Author: Vincent Diepeveen
Date: 11:46:25 02/04/01
Go up one level in this thread
On February 03, 2001 at 05:53:20, Tony Werten wrote:
>On February 02, 2001 at 21:33:10, Robert Hyatt wrote:
>
>>On February 02, 2001 at 14:42:23, Alvaro Jose Povoa Cardoso wrote:
>>
>>>On February 01, 2001 at 18:37:48, Robert Hyatt wrote:
>>>
>>>>On February 01, 2001 at 16:27:29, Tony Werten wrote:
>>>>
>>>>>On February 01, 2001 at 14:19:25, Alvaro Jose Povoa Cardoso wrote:
>>>>>
>>>>>>Does any one tried to use just Upperbounds and Lowerbounds in hashing, ignoring
>>>>>>ExactScore entries?
>>>>>>My question has to do with the fact that if we use ExactScores we don´t get long
>>>>>>PVs to feed in the next iteration, even if we extend PVs from hash.
>>>>>>My testing shows that if we do not use ExactScore entries in the hash table, we
>>>>>>end up with very long PVs that can be fed into the next iteration in order to
>>>>>>aproximate the minimal tree. After all, ExactScores are rare compared to
>>>>>>Upperbounds and Lowerbounds.
>>>>>>
>>>>>>Can someone comment on this?
>>>>>
>>>>>One of us is missing something. The PV is always exact. If you don't store
>>>>>exact, you don't have a pv. I don't know what you get from the hashtable, but it
>>>>>isn't a PV.
>>>>>
>>>>>I take my PV from hashtable and I (almost) always get a long pv, at least as
>>>>>long as the search depth.
>>>>>
>>>>>Tony
>>>>>
>>>>>>
>>>>>>Thank you
>>>>>>Alvaro Cardoso
>>>>
>>>>
>>>>He is talking (I think) about a PV that is cut short when you get an EXACT
>>>>hash hit. IE in crafty, where a PV ends with <HT> which sometimes happens
>>>>at ply=2/3/4 and results in a very short PV. I use internal iterative deepening
>>>>to help search around this the next iteration, however..
>>>
>>>
>>>That is exactly what I was talking about. When I get an EXACT hash hit the PV is
>>>cut short, and extending it from the hash table doesn't help much, I just get
>>>1-3 moves from the hash, perhaps there is something wrong with my program.
>>>Also, and in response to Dr. Robert Hyatt, I can´t execute the assignment
>>>because my program is the portuguese version of checkers. I hope you don't mind
>>>if I ask questions not related to the game of chess.
>>>
>>>Thank you
>>>Alvaro Cardoso
>>
>>That is a perfectly normal problem. My PV is often shorter than it should
>>be due to an EXACT hash hit. Nothing you can do about it, not that there
>>is anything wrong with it.
>
>I still don't get how you can have a short pv if you have a exact hashhit.
>
>Maybe run a test where you add in your pv, the bounds it got. My guess is you're
>last move is a BELOW_ALFA hit. ( That was a problem I had before )
>It was caused because I raised alfa after nullmove, sometimes making it the best
>score and having a bogus best_move which couldn't be found in hashtable.
Improving alpha with score of nullmove is pretty dubious
depending upon what values you search with the effect will
be more dubious:
If your nullmove searches with [beta-1,beta] then it's asking
for big trouble.
If you search with [alfa,beta] then the number of cases where it
will be dubious will be way less, but of course you might need
more nodes to nullmove then!
In the first case
you're going to get exact scores which might be higher
as the real exact score in that position. especially
scores of beta-1 might hurt bigtime as if you get those
The second case with [alfa,beta] is less painful as it doesn't
screw the PV up too bigtime, but instead will mess up the hashtables
slowly.
In both cases you will however suffer from the kind of position where
the reduction of the nullmove achieves to delay a kind of big threat
which otherwise gets picked up as you are forced to try moves!
Those positions are not exactly in testsets but rather are more
common in games, where in difficult positions you can do things to
delay problems!
>cheers,
>
>Tony
>
>>
>>And you can _still_ turn in your assignment on time. Checkers is perfectly fine
>>to test your hypothesis. :)
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.