Author: Robert Hyatt
Date: 20:28:30 12/24/99
Go up one level in this thread
On December 24, 1999 at 22:14:14, Vincent Diepeveen wrote: >On December 24, 1999 at 21:34:34, Robert Hyatt wrote: > >>On December 24, 1999 at 17:20:54, Vincent Diepeveen wrote: >> >>>On December 24, 1999 at 16:54:39, Robert Hyatt wrote: >>> >>>>On December 24, 1999 at 16:32:49, Vincent Diepeveen wrote: >>>> >>>>>On December 24, 1999 at 15:53:49, blass uri wrote: >>>>> >>>>>>On December 23, 1999 at 19:24:48, Vincent Diepeveen wrote: >>>>>><snipped> >>>>>>>Bob i've got proof of bad working of singular extensions. >>>>>> >>>>>>I think it is possible that they had a rule of not extending >>>>>>wasting tempo moves so deep thought could not see when that 18.Rg1 is >>>>>>losing when it played 16.c4 because black waste tempo in the line Qh4-h3 Rg1 >>>>>>Qxh2+. >>>>>> >>>>>>It is also possible that they did not extend this line because they did not >>>>>>extend lines that give the queen for a pawn. >>>>>> >>>>>>You need to see 14 plies from the move 16.c4 without extensions to find >>>>>>22...Qxh5# >>>>> >>>>>Don't forget check extensions in qsearch and exending on check, >>>>>apart from that singular extensions. >>>> >>>>Why are you still stuck on the deep thought hardware problems? DB 1 and >>>>DB 2 were _far_ different. DT didn't do nor understand checks/mates in the >>>>q-search. It didn't understand repetitions in the hardware either. DB 2 >>>>fixed both of these problems as well as added an order of magnitude more >>>>gates for the evaluation. >>> >>>This was hong kong: deep blue 1. >> >>Vincent: Listen _carefully_. It was _not_ Deep Blue 1. Deep Blue 1 did >>_not_ exist at that point in time. Either email Hsu or Campbell and ask, >>or check the tournament publicity stuff. It was _very_ clear. "Deep Blue >>Prototype". >> >>Or wait for Hsu's book and notice exactly _when_ the first DB 1 chips were >>delivered to them. Hint: It was _way_ after Hong Kong. You are talking about >>the original chiptest/deep thought chess processor. DB prototype used the new >>software search that would be used in Deep Blue. But it used old chess >>processors with lots of known bugs. >> >>This is all public knowledge. It is all old news. Forget about "Deep Blue 1 >>in Hong Kong" It didn't happen. >> >> >> >> >>> >>>Even deep thought should have seen of course. Put crafty to fullwidth >>>and see how little plies it needs to see c4 is very bad. Just give >>>a big bonus to not having a pawn on c2 or c3. See my previous post. >> >>And what is the purpose for funking up the eval? And what is the purpose >>for comparing to a machine that hasn't played chess since Hong Kong? DB1 and >>2 are _far_ different. Trying to extrapolate based on things Deep Thought >>did is about as useful as tryint to extrapolate based on things Crafty did 4 >>years ago. It is _nothing_ at all like it was back then. Neither is DB1 or DB2 >>anything like deep thought in many respects... >> > >Didn't Hsu write about singular extensions *many* years before hong >kong? > >I remember 1989 or something? yes... > >Funking up the evaluation shows that even a stupid program that's piece >square table needs only 9 plies here even without singular extensions. >doing checks in qsearch won't need much more plies in this position. perhaps >1 ply extra at most. > >I'm clearly doing some eval experiments here. > >What i try to show is the use of singular extensions. >sometimes singular extensions see a trick a ply sooner. >sometimes even 2 ply. Some positions kingside are even made for that >type of extensions. That's no surprise. _ANY_ extension is two-edged. You see some things quicker, others take longer... Nothing magic there as the search space (number of nodes) is constant, but the search tree shape is not. > >It is not gonna solve chess however. Of course it isn't.. but it is a significant leap. As I said, ask anybody that has gotten it working in any form. Myself. Hsu. Bruce. Lang. Kittinger. To name the ones that I know have used a form or forms of singular extensions with success. > >If that was the case, then deep junior we could rename as deep blue 3. >If it gets like 18 plies or something, then it in fact sees everything >that has to do with attacks and checks till 18 plies of depths, excluding >some checks (not counted in search depth) and some recaptures (not counted >in search depth). Yet it didn't solve chess yet. > >Now deep blue getting in a simple position (from branching factor >viewpoint) only gets 11 ply against kasparov. > >How do you *ever* want to make your claim true that a hardware guy >can extend 40 ply in the right lines, where i cannot? I don't even have to answer that. :) I didn't discover how to split the atom. But that doesn't mean someone didn't. Someone a bit smarter than I am. :) > >This though all practical experiments show the opposite! > >Everyone has thrown out singular extensions, and replaced them with >some threat extensions. Not to see mainlines deeper, but only to >see a few checking lines deeper which have to deal with mate. Everyone? Lang? Bruce? Cray Blitz? Kittinger? > >So basically i see extensions as something that's needed to prevent >horizon behaviour. Horizon behaviour i only see happening with >passed pawns and especially with mates. > >Using those extensions will in positions where i can put my >opponent to mate sometimes see a shot, or prefer big misery for me. > >That however OBVIOUSLY doesn't happen every game. So basically the >average move is not getting helped by threat-extensions. Considering >singular extensions are a part from what we call threat-extensions, >then this also applies to singular extensions. > >Now some objective calculations. Let's compare 2 programs. Deep Junior, >basically looking deeper at tactics, with deep blue. If deep blue >gets an 11 ply search in a position where i get with a few tens of >millions of nodes already a lot deeper, >then obviously in most middlegame positions we can expect deep blue >to have searched at most 10 ply. > >When i include root move as apossible to extend move, then we have >from that 10 ply only 6 ply to extend our moves on. > >You asked in previous posts here loudly whether that matters. >YES that matters a lot. > >That means that in those 6 ply left even in the biggest forced line >we have very likely AT MOST 3 ply to extend (of course not counting >checks in the line as those are extended by any program). So all >possible singular extensions that might apply to the leafs, WON'T >apply to the leafs. As the singular extension algorithm never >considers them to possibly extend. > >So the average forced line case : 10+3 = 13. >And that only for tactical tricks which i might already >evaluate correctly and extend for some reasons at the leafs already. Your math is simply bad. If Crafty can find a mate in 10 with a 6 ply search, with _no_ checks in the q-search. Without recognizing check or mate in the q-search. Then obviously I am extending more than 3 plies for a 10 ply search... > >Now we still didn't take into account the values you told me. > >You told me that for S you're using quite a big margin and expected >deep blue to use quite a big margin too. Secondly the most interesting >singular extensions you're only using 2*S difference before considering >to extend it. You can find the JICCA paper. They didn't use a "big value" It was a fraction of a pawn, only. I don't remember the exact value without having the Journal handy. I seem to remember 38, with a pawn=128. In Cray Blitz, we settled on roughly 1/3 of a pawn for our SE value. That allowed us to play what I thought was the best chess, even though a bigger value worked better on tactical tests because it didn't extend as much on non-tactical branches. > >This obviously means that only on material win or loss stuff gets extended. That only means you implemented it wrong. I could extend by getting a rook on the 7th, or damaging the opponent's king safety, or creating a passed pawn, or lots of other significant positional changes.... Also remember I reported that they did a new modification to this extension for DB-2... if there was one singular move, they extended as in the JICCA article. If two moves were better than the rest, they extended each, but not as much as when there was only one singular move... > >That exactly is what i wanted to go to. > >Deep Junior is already getting like 17 plies in those lines... > >With the selective search versions of DIEP where i tried to search >deep in paderborn i also got to 14 or 15 ply, just seeing tactics bigtime >deeper, and the higher evaluated leafs (which is already seeing more >than singular extensions as i didn't use a margin S, but simply the >best few evaluated leafs which is a kind of bestfirst search, but >also INCLUDING singular extensions). > >It missed the world actually. > >I learned it the hard way after like 2 years of bigtime experiments >that extensions are there to see a few lines deeper which are NEEDED >to see deeper. The important lines are for the biggest part >unforced though and definitely there is no way to see them deeper, >especially not something that's just alfabeta dependant and not taking >any chess heuristics into account. > >My own experiments were past 2 years already a lot further than just >alfabeta depencencies. I wrote hundreds of lines of code to select moves >that were also plausible to see deeper (so to not prune). > >Even that combined with alfa beta values and windows didn't bring me >anything further. > >O yes, it was cool to solve some tactics, but it played positional >simply not as well as the same program without the pruning. > >This because chess is so much harder than the few rules i had to select >moves. > >>>We talk about Deep Blue a few months before it plays Kasparov. >>>This is not an evaluation issue only. My eval sees it 1 ply sooner than >>>a simplistic piece square table program. The piece square table >>>version of DIEP needs only 9 plies to see this. >>> >>>I don't DOUBT that extending all checks sees it even sooner. >>> >>>>What is the purpose of trying to find errors in late 1980's hardware and then >>>>from that extrapolate things about DB2? They have very little in common. >>>> >>>> >>>> >>>> >>>>> >>>>>That all rips plies from it. apart from that you don't need to >>>>>see all lines to already not play c4. >>>>> >>>>>If evaluation of Deep Blue I was really that bad, then it still should >>>>>see all this tactical. If it didn't , then obviously its fullwidth >>>>>search with singular extensions and some other crap didn't work for >>>>>this obvious position. >>>> >>>> >>>>That was _not_ deep blue. It was deep thought hardware. This has been >>>>well-documented for several years now. It was called (at one point) >>>>deep blue prototype. With the note that "this is the DB software search >>>>but using the original deep thought hardware." >>>> >>>> >>>> >>>> >>>>> >>>>>It clearly refutes the idea that singular extensions solve the game. >>>>>I see it very simple. A professional hardware designer makes a >>>>>special chip called deep thought, deep blue, deep blue II. >>>>> >>>>>An incredible achievement on its own. >>>>> >>>>>Then he's inventing singular extensions. >>>>>Though all commercial programmers who >>>>>dedicate fulltime to their program and many other researchers >>>>>can't get it to work, or make it work in such a way that it only >>>>>gives them sometimes in tacitcal position 1 or 2 ply more, >>>>>this hardware designer, can? >>>> >>>> >>>>I got it to work. Bruce got it to work. Lang got it to work. Kittinger >>>>got it to work. >>>> >>>>> >>>>>And apart from this claim that it's seeing the right plies, >>>>>but then it falls into an easy joke? >>>> >>>> >>>>not the same hardware _at all_... >>>> >>>> >>>> >>>>> >>>>>I'll put now evaluation off from DIEP (not smart in itself, but i'll >>>>>just let it do material) and see how many plies after c4 it sees >>>>>serious trouble, besides seeing whether it plans c4 at all. >>>>> >>>>>>I do not think that the reason of not seeing this tactics was singular >>>>>>extensions because I believe that with singular extensions and some check >>>>>>extensions it is possible to see that 16.c4 is bad. >>>>>> >>>>>>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.