Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / "infinite" analysis in Houdini?
- - By MaLaoshi (*) Date 2012-12-13 13:00
When I do infinite analysis in Houdini 2.0c or Houdini 3 (both 64b PRO versions),
the computer always stops after 277 hours plus a bit, which seems to work out
to 1 million seconds. Comparing with the behaviour of Rybka 2 and Houdini 1.5a,
it seems this is an issue with the engines and not with my Arena 3.0 GUI. Now
I know that with these Houdini's you're supposed to be able to save&resume,
but frankly that isn't implemented that well and in any case I don't see how such
a hard, undocumented limit would benefit the user.

Any thoughts? Sorry if this is an ignorant question, many people must've come
across this before.
Parent - By Carl Bicknell (*****) Date 2012-12-13 13:26

> When I do infinite analysis in Houdini 2.0c or Houdini 3 (both 64b PRO versions),
> the computer always stops after 277 hours


:eek:
Parent - - By shrapnel (***) Date 2012-12-13 13:31
277 hours:eek:.
Get a life !:roll:
Parent - By mocha1961 (***) Date 2012-12-13 14:04
lolz
Parent - By Peter Grayson (****) Date 2012-12-14 23:29

> 277 hours:eek:.
> Get a life !:roll:


Probably not the answer he was looking for. :grin:
Parent - - By RFK (Gold) Date 2012-12-14 23:46

> 277 hours:eek:.
> Get a life !:roll:


Umh....?

He may not need to get a life. :wink:

But his computer isn't doing much of anything! :yell:

Unless you're directing your post to his computer! :twisted:
Parent - - By keoki010 (Silver) Date 2012-12-15 01:38
Looks like it's busy for 277 hours plus a little bit! :twisted::wink:
Parent - By RFK (Gold) Date 2012-12-15 17:29 Edited 2012-12-15 18:04

> Looks like it's busy for 277 hours plus a little bit! :twisted::wink:


With a diagnoses of  Pathological obsession  :grin:

[edit] (not if it  stops after 277 hours plus a bit :wink: Crash -human terms-nervous breakdown!)
Parent - - By abtp (*) Date 2012-12-13 13:47
"the computer always stops after 277 hours plus a bit, which seems to work out to 1 million seconds."
to me 277 hours translates to 11 days and 13 hours ...

"many people must've come across this before. "
Not me. I'm quite satisfied with H3's output after 6-8 hours or depth 30, whichever comes first.
Parent - - By Eelco de Groot (***) Date 2012-12-13 15:30
I am of course glad that you are satisfied with Houdini. With Stockfish "fork" Rainbow Serpent however (bugfix version is now available but with only a slow 32 bit compile) you should be able to do that depth 30 in not much more than 6-8 minutes on a quad core machine that is, certainly with 64 bits. If the same depth output is however obviously 60 times worse than that of Houdini then I would like to see some good examples :neutral: In other words does depth 30 Houdini beat depth 30 RS with 2^6 amount of time = 6*70 elo? Well, probably so when the difference with equal time is already that much :fat: But what on earth does H3 do to need that much time... I think the PV of Houdini is going much deeper than 30 plies, but that goes for most engines. How much more than for other engines though?

Regards, Eelco
Parent - By Eelco de Groot (***) Date 2012-12-14 01:56 Edited 2012-12-14 08:14
José Mº Velasco made a 64 bit compile!!

AND a better 32 bit I think, mine is not much optimized. See also the Immortal Chess forum. The engine really could profit from 64 bit, because it uses a bit more bitboards, not a whole lot more, not like Komodo does use massive amounts of bitboards, but more than Stockfish and it is still a "Michael Schumacher" style engine that does quite a bit of iterations. Like Ed Schröder once made a Rebel beta that would "fly through the plies" :grin:, he coined that term. Not that 64 bit does that better but it should look good :evil: Can't test the 64 bits version from José myself. I wonder if it will be accepted in the Chessbase Cloud with its real name, and the extra iterations you can do might help a little bit getting a new entry if you are interested in adding analysis to the cloud. (It is not completely depending on the number of iterations of course)

Thanks José!

Eelco
Parent - - By RFK (Gold) Date 2012-12-13 17:37

> after 277 hours plus a bit


Clearly, the problem is related to the "plus a bit"  an analytic worm hole -in short your analysis has entered into an Einstein-Rosen Bridge! :cool:
Parent - - By MaLaoshi (*) Date 2012-12-14 04:16 Edited 2012-12-14 04:18
OK, so I'm being thought a deviant; guess I'm fine with that. :lol:
In a science context it doesn't feel so weird to let a computer crunch
for a month, and I have the hardware. So why not try to see 2015's
analysis--today? After all, few would settle for 18 ply on Fritz 6 in this
day and age, and either that or 40 ply on Houdini surely is a drop in the
ocean compared to the true complexity of chess. To each his own I
guess.

Eelco, great to hear you're hard at work on your engine project, but
wouldn't you get more relevant feedback in a thread which is actually
devoted to that topic?
Parent - By RFK (Gold) Date 2012-12-14 09:17
You're no crazier than I am! You just express your insanity differently than I do! But all in all, you are in good company! :smile:
Parent - - By Antares (****) Date 2012-12-14 12:17

> So why not try to see 2015's analysis--today? After all, few would settle for 18 ply on Fritz 6 in this day and age, and either that or 40 ply on Houdini surely is a drop in the ocean compared to the true complexity of chess.


The problem is that -for most positions- Houdini gets blind over ~d30 pretty quickly, and you can easily find better "continuation" moves (it already pruned out) when manually interacting with it. The idea of "the longer i run the engine the better the resulting line will be" is simply -due limited resources- not correct, in corr-chess there are billions of positions which a chess-engine can't solve on their own - most often because it is attracted by high evals into drawn endgames where lower evals would provide a way to win.
Parent - - By MaLaoshi (*) Date 2012-12-14 12:46
Not related to my original question, but interesting; thanks. Kinda suggests that there could be
a special engine variant/setting for these long-time analyses, like Houdini already got a dedicated
tactical mode. Perhaps give the machine a couple TB high-latency disk storage to play with in
addition to the hash tables in RAM. But without chess programming experience of my own I'm just
speculating here.
Parent - By Waschbaer (**) Date 2012-12-14 12:49
It sounds like Aquarium could be the solution

/Waschbaer
Parent - - By Antares (****) Date 2012-12-14 20:39

> Not related to my original question [277 hours plus a bit]


Well, 277 hours(+a bit) are pretty near to 100.000 seconds, so obviously there is a hard limit... which wouldn't surprise, because there were already way more limiting timing-issues with Houdini 2.0a., but maybe the author Robert Houdart/"Stonehenge" will drop in and explain it to us.

> Kinda suggests that there could be a special engine variant/setting for these long-time analyses,


I suggested this also as a "corr edition", but got no answer. To problem obviously is that the chess-engine doesn't have the structures necessary to store all informations it gathers while the search, while tools (see below) can handle this better.

> Perhaps give the machine a couple TB high-latency disk storage to play with


This is just useful for positions near tablebase-endgame, else it makes no difference as tb-hits are pretty rare.

I don't know what kind of positions you want to "research", but tools -like Waschbaer pointed out- like Aquarium, ChessBase Deep Position Analysis and such a like easily improve on a standalone engine... Houdart stated this himself btw..
Parent - - By MaLaoshi (*) Date 2012-12-15 05:02

>> Perhaps give the machine a couple TB high-latency disk storage to play with
>This is just useful for positions near tablebase-endgame, else it makes no difference as tb-hits are pretty rare.


Oops I meant terabyte not tablebase, should not have abbreviated with the obvious potential for confusion.
(BTW I avoid tablebases on magnetic disks like the plague, even 5-man TB's on a cheapo USB drive is better
than that constant rattling)

Anyway, thanks for the comments, let me read up on this stuff.
Parent - By Antares (****) Date 2012-12-15 16:38

> give the machine a couple TB high-latency disk storage to play with [...] Oops I meant terabyte not tablebase


And for what reason? For a 277h infinite analysis-session the performance-impact of harddisk-speed&size -when not talking about tablebase-accesses- is nearly non-existant (in case of Houdini >=2 maybe for writing the learning-file, or just created by OS- or other processes running at the same time)... the over-all dominant factor is processor-speed, and of course RAM for hash doesn't hurt...
Parent - By mindbreaker (****) Date 2012-12-15 05:19
You can achieve more by making a little tree and running it at those branches.  How long were you planning to run it.  Clearly longer than 277 hours.

For every move there are a limited number of responses.  If you run each for 48 hours then eliminate the ones that are clearly stupid then you could again go one more branching with the fraction that remain.  The end result should be much better than sitting on the same position because the hash that is allocated will be more specifically useful in each line rather than divided.  This can be automated to some degree with "Deep Position Analysis" in the Fritz GUI.  To be more focused at the very beginning you can have 1 for other branching levels.  There are also more advanced ways of discovering a good move in IDEA. 

There is what is called a search horizon effect.  Eventually, on Infinite you are just stuck or have to wait weeks/months for the next ply update.

It could just be a coincidence that it stopped or the GUI.  All it takes is one cosmic ray to hit your CPU or an important memory location.  The sun also is going into a more active phase which increases the number of disruptive particles/photons that could corrupt or disrupt a computer.  Unless you put your computer under the ocean it is not going to be anywhere near immune from that.  You could travel so as to be in perpetual night but that could get pricy ;)
Parent - - By šachista (*) Date 2013-01-10 10:47

> When I do infinite analysis in Houdini 2.0c or Houdini 3 (both 64b PRO versions),
> the computer always stops after 277 hours plus a bit, which seems to work out
> to 1 million seconds.


You mean something like this?

1.Ke3 Le8 2.Lc4 Kf7 3.Kd2 Ke7 4.Kc2 Lf7 5.Kb2 Kd7 6.Kb3 Le8 7.Ld3 Ke7 8.Ka4 Kd7 9.Le4 Kc7 10.Ka5 Lf7 11.Lb1 Le8 12.Lc2 Lf7 13.La4 Le8 14.Lb3 Lf7 15.Ld1 Lg8 16.Lf3 Lf7
  ²  (0.44)   Tiefe: 49/97   111:09:16  3951893843kN
1.Ke3 Le8 2.Lc4 Kf7 3.Kd2 Ke7 4.Kc2 Lf7 5.Kb2 Kd7 6.Kb3 Le8 7.Ld3 Ke7 8.Ka4 Kd8 9.Ka5 Kc7 10.Le4 Lf7 11.Lb1 Le8 12.Lc2 Lf7 13.Le4 Le8 14.Ka4 Kd8 15.Kb3 Ke7 16.Kc3 Kf7
  ²  (0.41)   Tiefe: 50/103   239:20:58  4254127462kN
1.Ke3 Le8 2.Lc4 Kf7 3.Kd2 Ke7 4.Kc2 Lf7 5.Kb2 Kd7 6.Kb3 Le8 7.Ld3 Ke7 8.Ka4 Kd8 9.Ka5 Kc7 10.Le4 Lf7 11.Lb1 Le8 12.Lc2 Lf7 13.Le4 Le8 14.Ka4 Kd8 15.Kb3 Ke7 16.Kc3 Kf7
  ²  (0.41)   Tiefe: 50/111   277:46:40  1365104189kN

Your assumption appears to be correct, as 277 hours, 46 minutes and 40 seconds add up to exactly 1000000 seconds and as you said, the analysis simply stops at this point.
Parent - - By ernest (****) Date 2013-01-11 17:47

> Your assumption appears to be correct, as 277 hours, 46 minutes and 40 seconds add up to exactly 1000000 seconds and as you said, the analysis simply stops at this point.


Well, as I said, not sure it's due to Houdini or due to the GUI...

It seems you are using some Fritz GUI (which one?): strange that in your example I see nodes in kN (kilo nodes) and not in mN (million nodes)?
Parent - - By šachista (*) Date 2013-01-11 20:39

> Fritz GUI (which one?)


9
Parent - - By ernest (****) Date 2013-01-12 20:18 Edited 2013-01-12 20:23
OK, that explains why kilonodes kN and not million-nodes mN (Fritz 11 and later GUIs).

Now, to test if the 1000000 sec analysis limit also happens
with GUI9 and other engine (for instance Critter)
with later Fritz GUI (GUI 13) and Houdini
...takes a looooong time (11+ days)   :cool:
Parent - By šachista (*) Date 2013-01-13 02:33

> Now, to test if the 1000000 sec analysis limit also happens
> with GUI9 and other engine (for instance Critter)
> with later Fritz GUI (GUI 13) and Houdini
> ...takes a looooong time (11+ days)


I'll first test with Houdini 1.5a as i read MaLaoshi's as if there was a difference between Houdini 2/3 and Rybka 2/Houdini 1.5a and i was originally running that analysis to see if this unsolvable position is really unsolvable for H3 given enough time. So far, that seems to be the case with a mediocre i7-950.
Parent - By ernest (****) Date 2013-01-11 17:41

> When I do infinite analysis in Houdini 2.0c or Houdini 3 (both 64b PRO versions),
> the computer always stops after 277 hours plus a bit


Well, not sure it's due to Houdini or due to the GUI...
Parent - - By Stonehenge (***) Date 2013-01-13 11:57
Sorry for my late reply, I only now notice this thread. If you have a real support question it is usually more efficient to send an e-mail to houdini@cruxis.be.

"Infinite analysis" is hard-coded as 999,999,999,999 microseconds in Houdini, I never imagined this would be a problem.
You could work-around this limit by specifying a longer time in the "go" instruction, but this is probably not straight-forward in most GUIs.
In the next release I'll add another "999," so that with the default "infinite" you can run 11,500 days uninterrupted :).
Parent - - By Akbarfan (***) Date 2013-01-13 17:07

> In the next release I'll add another "999," so that with the default "infinite" you can run 11,500 days uninterrupted :).


will take 31 years  :eek::eek::eek:  :wink:
Parent - - By šachista (*) Date 2013-01-13 17:56
In other words, the next Houdini release won't have infinite analysis either.
Parent - By Banned for Life (Gold) Date 2013-01-13 18:01
:lol:
- - By rocket (****) Date 2013-01-11 16:46
When I put a a tactical position through infinite analysis, the  houdini engine finds the solution a few seconds slower than if I just press "go" and the usual playing option,
I have tried this twice to see that it wasn't a coincidence because of the random generator(of which particular move the engine prefers for each try).

It should be exactly the same speed putting it through infinite analysis!
Parent - - By Barnard (Bronze) Date 2013-01-11 16:58
why dont you ask directly to Houdart?

maybe he will be able to help you and explain that to you

regards
Parent - - By rocket (****) Date 2013-01-11 17:00
I wanted to check if other people have experienced the same thing
Parent - By Barnard (Bronze) Date 2013-01-11 17:05
at least me,i never experienced that

but maybe,and just maybe,analizing it acces to learn file,and playing not...and i recall MAYBE,because i dont know,is just an idea
Parent - - By ernest (****) Date 2013-01-11 17:38

> if I just press "go" and the usual playing option,


Then you are not doing infinite analysis.
The engine will stop analysing after a time control defined by the Level which you defined.
For instance, if you set the Level at "Blitz Game" 4min+2sec
(and if you are not following the Book, in which case Go will immediately play the next Book move),
Go will analyse only around 10 to 15 sec.

Besides, in order to compare, you need to be in 1 Thread mode, because you need reproducibility.
Parent - By rocket (****) Date 2013-01-12 23:39
<"Then you are not doing infinite analysis.>"

I  wrote than IF I press "go button", I am well aware that one is game playing option, and the other is infinite analysis.
Up Topic The Rybka Lounge / Computer Chess / "infinite" analysis in Houdini?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill