Author: Albert Silver
Date: 13:18:24 07/14/00
Go up one level in this thread
On July 14, 2000 at 11:45:20, Christophe Theron wrote:
>On July 13, 2000 at 17:44:25, Dann Corbit wrote:
>
>>On July 13, 2000 at 17:02:47, Christophe Theron wrote:
>>[snip]
>>>I understand your point.
>>>
>>>My feeling is that in the current state of the art, we are too much ignorant to
>>>decide so easily what is knowledge (or chess wisdom) and what is not.
>>>
>>>Most people, when they use the words "chess knowledge", are thinking about
>>>"position evaluation". I find this very short minded (I'm not saying you think
>>>this way, of course).
>>>
>>>Position evaluation is not the only part of a chess program where you put "chess
>>>knowledge". A lot of chess knowledge is put somewhere else in a chess program.
>>>
>>>There is a LOT of chess knowledge in the search algorithms. The concept of
>>>recapture extension, for example, is pure chess knowledge. Move ordering is, as
>>>well.
>>
>>Actually, I am forced to agree with you here. We might think also of a GM
>>looking at the board. He is also performing a search, since he will consider
>>several possibilities and choose the best one. He will have specific move
>>ordering when he considers a particular move. I think that a big difference is
>>that human players decide on a goal and then think about moves that aim towards
>>achieving that goal.
>
>
>
>That's right. When I try to use the knowledge of strong chess players (well at
>least much stronger than myself), I always find this problem: most of the time
>they evaluate MOVES not POSITIONS.
>
>Of course they also know how to evaluate positions (but they are unable to give
>numerical values to them, of course), but that's not how they "search" their
>tree. They try to build "best lines" that have as many "good moves" in them as
>possible.
>
>One interesting thing that comes from this is: if it is possible to reach a
>given position by two different paths (that's what we call a transposition),
>they will almost always prefer one path over the other one.
>
>A chess program would evaluate both lines as totally equivalent. A human player
>would not.
I don't know about that. It's true that many times a position can be transposed
to via more than one move order, all the same, the trees involved are not the
same, at least for humans. The problem is in limiting the size of the tree.
Sure, in a given position I can begin a particular line with one move order that
is more or less forcing, with a few branches to be calculated and a quiet move
or two, or I can reach it with the very same moves, but in a different order. In
one move order, some of the moves may require more calculations because of
possibilities dependent on the move order. Suppose there are three checks that
have only one or two direct possibilities, and I will also need two quiet moves
that lead to the desired position and furthermore, there are two ways of doing
this. In the first move order, I can interpose the quiet moves a bit earlier,
but this also means that I have to take into account several possibilities,
which may not work but that need to be calculated, that wouldn't be there if I
were to first play the checks. Perhaps for the computer, the calculating
necessary results in the same thing as it is more or less indifferent to many
factors such as visual impressions, but not so for the human. I have seen this
quite often in certain combinations, and I too will speak of the 'cleaner' move
order as it leaves less doubts in its process. I learned this, and I agree with
it.
This also eventually leads to the more evolved form of handling positions, where
players (human) choose the 'simpler' path. Again, this is highly debatable in
theory as the 'simpler' solution can often be objectively longer, but this
methodology is entirely human and was never designed for computers so while it
may be debatable in theory, in practice it is not.
Albert Silver
>
>That's why I think that human players are not using alphabeta actually. They are
>using something close, but evaluating paths rather than final positions is
>definitely not alphabeta.
>
>
>
>
>>>They are both very related to chess, and quite unrelated. A grandmaster could
>>>find these concepts useless maybe. But maybe a grandmaster is not a reliable
>>>source of chess knowledge.
>>>
>>>Also, very often "knowledge" is associated with "memory" (storing data). This is
>>>misleading as well. Most of the knowledge in a chess program exists in the form
>>>of algorithms, not tables, databases or files.
>>>
>>>I tend to define knowledge (actually "relevant knowledge") as the set of
>>>information processing procedures (algorithms) that help a given entity to
>>>survive or even to grow in a given environment. Maybe I'm melting several
>>>concepts together in this definition, but anyway these separate concepts are
>>>really hard to differenciate. So maybe it is not so useful to differenciate
>>>between them.
>>
>>I think that this is probably the most interesting area of chess program
>>development. Struggling with ideas and definitions may spring forth something
>>new.
>
>
>
>Yes, and discovering a new algorithmic idea that works actually means
>discovering something new about the nature of chess.
>
>This is obvious for example with the idea behind "null move". The fact that
>"zugzwang" is a very rare event in chess and that it allows to avoid a lot of
>search effort is part of what makes the game interesting. All human players use
>this idea without actually realizing what they are doing and why.
>
>This property does not exist in other games such as checkers and
>Othello/Reversi. This is as important as the fact that the chessboard is an 8x8
>square. But it is less obvious, and not stated explicitely in the rules of
>chess.
>
>
>
> 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.