Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Oops...

Author: Robert Hyatt

Date: 06:58:44 05/02/00

Go up one level in this thread


On May 02, 2000 at 01:41:23, Christophe Theron wrote:

>On May 02, 2000 at 00:15:44, Christophe Theron wrote:
>
>>On May 01, 2000 at 23:36:30, Robert Hyatt wrote:
>>
>>>On May 01, 2000 at 22:50:50, Christophe Theron wrote:
>>>
>>>>On May 01, 2000 at 21:59:05, Robert Hyatt wrote:
>>>>
>>>>>On May 01, 2000 at 20:39:13, Christophe Theron wrote:
>>>>>
>>>>>>On May 01, 2000 at 18:50:24, Robert Hyatt wrote:
>>>>>>
>>>>>>>On April 30, 2000 at 22:34:17, Christophe Theron wrote:
>>>>>>>
>>>>>>>>On April 30, 2000 at 18:42:57, Amir Ban wrote:
>>>>>>>>
>>>>>>>>>On April 30, 2000 at 16:39:04, Christophe Theron wrote:
>>>>>>>>>
>>>>>>>>>>On April 30, 2000 at 06:47:29, Amir Ban wrote:
>>>>>>>>>>
>>>>>>>>>>>On April 29, 2000 at 11:31:09, Eran wrote:
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>I am sorry if I said it. Okay I believe you that Junior6 has underpromotion code
>>>>>>>>>>>>and that's wonderful. Maybe I will consider buying it. Does Junior6 consume
>>>>>>>>>>>>hashtable memory as large as fritz does? Is having large hashtable memory
>>>>>>>>>>>>important for Junior6? Is 40 MB hashtable enough for 40/120 games?
>>>>>>>>>>>>
>>>>>>>>>>>>Eran
>>>>>>>>>>>
>>>>>>>>>>>More memory for hash is good, but Junior is not very sensitive to it and you can
>>>>>>>>>>>change memory size by order of magnitude without obvious effect on playing
>>>>>>>>>>>strength.
>>>>>>>>>>>
>>>>>>>>>>>The comparison to Fritz is interesting and backward: I believe that Fritz 6
>>>>>>>>>>>(new) consumes less memory than previous versions and the reason may be a
>>>>>>>>>>>conversation I had with Frans about this in Paderborn, from which he may have
>>>>>>>>>>>decided that he doesn't need so much memory.
>>>>>>>>>>>
>>>>>>>>>>>Amir
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>I must admit I don't understand what you say...
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>    Christophe
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>Then why don't you ask :)
>>>>>>>>>
>>>>>>>>>I understood from Frans that he's hashing quiescence nodes. I told him I don't,
>>>>>>>>>and that I consider it a waste of time.
>>>>>>>>>
>>>>>>>>>Amir
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>I hash quies nodes. I tried both methods (hashing them and not hashing them) and
>>>>>>>>found that hashing QSearch nodes was definitely better, but not by much. I did
>>>>>>>>hours of experiments and drew a lot of curves with my spreadsheet in order to
>>>>>>>>find this.
>>>>>>>>
>>>>>>>>It works better even if the hash table is highly saturated.
>>>>>>>>
>>>>>>>>That's how it works for me. I don't think that I have a better hashing/replacing
>>>>>>>>strategy, actually it is rather simplistic. Maybe it's because I do more in
>>>>>>>>QSearch than both of you do, although I cannot know for sure.
>>>>>>>>
>>>>>>>>Maybe I should check again... :)
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>    Christophe
>>>>>>>
>>>>>>>
>>>>>>>I used to hash q-search nodes.  I found it reduced the size of the tree by about
>>>>>>>10%.
>>>>>>
>>>>>>
>>>>>>
>>>>>>Yes, that's approximately what I measured.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>  I also found that by taking it out, I reduced the total search time by
>>>>>>>15 %.  A net gain of 5%.
>>>>>>
>>>>>>
>>>>>>
>>>>>>That's where I differ. Taking hash probe/updating out of QSearch results in a
>>>>>>very marginal speed gain (in NPS) for me.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>  More importantly, I don't need nearly the hash memory
>>>>>>>now since over half of the search is not hashed.
>>>>>>
>>>>>>
>>>>>>
>>>>>>I think there is no saving in memory. In QSearch, if the hash table slot is
>>>>>>available, I use it and will maybe gain something, and if it's not available
>>>>>>(because a more important information is stored) I end up not using the hash
>>>>>>info in QSearch, as you do.
>>>>>>
>>>>>>Not hashing in QSearch means you are not taking advantage of all the memory you
>>>>>>have to help you to decrease the size of your tree. It does not mean that you
>>>>>>will need less memory.
>>>>>>
>>>>>>You do not take advantage of a resource you have, as your result of a 10%
>>>>>>smaller tree when using HT in QSearch shows.
>>>>>>
>>>>>>Of course, the increased NPS in your case justifies your choice.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>My next task is to save some time by getting rid of the hashing/unhashing code
>>>>>>>in the q-search as well, since it isn't used...
>>>>>>
>>>>>>
>>>>>>Yes, I thought of this possible speed gain. But then you have to write a special
>>>>>>version of make/unmake that you will use exclusively in QSearch, and design
>>>>>>another system to check for repetitions for example (as I assume that you use
>>>>>>the hash key to detect repetitions).
>>>>>>
>>>>>
>>>>>
>>>>>Not possible in my q-search.  I don't do checks.  Only captures.  As a result,
>>>>>I don't do repetition checks at all since they can't possibly happen.
>>>>
>>>>
>>>>
>>>>I understand. So obviously you'll earn some extra % of speed by taking out the
>>>>hashing code.
>>>>
>>>>
>>>>
>>>>
>>>>>>This is maybe worth the effort for you, because you just generate captures in
>>>>>>QSearch (that means no repetition can happen in QSearch). But my QSearch does
>>>>>>more than that, so for me it's a real problem. My QSearch catches a lot of draws
>>>>>>by repetition.
>>>>>
>>>>>I did this is CB.  I am not sure it is the right thing to do, as several good
>>>>>programs are using a restricted (or non-existant) capture search.  Ferret is
>>>>>one example, Junior is another.  The benefit right now is that I use all the
>>>>>extra time to extend in the basic search, where I generally encounter fewer
>>>>>errors than in the q-search.
>>>>
>>>>
>>>>
>>>>The benefit of the kind of QSearch I do is not big. I do it mainly because I
>>>>find it more elegant.
>>>>
>>>>If it was worse than doing a simplistic QSearch, I would not do it. But as it is
>>>>not worse, I tend to use the most "elegant" approach (which is a matter of taste
>>>>in this case).
>>>>
>>>>
>>>>
>>>>
>>>>>>You also need to keep the hashing of pawn structures anyway, so all hashing will
>>>>>>not be taken out of the QSearch make/unmake.
>>>>>
>>>>>
>>>>>right... although that is (in my code) a 32 bit hash, rather than a 64 bit
>>>>>one.
>>>>>
>>>>>>
>>>>>>
>>>>>>But I understand that not hashing in QSearch can be a good decision for some
>>>>>>programs.
>>>>>>
>>>>>>
>>>>>>
>>>>>>    Christophe
>>>>>
>>>>>
>>>>>Bruce and I flipped and flopped on this one.  I started off hashing, while I
>>>>>don't think he did.  Then I took it out but he added it.  It is very close
>>>>>to 'equal' in my code.  And if it is equal, I prefer not doing it as it doesn't
>>>>>stress memory nearly as much, and lets me do long searches without needing huge
>>>>>hash tables...
>>>>
>>>>
>>>>
>>>>I understand the "stress memory" point.
>>>>
>>>>But I don't buy the "need less hash tables" argument. This is just not true.
>>>>
>>>>The hash table slots you fill during QSearch will never compete with the slots
>>>>you fill during the "normal" search (I do not know how you call it).
>>>>
>>>>If you use the standard replacing scheme which uses the depth of computed
>>>>subtree as the priority factor of a given slot, then it's obvious that a node
>>>>computed in QSearch will never replace a node computed during the normal search,
>>>>nor prevent a node computed in the normal search to replace it.
>>>>
>>>>By not sending QSearch nodes into the hash table, you are NOT allowing more
>>>>space for the other nodes. You are just wasting memory space. I understand that
>>>>the time saved in the process can justify it, but it's wrong to think that you
>>>>are saving a single byte of memory.
>>>>
>>>>
>>>>
>>>>    Christophe
>>>
>>>
>>>
>>>This is a well-known bug.  If you use totally depth-prefered replacement, you
>>>run into a big problem, in that the table (in a deep search) gets filled with
>>>'deep positions' and the positions near the tips don't get stored.
>>
>>
>>
>>I guess that now you know which replacement strategy I'm using! :)
>>
>>
>>
>>
>>>  Yet they
>>>are _critical_. My approach is the one used by Ken Thompson, that of using
>>>two tables, one depth-prefered, one always-store.
>>
>>
>>
>>OK, so in this case you are right, storing QSearch nodes does reduce the space
>>for the other nodes, in the "always replace" table at least.
>>
>>
>>
>>
>>>Probably the best way to compare two programs is to do the following:  Pick a
>>>position, and search (using the same hardware) for 300 seconds, while varying
>>>the hash table for both from something very small, to something very big.  For
>>>each program, you will find a point where the improvement slope drops sharply
>>>and levels off.  We need to test a q-search prober vs a non-q-search prober.
>>>
>>>I can run the test for Crafty if you want...  and can run from something tiny
>>>up to 384M max...
>>
>>
>>
>>I can do the same, but I'll not be able to use more than 16Mb of RAM, as my
>>computer has only 32Mb.
>>
>>I have already done this kind of experiment and plotted the curves with a
>>spreadsheet. Actually I have done the experiment with 2 versions of my own
>>program: one using HT during QSearch, the other one not using HT during QSearch.
>>The programs are otherwise totally identical.
>>
>>The experiment was not done on only one position, but on 40 positions. The
>>positions were taken from 2 actual games (20 positions of the first game, 20
>>positions of the other game) played by an old version of Chess Tiger against
>>Rebel Decade, one with white the other one with black. The positions are
>>consecutive ones, and range from early middlegame to early endgame.
>>
>>The positions have all been searched to a fixed depth of 8 plies. The size of
>>the hash table ranged from 32Kb to 8Mb. I have checked that the point where the
>>slope of efficiency drops exists in this range of hash table sizes.
>>
>>The curves are interesting and show that, for my program, using the HT during
>>QSearch is always better, and surprisingly is much more interesting when... the
>>HT begins to be saturated!!
>
>
>
>Oops... Sorry, I'm tired.
>
>The above is complete bullshit. The curve of course shows the opposite: using HT
>during QSearch is better when the HT is not saturated.
>
>Which is not surprising.
>
>
>    Christophe
>
>
>
>
>>If you are interested, I can send you the small Excel4 file (or export it to
>>text file).
>>
>>I would be interested in the curve that could be drawn from the experiment you
>>are describing.
>>
>>
>>    Christophe



I will run this test.  I did it a couple of years ago and posted in r.g.c.c
in response to a request by "Komputer Korner".

The thing I worry about is this:  for a hash table of size N, there is obviously
a specific number of nodes that will fit in it at that size.  Suppose you fill
it by depth=12.  What happens when you go to depth=20?  The last 8 plies can't
get in, as their depth is too low.  And you will see the search die.  And die
badly, performance-wise.  That is the purpose of the 'always-store' part.  I
have 1/3 of my table holding deep search results.  And 2/3 of the table is
always store so it holds the "local search" results.

It is those way-deep searches that make this a problem...  IE overnight, or
at todays hardware speeds, even 10 minutes.  I can easily hit 1M nps on my
xeon.  Using 384M of hash, I get about 24M hash entries.  I can overrun that
so fast it would be a joke.  Even not hashing q-search sees this fill up in
searches of < 1 minute...



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.