Author: Peter Fendrich
Date: 08:12:50 04/06/04
Go up one level in this thread
On April 06, 2004 at 08:15:53, Tord Romstad wrote:
>On April 06, 2004 at 07:24:56, Peter Fendrich wrote:
>
>>On April 06, 2004 at 05:18:00, Tord Romstad wrote:
>>
>>>On April 05, 2004 at 18:58:57, Andrew Wagner wrote:
>>>
>>>>On April 05, 2004 at 18:42:57, rasjid chan wrote:
>>>>
>>>>>On April 05, 2004 at 15:59:40, Dann Corbit wrote:
>>>>>
>>>>>What fruits! I can't yet digest the apple.
>>>>>
>>>>>On a more serious note, it seems there MAY BE much more in hashing
>>>>>than what I know - UB, LB, EX. I need time to see what all these mean.
>>>>
>>>>UB = Upper bound, LB = Lower bound, EX = exact.
>>>>
>>>>When you store a value in the hash table, sometimes it will not be exact, so you
>>>>store some flag along with it that says what kind of position it is. If you just
>>>>failed high, all you know is that the score is at least X. If if failed low, all
>>>>you know is the score is at most X. And if the score is between alpha and beta,
>>>>it's exact.
>>>
>>>Another option is to store _two_ values in the hash table entries, an
>>>upper bound and a lower bound. You will probably also need to store two
>>>depths, one for each bound.
>>>
>>>This is of course more expensive in terms of space, but it will also give
>>>you a bigger number of hash table cutoffs. Whether it is worth the price
>>>probably depends on the engine. In my engine, two bounds work much better.
>>>
>>>Tord
>>
>>This must be some kind of MTD thing. In PVS I don't see how it would help where
>>almost every window is a null window but maybe I'm missing something...
>
>You are almost certainly right that using two bounds is much more advantageous
>in MTD than in PVS, but as far as I can see it should help in PVS, too. I
>don't understand why it is relevant that almost every window is a null window
>(in an MTD search, of course, *all* windows are null windows).
Yes you're right. The window size has not much to do with anything here but
I'm still sceptical to it's usefulness in PVS!
>This is how the code for hash table cutoffs look in my engine:
>
>/* 'he' is a pointer to a hash table entry. */
>
> if(he != NULL) {
> if(he->lower_depth >= depth && he->lower_generation == HashGeneration) {
> if(lower_bound(he) >= gamma) return lower_bound(he);
> }
> if(he->upper_depth >= depth && he->upper_generation == HashGeneration) {
> if(upper_bound(he) < gamma) return upper_bound(he);
> }
> }
>
>If my understanding of PVS and other traditional alpha beta variants is
>anywhere near correct, the code would be very similar in a PVS search
>(except that the first occurence of gamma would be replaced by beta,
>and the second by alpha). Having two bounds should increase the
>probability of a hash table cutoff.
Agree, but most interesting is how often. It must be much more often in MTD.
I just have presentiment of that the increased entry size wont pay off in PVS
with that much more cut offs.
>Under the (unrealistic) assumption that the number of hash table entries
>is the same in both cases, a search with a two-bound hash table should
>always consume fewer nodes than the same search with a one-bound hash table.
>In practice, of course, the number of hash table entries will be lower
>with two bounds, and it is hard to say whether one or two bounds is
>optimal without testing.
Yes, testing is the only way to know but arguing is part of the fun!
In this case however I don't have much more than a feeling to base my
arguments on. When the speed is x Knodes per seconds it's hard to have a
complete picture of what's going on during the search. Most variants are
really weird and would never occur in a normal persons brain...
/Peter
>It is perfectly possible that you are right, and that my understanding of
>the complexities of PVS is still too limited to enable me to understand the
>problem. I've only been a PVSer for two days, and my engine still doesn't
>have hash tables.
>Tord
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.