Author: Christophe Theron
Date: 11:46:22 08/11/00
Go up one level in this thread
On August 10, 2000 at 11:06:48, Robert Hyatt wrote:
>On August 09, 2000 at 16:42:26, Andrew Williams wrote:
>
>>On August 09, 2000 at 16:01:49, Christophe Theron wrote:
>>
>>>On August 09, 2000 at 05:41:22, Andrew Williams wrote:
>>>
>>>>On August 08, 2000 at 15:56:04, Christophe Theron wrote:
>>>>
>>>>>On August 06, 2000 at 16:36:15, Vincent Diepeveen wrote:
>>>>>
>>>>>>Show me an MTD program that uses less nodes a ply as DIEP does.
>>>>>>
>>>>>>What diep is doing is very simple in search:
>>>>>>
>>>>>> PVS (starting with -infinite)
>>>>>> check extensions
>>>>>> checks in qsearch
>>>>>> nullmove R=3
>>>>>> no other crap. no pruning. Perhaps at WMCC i prune a bit,
>>>>>> but that's because against computers playing is different.
>>>>>>
>>>>>> Yet i'm missing programs using less nodes a ply with MTD.
>>>>>> I"m missing *any* deep searching program that uses MTD actually.
>>>>>
>>>>>
>>>>>Anmon, a french chess program, uses MTD(f). It is a strong program.
>>>>>
>>>>>If you are not pruning in the tree, then MTD(f) should be better for you. I
>>>>>don't use MTD(f) because I use the value of alpha and beta to prune in the tree,
>>>>>and with MTD(f) this kind of pruning makes the search really unstable (you get a
>>>>>fail-high, and when you re-search with a higher window you get a fail-low,
>>>>>oops).
>>>>>
>>>>>
>>>>>
>>>>> Christophe
>>>>
>>>>
>>>>What you can do in these cases is to use the external bounds instead of
>>>>alpha and beta. By "external bounds" I mean the bounds that have been
>>>>established in the mtdf() loop which is driving the alphab-beta search.
>>>>
>>>>Cheers
>>>>
>>>>Andrew
>>>
>>>
>>>Thanks for the idea.
>>>
>>>
>>> Christophe
>>
>>I'd like to claim credit for it, but I got it from a post Don Dailey made here
>>ages ago.
>>
>>Andrew
>
>
>I think that solves the 'lazy evaluation' problem. But I am not sure it is a
>cure-all for pruning based on alpha and beta...
As I understand it is like using an artificial alpha and beta. They would be set
a little bit the way you set alpha and beta for aspiration search, and they are
not changed during the search (at least they stay unchanged for several
re-searches).
This should adds stability and the behaviour should be close to what you get
with standard alphabeta and aspiration search.
I have not tried yet, but that's how I understand it.
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.