Author: Robert Hyatt
Date: 05:50:51 04/27/01
Go up one level in this thread
On April 26, 2001 at 23:36:55, Albert Silver wrote:
>On April 26, 2001 at 22:52:20, Robert Hyatt wrote:
>
>>On April 26, 2001 at 20:29:07, Uri Blass wrote:
>>
>>>On April 26, 2001 at 20:06:38, Robert Hyatt wrote:
>>>
>>>>On April 26, 2001 at 17:30:11, Uri Blass wrote:
>>>>
>>>>>On April 26, 2001 at 17:24:09, Robert Hyatt wrote:
>>>>>
>>>>>>On April 26, 2001 at 16:57:53, Christophe Theron wrote:
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>If it is just that, then there are ways to avoid these problems. Expensive, but
>>>>>>>less expensive than brute force.
>>>>>>>
>>>>>>>The problem I abhor is when my program gets oursearched. This by far outweigths
>>>>>>>any other minor problem. Brute force always has this problem, not just in one
>>>>>>>game out of 1000.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Christophe
>>>>>>
>>>>>>
>>>>>>Maybe that was Hsu's point. At 200M nodes per second you probably won't
>>>>>>get outsearched if you search every node twice. :)
>>>>>
>>>>>Hsu is wrong.
>>>>>
>>>>>Deeper blue made a tactical mistake in the second game against kasparov because
>>>>>it did not search deep enough.
>>>>>
>>>>>It did not see that the final position is drawn and it proves that search is
>>>>>important also at 200M nodes per second.
>>>>>
>>>>>Uri
>>>>
>>>>
>>>>By that definition every lost game makes a tactical mistake.
>>>
>>>The point is that it is not a mistake because of a zunzwang so recursive null
>>>move could help to find the right move faster.
>>>
>>>Deeper blue
>>> Since _nobody_
>>>>has shown a draw in that position with a computer, I can personally forgive
>>>>deep blue as well.
>>>
>>>I remember that Diep could see enough in order to play 44.Kh1 and not 44.Kf1 in
>>>the position some plies before the drawn position.
>>>
>>>I guess that other programs can also do the same if you give them enough time.
>>>
>>>Here is the relevant position
>>>[D]R7/1r3kp1/1qQb1p1p/1p1PpP2/1Pp1B3/2P4P/6P1/6K1 w - - 0 1
>>>
>>>Deeper blue searched 192 seconds and played 44.Kf1
>>>I am interested to know what programs can see after 192*200,000,000 nodes.
>>>
>>>Uri
>>
>>
>>Here is crafty's output on my notebook. Lots of mind changing...
>>
>>I will guarantee you that if a program plays Kh1 over Kf1 it is
>>_not_ because it is avoiding a draw. They are not going to see that
>>from here...
>>
>>The following is the best 4 moves and scores, searched for
>>60 seconds. NOthing marks the first 2 as being much different.
>>the third is close...
>>
>> ({14:+1.85} 1. Kf1 Rb8 2. Ra6 Qxc6 3. dxc6 Kf8 4. Ra7 Rc8 5. Rb
>>7 h5 6. Rxb5 Ke7 7. Ra5 Rc7 8. Ke2 h4 $18)
>> ({14:+1.88} 1. Kh2 Rb8 2. Ra6 Qxc6 3. dxc6 Kf8 4. Ra7 Rd8 5. Rb
>>7 Rc8 6. Rxb5 $18)
>> ({14:+1.54} 1. Kh1 Rb8 2. Ra6 Qxc6 3. dxc6 Kf8 4. Ra7 Rc8 5. Rb
>>7 Ra8 6. Rxb5 Ra1+ 7. Kh2 Re1 8. Rd5 Bxb4 9. cxb4 Rxe4 $18)
>> ({14:+0.39} 1. Qxb6 Rxb6 2. Ra7+ Kf8 3. Kf1 Bb8 4. Rd7 Bd6 5. K
>>e2 Be7 6. Ra7 Bd6 7. h4 Rb8 $14)
>>
>>
>>I don't want to even talk about Diep or any other program playing Kh1 or
>>not playing Kf1 until I see real PVs and scores to prove they know that Kf1
>>leads to a draw...
>
>Frankly, I don't even understand why this topic keeps coming up still. I already
>showed Vincent, and others who saw the posts, that there were many lines that
>the program considered a draw only because it hadn't seen certain potential
>threats, and only AFTER it had seen how to counter those threats could one say
>the program saw a draw. Otherwise it is merely a horizon effect. It sees a draw,
>because it _hasn't_ seen the potential to complicate. When it does, the eval
>will swing again, until it has neautralized those attempts, after which one can
>safely say the program sees a draw. So if Diep announces it's a draw nice and
>early, one can safely say it is a horizon effect, nothing more.
>
> Albert
Too many folks, some program authors included, don't understand how a draw
score gets propagated back up the tree. Which makes them assume that when a
program says "draw" then the draw is absolutely forced. This is completely
false. A program will propagate the draw score backward when it thinks one
side is ahead and the other side can force a repetition. But it may well find
later that the side it thought was ahead was not. And the draw score becomes
invalid.
The only absolutely correct scores we generate are mate in N.
This page took 0.01 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.