Author: Vincent Diepeveen
Date: 08:13:44 01/17/00
Go up one level in this thread
On January 17, 2000 at 09:32:40, Robert Hyatt wrote: >On January 17, 2000 at 03:11:34, Jouni Uski wrote: > >>Yes stunnigly they are available at: >> >>http://www.research.ibm.com/deepblue/watch/html/c.html >> >>First thing I notice is, that search depth is not very convincing at all - >>mostly 11-12 ply (also in endgame part)! Have I read log wrong way? Most >>Chessbase engines got same ply in standard PC... >> >>Jouni > > >It is a "different" kind of ply. DB didn't use null-move of any kind. They You can also search a lot less deeply without alfabeta pruning. Even with an incredible amount of nodes you hardly will come deeper than a few plies then. >also used singular extensions which cause the typical chess ram to search >1-2 plies less deep in the typical case, although it will probe much deeper Nah, singular extensions were introduced when chiptest searched up to 8 ply at tournament level. From experiences at ICC server i can ASSURE everyone that 8 ply is not enough to see even some simple tactical tricks. SURE singular extensions gave a lot back then, as no one got to 8 ply then, so everyone was suffering from major tactical problems. Those get away when slowly getting above 10 ply, which is near to 6 moves with some extensions, as was already estimated by De Groot back in 1966, as being the 'basic plan depth of a master'. Later when they had deep thought II, getting 10 ply with an incredible number of nodes a second, then the deep thought team themselves CLAIMED that singular extensions didn't give much anymore. Let me quote: "the main conclusion that must be drawn from the experiments onse lective search reported in this article is that while selective search, in particular domain-independant selective search, does indeed improve the performance of a chessprogram, no single heuristic is able to contribute a large improvement. Moreover, many heuristics or variations that might appear promising in theory turn out to be of little or no value in practics". I can support to some extend this opinion. Nearly everything a human does is extremely knowledge based. I'm never gonna look at a certain forced line if my brain doesn't feel like wnating to look like it. alfa nor beta have to do with this (as far as you can as human be aware of any values, i'm not). Then conclusion goes on: "the total improvement of the best combination of heuristics discovered so far over brute force is 72 points (86 USCF). Of this 49 points is due to threat extensions, 7 points is due to singular extensions, 5 points due toe PV lookahead extensions, with the reaminder due to a number of less important extensions". I could not assigne values to the worth of extensions, and I'm missing recapture extensions btw. Note that i could never proof them working in tournament games either, but in blitz they kick butt, just like mating extensions. I agree however with the conclusion. Everything a human does is knowledge based. If we use a bit of knowledge, either dynamic or static, then this means that search is gonna miss a lot. obviously you need something to see deep threats/checks a bit deeper, to not lose all games because of a mate in 7. However, i feel only evaluation is important. My current version of DIEP is tactical simply weak. However if i turn back on all extensions it used to have, then it's gonna score like 300 points at least more at a given testset than the current version (the current version does it really weak on those testsets, only win at chess gets still solved with 299 out of 300, which btw is bigtime better than chiptest with 500k nps did, though evaluation based). I feel however that at slow level (where diep gets >= 10 ply) score of diep is not gonna be different a lot. How much difference is at blitz we'll see within some weeks/months at the icc server against other programs and against humans. >in the traditional sense. One other unknown is whether their 10 plies includes >the 4 plies of hardware search. They don't count total nodes that I noticed If they print '11' then they did get 11 ply brute force, which INCLUDES the chessprocessors. It's impossible to get fullwidth without hashtables at chessprocessors, with the way parallellism was set up, and without the chessprocessors communicating hash etcetera with each other to get 15 ply fullwidth, WITH a load of extensions. Note that Hsu describes that a typical 12 ply search was first 4 ply at processor 1, then 4 ply at 30 different processors, then 4 ply in hardware. >anywhere, so it isn't easy to decide. For me, 1M nodes per second gives a >depth of 13-14 in the middlegame. Going 200M should drive that to roughly >log3(200) which is about 5 plies deeper. So call that 18-19 plies. Removing >null-move would subtract 2, so 16-17 plies. Singular extensions 1-2 more plies, >so that would be (14-15) to (15-16) plies. >Somewhere on their web site they mentioned 14 plies as "normal". So their >iteration number might be slightly different from "ours". IE when I search >until depth is reduced to zero, I go to the quiescence search. They might >go to their hardware, which is not exactly a "quiescence search" since it looks >at all moves for 4 plies, but doesn't do any of the singular stuff and so forth. >I'll try to ask next time I hear something from Hsu. He implicitly answerred that in april this year, apart from that it's fullwidth not possible given the way it was implemented with all the conditions and extensions. Deep thoughtII searched many tens of millions nodes a second and got 10 ply. deep blue was more NPS optimized, and got just a ply or at most 2 ply deeper. very logical. Vincent
This page took 0.01 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.