Author: Albert Silver
Date: 05:57:31 11/13/05
Go up one level in this thread
On November 12, 2005 at 11:48:36, Dieter Buerssner wrote:
>On November 12, 2005 at 09:20:07, Albert Silver wrote:
Food for thought. Thanks for the detailed reply.
Albert
>
>>Over at CSS, the latest results of CEGT were shown with both Fruit 2.2 and Fruit
>>2.2.1 having achieved the exact same rating. The poster concluded this reflected
>>the myth that TBs added strength, whereas the ratings proved otherwise. I view
>>it differently of course. What *can* be concluded is that the speed lost
>>searching the TBs on the HD is compensated by the knowledge brought.
>
>I think, this cannot be concluded. Robert Allgeuer has reported some interesting
>tests some time ago, where the results are similar. However, he did not see much
>slow down due to TBs in game phases, where the game was not already decided.
>
>I believe, things are more complicated here. I suspect, that TB accesses change
>the search tree in some unfavorable way. Some random observations from selected
>positions (out of my memory):
>
>- I have seen positions, where the result was a win for one side - with TBs the
>engine showed draw score for larger depths, while without TBs the score
>continually increased
>
>- I had tried "fail hard" TB scores vs. "fail soft" TB scores. Fail hard means:
>return beta or alpha in case where the TB score is out of the search window.
>Fail soft means: return result from the TB. To my surprise, fail hard often was
>faster. Hard to explain.
>
>Of course, there are many positions, where the slow down due to the TB-accesses
>makes the engine slower to find the best move. But because of Robert Allgeuers
>results, I don't believe this is the main reason to explain the results, that
>have been reported.
>
>The testing environment of the results of Helmut Conrady in CSS some time ago
>was also not described in all detail. For example, it is not known (to me - I
>had asked at the time) in which order the tests had been done, whether learning
>was cleared between the tests, and I seem to remember vaguely some more issues.
>It is clear, that TBs should always help for not too short time control. Only
>using them at the root can never hurt (just recently Yace lost a game vs. Crafty
>in KRPPKR, which was a TB draw, Crafty used this TB, Yace didn't). One can tune
>the engine in a way, that there is practically no slow down, by only accessing
>TBs very close to the root position.
>
>BTW. There are also many (?) (study like) position, where one can find, that
>using TBs extremely aggressively (during the whole search) will help to find the
>solution much faster. Even when nodes/s goes down to perhaps 5% of the usual
>value.
>
>Regards,
>Dieter
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.