluation but also knowledge which lines to prune and which lines to extend(it is certainly chess knowledge and one of the advantage of strong players is that they know to search the relevant lines). Uri
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.