Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Bob Hyatt's views about Cluster Rybka
- - By ernest (****) [fr] Date 2008-11-10 19:28
Come on, boys, let's have a healthy discussion :-p on Bob's comments in CCC/Talkchess, concerning Cluster Rybka (from what he saw in the ACCA Panam tournament)

http://www.talkchess.com/forum/viewtopic.php?t=24807

"At least the hyperbole about the "40 core monster" has been discovered. Yes, Rybka is extremely strong. But the cluster Rybka is absolutely no stronger and perhaps a bit weaker due to the unsynchronized depth issue."
Parent - - By Ray (****) Date 2008-11-10 20:08 Edited 2008-11-10 20:31
I suspect he is quite right - in WCCC Rybka did not seem to have the brute force of 40 CPUs, more like an Octal at most. That is just an opinion and is very subjective of course...
Parent - - By Dragon Mist (****) [hr] Date 2008-11-10 21:52
Wow! This one goes directly to the second place of my "impossible claims" list! Almost got to No.1, which is that "Rybka didn't play as 40 core should, unlike Hiarcs did", or something like that anyway.
Parent - - By Ray (****) Date 2008-11-10 23:01

> Wow! This one goes directly to the second place of my "impossible claims" list!


Indeed, that's why I said it was just my opinion and very subjective :-)
Parent - By Dragon Mist (****) [hr] Date 2008-11-10 23:36
Ok, fair enough. :-)
Parent - - By InspectorGadget (*****) [za] Date 2008-11-11 08:40
Crappy algorithms, poor end game, bad evalution of blind Bishops, no Bishop under promotions.

Beats the crap out of every program. :)
Parent - - By oudheusa (*****) [nl] Date 2008-11-11 11:21
A bit harsh for the #1, but I fully agree on the endgame handling. On average I would estimate Rybka 3 to be about 0.20 points too optmistic in pure endgames. This is indeed BAD!
Parent - By InspectorGadget (*****) [za] Date 2008-11-11 14:15
No, I just mentioned the things that people like to complain about. Despite all those things, miss Fish still beats everyone who is better at the things I mentioned.

My main complaint is the Blind Bishop.
Parent - - By boo! (**) [no] Date 2008-11-10 22:58
With the very limited time between the R3 launch and the WCCC I wouldn't be surprised if the cluster implementation used was somewhat crude. However, knowing how seriously the Rybka team take their testing, I couldn't imagine the cluster being worse than R3 on an octal. Why take the risk?

The funny thing with Bob's rybka posts is the way he talks about her as he would a 6-ply searcher of one of his students.
Parent - - By zwegner (***) Date 2008-11-10 23:16

> The funny thing with Bob's rybka posts is the way he talks about her as he would a 6-ply searcher of one of his students.


Especially when he says things like "Rybka is extremely strong".

Personally I'd say that Rybka is probably stronger on the cluster, but by a small margin.
Parent - By Carl Bicknell (*****) [gb] Date 2008-11-10 23:34

> Personally I'd say that Rybka is probably stronger on the cluster, but by a small margin.


Vas has already answered this. He said it was 100 elo > octal.
Parent - By NATIONAL12 (Gold) [gb] Date 2008-11-10 23:32
this is rubbish.i have a Skulltrail and overclocked as well,i can state that Lukas 32 core is faster than my Skully.
Parent - - By BB (****) [au] Date 2008-11-10 23:34

>Come on, boys, let's have a healthy discussion :-p on Bob's comments in CCC/Talkchess, concerning Cluster Rybka (from what he saw in the ACCA Panam tournament)


Reading the whole thing, he actually does have some useful observations, though the conclusion "absolutely no stronger" might be a bit stretching it. His main claim is:

Several were noticing that Rybka's depth and score were bouncing all over the place. I finally figured out why on a position where Crafty played QxQ and Rybka had one recapture move (RxQ) and several other moves. We were seeing all sorts of -9.0 scores being kibitzed. As I went through them, what I discovered was this: Rybka uses an algorithm first tried (and then discarded because it is useless) back around 1978 or so by Monty Newborn. What is happening is this: Each root move is given to a different cluster node when the first iteration of the search starts. Each move is searched independently until time runs out. The depth is not synchronized in any way...

From which he derives the conclusion that the Cluster Speed-Up is zero.

[...] why is the depth bouncing around going forward and backward all over the place? But if you look more carefully, and find just the kibitzes for the same move and ignore the others, you will find that the depths and scores now look sensible.
The primary fault with this is that the speedup is zero. If you watch the typical program, 90% of the search effort is on the _first_ move. Here, each node is searching a "first move" which breaks alpha/beta completely. Even worse, there is a depth issue in that each of the root moves are often searched to different depths. Which is better, +1 at depth 17 or +.01 at depth 21? That was the problem Monty discovered and that is why this is no good, without even considering the issue that the distributed search is not speeding up the process one bit.
I have no idea why such an algorithm would be used. There is no up-side to doing this and using a single node would be no slower. And it actually would be a bit faster since all root moves get searched to the same depth so that there is no issue in comparing scores from different depths.


I leave it to the reader to try to determine whether Hyatt has made a correct diagnosis of the ClusterRybka methodology.

The VR evidence claimed about 100 ELO from: Sure, it's a ballpark figure. We didn't try to measure it precisely. At 10' + 5" I got a score of 17-7 (+12 =10 -2). About which, one can note this result is a difference of 154 ELO, with a standard deviation (ignoring colour) of about 80 ELO.
Parent - - By Dragon Mist (****) [hr] Date 2008-11-10 23:42

> What is happening is this: Each root move is given to a different cluster node when the first iteration of the search starts.


But this is what some of us figured out a couple of days after the event, when Lukas and Vas were commenting lightly on the cluster thing. It is something alike as if using cluster speed-up not to deepen the search so much, as to widen it (as if constant PV mode of sort). So no points for Mr. Hyatt for that one.
Parent - By BB (****) [au] Date 2008-11-11 00:03

>But this is what some of us figured out a couple of days after the event, when Lukas and Vas were commenting lightly on the cluster thing. It is something alike as if using cluster speed-up not to deepen the search so much, as to widen it (as if constant PV mode of sort). So no points for Mr. Hyatt for that one.


Hyatt disagrees: :-)
I am absolutely certain of what he is doing, nothing else explains the output we saw yesterday and the data was obvious once I began to look at the output carefully. It was producing so much "noise" that at first I paid no attention until someone asked "why is its depth bouncing up, down, back up, and why are the scores all over the place?" A quick and careful look explained what was going on.
As I say, with only a few examples given, I can't examine his claims. Perhaps with a full log it might be more amenable to analysis. In any case, the Conspiracy Theorists will tell us that VR is intentionally obfuscating the Cluster Output so as to mis-lead the observer regarding what is going on... :-P :-P

Also, Hyatt muses that 100 ELO is a 4x speedup, so he is assuming 50 ELO per doubling --- my guess is that this is a tad low (even in the depths/ranges mentioned here), at least with respect to the VR testing method, as head-to-head tests have been noted (occasionally) to have higher ELO/doubling than elsewise.
Parent - - By AsosLight (***) [gr] Date 2008-11-10 23:47

>>As I went through them, what I discovered was this: Rybka uses an algorithm first tried (and then discarded because it is useless) back around 1978 or so by Monty Newborn.


Actually that's the point of his observation.
From that assumbtion concludes all the following.
I am not a software engineer i cannot comment uppon that.
From the other side i think that even if they managed to achieve a x4 speed up, which is impossible, they could not get +100 points,if they didn't make the search much much more efficient.
Parent - By Uly (Gold) [mx] Date 2008-11-11 03:46

> if they didn't make the search much much more efficient.


They possibly did, Vas has been talking about Rybka 3's search inefficiencies, perhaps because Rybka 4 Beta has them already fixed.
Parent - - By BB (****) [au] Date 2008-11-10 23:47 Edited 2008-11-10 23:52
Hyatt also throws down the gauntlet in a later post (hoping for a 35x speed-up on a 70-node academia-sized cluster):

I brought this up because of the various discussions and comments yesterday, particularly people asking about "why all the different depths? why is the depth bouncing up and down? How fast is this 5-node 8-core-per-node cluster? Etc." The answer is it is no faster than a normal 8-core 1-node box. I am finally making preparations to do a _real_ cluster approach. And I know I am not going to get 100 Elo from a 5-node parallel search. You lose too much. No shared hash table. No shared killers. No shared anything whatsoever. So the losses are large. But if you can get a real 2.5 speedup out of 5 nodes, that would be significant, as it would be like searching 50M nodes per second on the hardware I used yesterday when we were typically searching 20M nodes per second. Or more importantly, if I can get 35x on the big 70-node 8-core node, that will be serious numbers. 700M nodes per second is serious computing. I have no idea whether I can get that kind of speedup. I believe I can get 2.5x out of 5. But stretching that ratio to 70 is improbable. But anything > 1.0 is good...

I leave it to others to speculate how strong a 560-core Crafty might be... (of course, this being from academia, any methodological gains he finds will be published, eventually --- or perhaps sooner if there is no practical gain :-P)

In yet a later post, he seems to think fixed depth is a reasonable metric to measure multi-core speed-up: I still hope that one day someone will post some real numbers on Rybka's parallel speedup on an 8-way box, by running some "normal" positions to a fixed depth using 1, 2, 4 and 8 processors. He claims to scale better than any other program. Somehow I doubt it. Maybe "as good as" if he is lucky. But so far we just have urban legend to go on. Speedups for my program are quite easy to produce and anybody can do it. However, in contrast to this, one can note that VR (and myself also) thinks that a direct match with time odds is a superior measurement method (though I don't know if there will really be much observed difference between these methods, as the error margins might be too large --- from a philosophy standpoint, the direct match is obviously better).
Parent - - By zwegner (***) Date 2008-11-11 00:02
How do you get this:

> Hyatt also throws down the gauntlet in a later post (hoping for a 35x speed-up on a 70-node academia-sized cluster)


from this:

> I believe I can get 2.5x out of 5. But stretching that ratio to 70 is improbable.


??

> In yet a later post, he seems to think fixed depth is a reasonable metric to measure multi-core speed-up: I still hope that one day someone will post some real numbers on Rybka's parallel speedup on an 8-way box, by running some "normal" positions to a fixed depth using 1, 2, 4 and 8 processors. He claims to scale better than any other program. Somehow I doubt it. Maybe "as good as" if he is lucky. But so far we just have urban legend to go on. Speedups for my program are quite easy to produce and anybody can do it. However, in contrast to this, one can note that VR (and myself also) thinks that a direct match with time odds is a superior measurement method (though I don't know if there will really be much observed difference between these methods, as the error margins might be too large --- from a philosophy standpoint, the direct match is obviously better).


That is the _only_ way to measure speedup. He's not talking about strength.
Parent - - By BB (****) [au] Date 2008-11-11 00:27

>>How do you get this:
> Hyatt also throws down the gauntlet in a later post (hoping for a 35x speed-up on a 70-node academia-sized cluster)


But if you can get a real 2.5 speedup out of 5 nodes, that would be significant ... Or more importantly, if I can get 35x on the big 70-node 8-core node, that will be serious numbers. 700M nodes per second is serious computing. I have no idea whether I can get that kind of speedup. I believe I can get 2.5x out of 5. But stretching that ratio to 70 is improbable.

Sigh, Hyatt hits me with high-sounding hype, I stop reading closely when he says he has no idea about practicalities, and then he turns pessimistic at the close of the paragraph... Perhaps, as with any good politician, he can give enough soundbites so as to be all things to all people.. :-P

>That is the _only_ way to measure speedup. He's not talking about strength.


I've lost what your antecedent to "that" is --- and perhaps also what "speedup" is as compared to strength in your nomenclature. I claim that: in order to get a useful measure of how much better a 4-core engine than the same engine in single-core, you should play them against each other at time odds (say 3.5x), and from this determine the time ratio needed to equalise the results. I would simply term the number derived from this as "speedup", and this would also seem to be of most value to a user in terms of a single number for multi-core comparison. Fixed depth numbers are of academic interest (see Hyatt's Cray Blitz paper), but I don't necessarily put much real-world value on them. However, as indicated previously, one might guess that the difference between these methodologies is not all that large, so the argument might be moot in the end.
Parent - - By zwegner (***) Date 2008-11-11 02:00

>>> How do you get this:
>> Hyatt also throws down the gauntlet in a later post (hoping for a 35x speed-up on a 70-node academia-sized cluster)
> But if you can get a real 2.5 speedup out of 5 nodes, that would be significant ... Or more importantly, if I can get 35x on the big 70-node 8-core node, that will be serious numbers. 700M nodes per second is serious computing. I have no idea whether I can get that kind of speedup. I believe I can get 2.5x out of 5. But stretching that ratio to 70 is improbable.
>
> Sigh, Hyatt hits me with high-sounding hype, I stop reading closely when he says he has no idea about practicalities, and then he turns pessimistic at the close of the paragraph... Perhaps, as with any good politician, he can give enough soundbites so as to be all things to all people.. :-p


This quote also puts things in perspective:

> I think 70 will be a challenge to use without a _lot_ of work. But I can see modest numbers (say 16) possibly.


As for hype, I would say Bob has done more to advance parallel theory in computer chess than any other person. So I'm definitely looking forward to any results he gets. I think this would be something long overdue: all distributed (non-SMP) approaches to chess haven't been very advanced. Two recent ones are Chessbrain and Rybka. Chessbrain sucked completely, and it seems Rybka's is overly simplistic. Bob has unrestricted access to an absolutely huge machine, which means plenty of time for tuning and experimenting (which can't be said for Diep or Zappa). Adding to the fact that Crafty is open source and we will be able to see this thing in action, I think we should give him the benefit of the doubt.

> I've lost what your antecedent to "that" is --- and perhaps also what "speedup" is as compared to strength in your nomenclature. I claim that: in order to get a useful measure of how much better a 4-core engine than the same engine in single-core, you should play them against each other at time odds (say 3.5x), and from this determine the time ratio needed to equalise the results. I would simply term the number derived from this as "speedup", and this would also seem to be of most value to a user in terms of a single number for multi-core comparison. Fixed depth numbers are of academic interest (see Hyatt's Cray Blitz paper), but I don't necessarily put much real-world value on them. However, as indicated previously, one might guess that the difference between these methodologies is not all that large, so the argument might be moot in the end.


I mean the only way to measure speedup is by searching a big set of positions to a fixed depth and comparing time to depth. Your idea of time odds is interesting (and perhaps more meaningful practically), but it's not so easy. There are variables that are dependent on time that wouldn't be covered there (time management and pondering, for one). You are probably right that there is not much difference between them. My impression was that you were measuring "speedup" by Elo gain.
Parent - - By BB (****) [au] Date 2008-11-11 02:36

> all distributed (non-SMP) approaches to chess haven't been very advanced.


ClusterToga might think otherwise. But I don't claim to know much here, other than that it is still a work in progress.

>I mean the only way to measure speedup is by searching a big set of positions to a fixed depth and comparing time to depth.


I'm just not convinced of this. Particularly as "depth 20" with 1 core and "depth 20" with 8 cores might have much different average tree breadths. Incidentally, VR/LK have used fixed depth in trying to determine value of tablebases, and there I have the same problem, in that TB usage can change the tree substantially, and it seems a bit unwarranted to assume that the situation can generically be projected into a fixed depth metrization.

> I think 70 will be a challenge to use without a _lot_ of work. But I can see modest numbers (say 16) possibly.


My guess is that 35x will still appear somewhere in the grant proposal... "Early experiments indicate we can reasonably expect a 2.5x speedup on 5 nodes, and even though a linear relation seems rather difficult, we might still hope for a 35x gain on the UAB 70-node cluster." :-P Incidentally, the DeepBlue speedup, independent on the metric, was rather low --- I think Hsu said that the 480 cores were only about 8-12% efficient or so, meaning that it was only about 48 times better than 1 core. See 6.3 of http://sjeng.org/ftp/deepblue.pdf

>There are variables that are dependent on time that wouldn't be covered there (time management and pondering, for one).


In an open-source setting, time management should be hackable. Or even more crudely, you could just do something like divide the gettimeofday by a suitable factor before passing it to the engine. The question of pondering is more subtle; I simply assume pondering off, and that this is a good surrogate for real chess. However, it might be that pondering can be handled in a more trenchant manner in the parallel setting.
Parent - - By zwegner (***) Date 2008-11-11 22:26

>> all distributed (non-SMP) approaches to chess haven't been very advanced.
> ClusterToga might think otherwise. But I don't claim to know much here, other than that it is still a work in progress.


Possibly. I was ignoring CT because nobody has any information about it. When the paper/source come out, then we can make judgments.

> I'm just not convinced of this. Particularly as "depth 20" with 1 core and "depth 20" with 8 cores might have much different average tree breadths. Incidentally, VR/LK have used fixed depth in trying to determine value of tablebases, and there I have the same problem, in that TB usage can change the tree substantially, and it seems a bit unwarranted to assume that the situation can generically be projected into a fixed depth metrization.


I can see how TBs would alter the search tree, but SMP search will (should) not. It will of course search more nodes (except in a few special cases), but most of these branches don't provide any new information, they are just searched because of inefficiency. Of course, if Vas is doing something that alters the fundamental search algorithm based on how many processors there are, then you can't really extract any meaningful information at all.

> My guess is that 35x will still appear somewhere in the grant proposal... "Early experiments indicate we can reasonably expect a 2.5x speedup on 5 nodes, and even though a linear relation seems rather difficult, we might still hope for a 35x gain on the UAB 70-node cluster." :-p Incidentally, the DeepBlue speedup, independent on the metric, was rather low --- I think Hsu said that the 480 cores were only about 8-12% efficient or so, meaning that it was only about 48 times better than 1 core. See 6.3 of http://sjeng.org/ftp/deepblue.pdf


I don't think he needs to write a grant proposal to use a machine that he already has access to. :) We'll see how well he does. Deep Blue is a special case though, as it only had 30 "cores". From there each node has 16 chess chips which are all tied to the main processor, and the parallel efficiency there isn't much (basically equivalent to a root splitting method).

> In an open-source setting, time management should be hackable. Or even more crudely, you could just do something like divide the gettimeofday by a suitable factor before passing it to the engine. The question of pondering is more subtle; I simply assume pondering off, and that this is a good surrogate for real chess. However, it might be that pondering can be handled in a more trenchant manner in the parallel setting.


I don't think that would be a very accurate measurement. Different engines have their strength scale differently over time (and parallel speedup also varies with absolute time). Add to this that measuring elo accurately requires thousands of games... Really, the only way to get some sort of universally comparable measurement of parallel efficiency is time-to-ply. But even then, if the search algorithm varies at all between single- and multi-cpu versions, any comparison doesn't mean much, just as depth and nodes can't be compared to different engines.
Parent - By BB (****) [au] Date 2008-11-11 23:37

>I don't think he needs to write a grant proposal to use a machine that he already has access to. :-)


You must not understand academia?! :) Indeed, grant proposals are also written for "research funding", so as to avoid teaching... :-P [You should realise that, as an academic of sorts, I am rather cynical about such things].

>It will of course search more nodes (except in a few special cases), but most of these branches don't provide any new information


This might depend on precisely what you mean by information. For instance, a wider 14-ply search can allow more "information" when searching at ply 15, as the backed-up hash values might be superior in that more nodes are considered completely (be it as a PV node, or as a lesser-ly pruned node). I simply find it difficult to dismiss such concerns out-of-hand w/o some experimental evidence. For instance, VR recently claimed that 4-way MultiPV R3 would beat R3 by 50 ELO at fixed depth, though that might just be guesswork.

>I don't think that would be a very accurate measurement.


I think many of your contentions regarding fixed time could be transported to fixed depth; for instance, the need to test at different depths. Also, while my back-of-the-envelope calculation does indeed find that one can achieve a greater level of statistical precision via a given amount of experiment using fixed-depth, I would guess that its relevance to expected chess results would be smaller than with the fixed-time approach.

>Really, the only way to get some sort of universally comparable measurement of parallel efficiency is time-to-ply.


I don't think one can compare efficiencies of different methods of parallelisation in a universal manner via time-to-ply, as outlined in the paragraph above. Incidentally, though it might be irrelevant with some engines, one could hope to discern the usefulness of a time-to-ply experiment via recording how often the single- and multi-core searches returned different best moves (or maybe just scores) for a given depth. Finally, on a lighter note, I can't find time-to-ply to be too universal for engines that don't use Iterative Deepening (and thus don't have ply counts). :-P :-P
Parent - - By BB (****) [au] Date 2008-11-11 04:16
Incidentally, for those who are interested in the current work on Crafty, Hyatt gives a broad outline here, including some comments about new testing methodology, and a paragraph about the Rybka game.
Parent - By Uly (Gold) [mx] Date 2008-11-11 05:02

> and a paragraph about the Rybka game.


For short:

"By move 33, [...] it looks easy (to the humans watching) to draw. [...] But we did make an evaluation mistake that eventually led to losing the game."
Parent - By Roland Rösler (****) [de] Date 2008-11-11 04:40
I see only Bobs praying for satori! :-)
And he does it in his way: thesis and accusal and damnation.
I don´t think, this is the best way to get best (or any) answers from Vas.

PS: I watched the game Rybka vs. Crafty. I´m sure, Rybka Skulltrail would have won too. Craftys play in the end was weak, that´s the whole story!
Parent - - By Nelson Hernandez (Gold) [us] Date 2008-11-11 13:23
I didn't watch the games but this is interesting stuff from Hyatt.  It perfectly matches what I would guess as well.  Parallel searches along different root moves is exactly what Vasik thought would work best in Freestyle.  He has mentioned it a few times so it would not be surprising if that is how he arranged to have the hardware set up.  After Rybka development we have to remember that Vas is an ardent Freestyler always looking for an edge.
Parent - - By Banned for Life (Gold) Date 2008-11-11 18:47
Its very hard to believe that Vas would implement parallel search at the root without sharing of state information between searches. The problems of independent searches from the root are well known and documented. I suspect that if Vas is indeed splitting at the root, he is looking at new ways of sharing critical information between nodes to achieve results closer to what one could achieve with shared alpha/beta values and hash table. His recent statements to the effect that his cluster algorithm is nearly as effecient as SMP, although hard to believe, might indicate that he has had some success in this endeavor to significantly reduce the communication bandwidth required to share important state information. Note that this is very similar to what is needed in an SMP configuration when the number of nodes gets large.

Regards,
Alan
Parent - - By Vasik Rajlich (Silver) [pl] Date 2008-11-18 07:59
These things are really complicated. The cycles are:

1) Think
2) Implement
3) Do other things
4) Think again
5) Realize that it can be done much better

Slowly, everything cycles forward.

It's true that centaur analysis during freestyle games has been an excellent source of ideas. I typically have several machines to coordinate and the machines have limited ability to share state. It's basically the same problem.

Anyway, the 100 Elo figure going from one node to five nodes is real. I'm willing to make a wager on this.

Vas
Parent - - By Uly (Gold) [mx] Date 2008-11-18 21:58

> Anyway, the 100 Elo figure going from one node to five nodes is real. I'm willing to make a wager on this.


Answer by Robert Hyatt:

"That's always the way to wriggle out of a discussion. You _can_ do as I usually do and produce real data and post it here. Or you can say "I'll bet you what I claimed is true..."

I'm not interested enough to make a wager and waste further time... I know what splitting at the root can gain. It has been documented for 30 years. I suppose I could run the test on Crafty easily enough since it splits at the root in addition to everywhere else. I could just make it split only at the root. Stay tuned and I'll post some real data since they obviously are not..."
Parent - - By keoki010 (Silver) [us] Date 2008-11-18 23:04
Vytron, where did you get this from?
Parent - - By Uly (Gold) [mx] Date 2008-11-19 00:19

> Vytron, where did you get this from?


Original thread at:

http://www.talkchess.com/forum/viewtopic.php?t=24807&start=60

The computer chess fora is divided because Robert Hyatt will not post in this forum and Vasik Rajlich will not post at the other forum. I'm trying to act as a bridge.
Parent - - By BB (****) [au] Date 2008-11-19 01:11

>The computer chess fora is divided because Robert Hyatt will not post in this forum and Vasik Rajlich will not post at the other forum. I'm trying to act as a bridge.


I think they are divided for a reason...? :) I should make a Bridge to Nowhere joke here... :-P
Parent - By Banned for Life (Gold) Date 2008-11-19 01:39
On November 16, 2005, Congress stripped the specific earmark allocation of federal funds for the two bridges in the final edition of the omnibus spending bill, without changing the amount of money allocated for use by Alaska.

Getting this much money without any strings attached was a grand slam. Alaska will miss Senator Stevens.
Parent - - By Uly (Gold) [mx] Date 2008-11-19 02:26

> I think they are divided for a reason...? :-)


Unless the people of one forum hate the people of the other forum, I don't see a reason.
Parent - - By Vempele (Silver) [fi] Date 2008-11-19 06:07
You mean they are people at *censored*? I thought they were some sort of degenerate, half-human...erm, things. :-p
Parent - By Uly (Gold) [mx] Date 2008-11-19 06:27
[Deleted]

I'm one of them and I'm running for moderator, so *no comments.*

[Deleted]

:P
Parent - - By BB (****) [au] Date 2008-11-19 06:32
Wow, in the Crafty 22.4 release (two days after 22.2), Hyatt says he got 14 ELO from dumping pawn islands from the eval. I shouldn't have thought that it could be anywhere near that much. The original value was 10/20 centipawns for 3/4 islands (0 for less), tending to 20/30 in the endgame (if I am reading the source correctly).
Parent - - By chessorpoker (**) [us] Date 2008-11-19 07:21
i for one, question everytime hyatt has something negative to say about rybka. crafty is pathetic and slow, rybka1.0 beats it 100/100 games what else is there to say?
Parent - By Uly (Gold) [mx] Date 2008-11-19 07:25
Does Rybka 1.0 still beats 100/100 Crafty 22.4? I've read that Crafty got big improvement (relative from version 22.1, not the 14 elo from 22.3.)
Parent - By zwegner (***) Date 2008-11-19 16:28
I, for one, question every time you say anything. Because... well, that should be obvious.

P.S. What is the win rate of your engine against Crafty?
Parent - By Vasik Rajlich (Silver) [pl] Date 2008-11-22 07:37
Ok, when we have some time we'll do a controlled test. I ran a total of something like 1000 blitz games with several versions over the span of several weeks and the result was something like +100 Elo, this won't be too far off.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Bob Hyatt's views about Cluster Rybka

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill