Author: Vincent Diepeveen
Date: 09:23:44 12/31/99
Go up one level in this thread
On December 31, 1999 at 11:03:45, Alessandro Damiani wrote:
>On December 31, 1999 at 10:30:12, Vincent Diepeveen wrote:
>
>>On December 31, 1999 at 04:51:27, Alessandro Damiani wrote:
>>
>>>Does someone have experience with Alpha-Beta-Conspiracy Search? I have read the
>>>draft of 1993 and it seems to me a very interesting scheme for selective search.
>>>
>>>Alessandro
>>
>>Ulf Lorenz (university paderborn) has made a program called: P.Conners...
>>...Parallel Conspiracy number search
>>
>>It plays pretty well. By far the best selective searching program made.
>>Nothing using entirely conspiracy or proof number search comes even close
>>to its strength.
>>
>>Vincent
>
>Alpha-Beta Conspiracy Search is an alpha-beta search. It avoids all deficiencies
>of Conspiracy-Number Search. Alpha-Beta Conspiracy Search uses conspiracy theory
>to extend the nodes, but works within alpha-beta. The memory cost is linear.
I'm not good in exact definitions. I find it all the same crap (apologies
to put it a bit rude).
Applying alfa beta or using a bound, it comes down to the same.
The really interesting thing is how you extend nodes.
>I like very much its behaviour to extend nodes near the minimax score and search
>less deep other variations. The future of selective search?
>Does someone have any experience with this interesting method?
I've kicked out all selectivity out of DIEP. It doesn't work for me.
I've experimented for 2 years with DIEP searching in depth limited way.
So at depth = n it would search at least 1/2 n depth,
other moves it would see if evaluation was happy
about extending a node. Further some tactical moves were seen further.
After measuring huge hashtable cutoff wastes because of applying
fractional depth i rewrote it just like junior to a very simple
principle (note my experiments started after junior won Paris):
- good move ==> 1 ply
- bad move ==> 2 ply
Basically i did 3 experiments
- good moves are positional moves
- good moves are tactical moves
- mix
The first approach was quickly banned after missing too much tactics.
If you search say 12 or 13 ply, but in reality see tactics up to 7
ply or something (Qsearch not counted of course), then you get really
sick of this type of search.
The second approach was very quickly banned as it played positional
too horrible.
Instead of simply stopping with the experiment i then started with
a mix experiment. I don't know why i didn't stop the experiment. I was
weak i guess. Too happy it solved many tactical testsets very quickly.
After Paderborn i drawed the conclusion that it was no good to search
selective after conceptual thinking about it.
I find it big nonsense to in general see tactics deeper and/or nodes
based upon your evaluation. It more or less assumes that your positional
understanding is enough of the game and that you only need to look
tactical a bit deeper.
Nullmove already adds more than enough selectivity to the fullwidth
search to take it serious.
It doesn't take away that continuing to experiment with such searches
is very interesting to do, and that some extensions are necessary to see
deep tactical lines. Lucky in chess when already searching around 12 ply
just an extension of 1 ply for check, a few threat extensions and afew
passed pawn extensions already covers those lengthy lines well enough.
Basically i try to get my DIEP to search a ply deeper without forward
pruning.
Why are we despite this obvious conclusion still interested in a different
form of searching?
I guess that the fact that there are strong human grandmasters answers that
desire.
How do humans actually search?
Humans basically try a few moves and extend lines they find most interesting.
The way in which they extend is however ONLY dependant upon their knowledge
or interest whether they see a pattern that fires a tactical shot or gives
them compensation.
So basically humans select a few moves based upon knowledge.
What is the big problem of mankind?
That is obvious to answer. The vaste majority miss a lot of moves which after
being told to the person he will usual qualify as: "o dear, i missed that
obvious move".
Now for computers we can draw the next conclusion if you try to search
in a selective way
- programs have a lot less knowledge than any A class human already
so they won't select the right moves to be searched
- humans have the ability to very flexible us a pattern. Sometimes
when i consider myself a single 'obvious' pattern, such a pattern then
translates into tens of patterns for the computer before it
understands it a bit. Still some positions with the pattern it will
do it wrong.
So all selective searches as implemented at the computer so far
do not follow th only form of selective search that has proven to be excellent
in the past.
>Alessandro
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.