Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Genius hash scores

Author: Christophe Theron

Date: 23:10:08 01/08/00

Go up one level in this thread


On January 08, 2000 at 23:55:58, Chessfun wrote:

>>>Cel 433
>>>Genius 5. 2kb hash = 1.f6 switches again to Kxg2 after 44 secs then solves
>>>1.f6 posting + score after 3 min 48 secs.
>>>Genius 5. 4kb hash = 1.f6 switches again to Kxg2 after 40 secs then solves
>>>1.f6 posting + score after 3 min 25 secs.
>>>Genius 5. 8kb hash = 1.f6 switches again to kxg2 after 1 min 19 secs solves
>>>1.f6 posting + score after 4 min 29 secs.
>>>Genuis 5. 16kb hash = 1.f6 switches again to kxg2 after 59 secs then solves
>>>1.f6 posting + score after 3 min 25 secs.
>>
>>
>>You mean Mb, not Kb, I presume?
>Correct.
>
>>
>>A very funny thing: Genius has almost the same problem than Tiger: at one point
>>increasing the hash table size gives a longer solution time. Then increasing
>>further yields again a better result.
>>
>I noticed that too, yet with Genius that is a little more extreme.
>Though I must admit I like these Genius programs I am a little
>surprized they didn't do better rating wise.


Hey, you never sleep? Looks like you are online all the time! :)


The Genius programs are handicaped by their branching factor. Each new ply takes
a little bit more than a new ply on other modern programs.

But they have a terrific pruning system in the last 3 or 4 plies of the search.

That means that they perform very well until, say, ply 8 or 9. After that,
almost every modern program is able to go deeper in less time, which is a
terrible handicap for Genius.

I'm sure you can check this experimentaly by comparing time to depth for Genius
and, for example, Fritz or Rebel-Tiger. Compare the time needed to complete each
new ply, and notice that the ratio (time for ply N+1)/(time for ply N) is always
higher for Genius (in average).

However I like the Genius programs a lot. Their evaluation function is one of
the best I have ever seen in term of stability. Maybe not the more accurate in
the absolute, but very homogenous. When Genius sees that it is 0.1 higher than
his opponent, it is going (generally) to keep this 0.1 until the endgame. When
it does not, it's because of lack of search depth. Genius then overlooks a
little trick, and to avoid losing more has to give back some of his positional
advantage.

In the endgame, his evaluation is simply unequaled until now. Modern programs
compensate by deeper search, but I'm not sure it's enough to equal Genius in the
endgame.

I use the term "modern programs" because I think that Genius has not seriously
evolved since several years. Probably since 1995.

Sorry for those of you who think that Genius6 and Genius6.5 are new programs.




>>>Genius 6 times are almost identical.


Strange... :)




>>>The other engines I think it is pointless exercise !.
>>>I just try one other engine MCP8 8kb hash = No solve after 18 mins.
>>>
>>>Tiger posts a + at 28 secs without going back? Genius picks 1. f6 always
>>>but not with a + score which is when it switches to Kxg2 before switching back.
>>>Trouble comes for Genius at depth 12/24 when it sees Kxg2 as better.
>>>2
>>>I'll trade yer two Genius's for a Tiger !! LOL.
>>>Thanks.
>>
>>
>>Here is the output from the Rebel-Tiger (K6-2 450MHz, 32Mb hash tables):
>>
>>
>>00:00:00.1	4.12	9	60006	a4  bxa3ep  bxa3  Kg3  a4  h5  f6  gxf6  a5
>>00:00:00.4	1.12	9	83505	a4  bxa3ep  bxa3  Kg3  c6  dxc6  f6  gxf6  a4  c5
>>00:00:00.3	-0.44	10	123820	a4  bxa3ep  bxa3  Kg3  c6  dxc6  dxc6  Kf4  a4  Ke4
>>Kxg2  h5
>>00:00:01.8	-1.33	11	213308	f6
>>00:00:01.5	-1.16	11	263647	f6  gxf6  Kxg2  Kg4  a4  bxa3ep  bxa3  Kf4  a4  Ke4
>>a5  Kxd5
>>00:00:02.1	-1.15	11	356087	c6
>>00:00:02.4	-1.14	11	445570	d6
>>00:00:03.7	-0.28	11	478497	d6  cxd6  Kxg2  Kg4  a4  bxa3ep  bxa3  Kxf5  a4  Ke5
>>a5  Kd5
>>00:00:04.7	-1.17	12	611167	f6
>>00:00:04.1	-0.20	12	699197	f6  gxf6  Kxg2  Kg4  a4  bxa3ep  bxa3  Kf4  a4  Ke4
>>d6  cxd6  a5
>>00:020:05.7	-0.19	12	766013	Kxg2
>>00:00:05.3	0.62	12	812882	Kxg2
>>00:00:06.4	4.16	12	1006462	Kxg2  Kg5  a4  bxa3ep  bxa3  Kf6  d6  cxd6  a4  dxc5
>>a5  c4  Kf3
>>00:00:06.0	4.16	13	1006467	Kxg2  Kg5
>>00:00:10.5	1.44	14	1609096	Kxg2  Kg5  a4  bxa3ep  bxa3  Kf6  a4  Ke7  a5  Kd8
>>a6  Kc8  Kh3  h6  Kg3
>>00:00:16.5	1.21	15	2565155	Kxg2  Kg5  a4  bxa3ep  bxa3  Kf6  a4  Ke7  a5  Kd8
>>c6  dxc6  dxc6  Kc8
>>00:00:27.9	1.22	15	4300739	f6
>>00:00:37.3	2.34	15	6108693	f6
>>00:00:43.0	2.33	15	6946271	f6  gxf6  Kxg2  Kg4  a4  bxa3ep  bxa3  Kf4  a4  Ke4
>>d6  cxd6  c6  dxc6  a5  Kd5
>>00:01:14.9	3.23	16	12181241	f6
>>00:01:56.7	5.33	16	19201142	f6  gxf6  Kxg2  Kg4  a4  bxa3ep  bxa3  Kf4  a4  Ke4
>>d6  cxd6  c6  dxc6  a5  c5  a6
>>00:02:29.3	5.33	17	24464726	f6  gxf6  Kxg2  Kg4
>>00:042:22.5	5.31	18	43205717	f6  gxf6  Kxg2  Kg5  a4  bxa3ep  bxa3  Kf4  a4  Ke4
>>d6  cxd6  c6  dxc6  a5  c5
>>00:09:45.2	5.00	19	96545193	f6  gxf6  Kxg2  Kg5  a4  bxa3ep  bxa3  Kf4  a4  Ke5
>>d6  cxd6  c6  dxc6  a5  c5  a6  c4  a7  c3
>>00:20:04.9	5.00	19	196144774	f6  gxf6  Kxg2  Kg5  a4  bxa3ep  bxa3  Kf4  a4  Ke5
>> d6  cxd6  c6  dxc6  a5  c5  a6  c4  a7  c3
>>
>>
>>I know this is not easy to read, but first column is time, second is score,
>>third is ply-depth, fourth is number of positions evaluated, and then you have
>>the best line.
>>
>>I have stopped the program exactly after 20 minutes, 4 seconds, 9 tenths. The
>>last line just repeats the last best line, as you see.
>>
>>So Tiger would play f6 very quickly, without knowing exactly why, then it would
>>play Kxg2, then sees that this is not as good as expected and would eventually
>>play f6 in about 28 seconds without changing his mind in 20 minutes.
>>
>>
>No it is clear enough 27.9 secs, I wouldn't have believed it would be that
>fast. So you got to 15 ply in only 4,300,000 positions, funny CM didn't see it I
>had that at 43,000,000.


I'm a little bit surprised too...

I think ply depth 15 does not mean the same thing in Tiger and in CM. Probably
CM does less pruning and sees more things at the same depth because it does much
more extensions? Just a guess, I don't have CM6000.



    Christophe



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.