Author: Christophe Theron
Date: 10:16:29 04/04/01
Go up one level in this thread
On April 04, 2001 at 10:17:02, Fernando Villegas wrote:
>Hi all:
>I have completed, until now, two games againts the beast, both 40 moves in 90
>minutes. And both I lost after tough fight. So what: I expected that and in fact
>I expect that againts almost every top program. What I go for is to get fun,
>FUN.
>
>Well, looking at the computer as he thought answers to my moves I realized some
>curious facts. First, the incredible great jumps in his evaluation from one
>second to another, from one ply to the next. I am talking of going from, say,
>+.90 to 0.00. My guess is that behind that there is not just a monolitic
>evaluation package of code, but a system of specific modules doing the job
>according to circunstances. I have seen changes in every program when going to
>deeper plies, but this changes by Gambit are a lot more deep and happens very
>fast.
This is because Gambit Tiger has a big "second order" evaluation function.
"Second order" means that it pays attention to the mutual geographic
relationships between the pieces.
"First order" is when you just pay attention to the position (square) of a piece
to decide what it is worth. First order is saying: "a knight at e5 is worth
+0.20", disregarding the position of any other piece on the board.
Many programs use heavily first order, because it is very simple to compute the
value of a position with this, as you just have to built tables and then update
the global score when you move a piece.
First order looks like a stupid thing to do in chess for a human player, but
believe it or not it has been working amazingly well in computer chess for
years.
Second order is much more expensive because in every position you need to look
at all the pieces and all the relationships between them.
Second order is also less stable than first order. There is almost never a big
surprise in first order. Whatever the sequence of move, if you play Nf3-e5, the
global score of the position is always going to change by the same amount (take
the global score of the position, remove the value of knight in f3, and add the
value of knight in e5 and you get the new global score).
When you do second order, a lot of unexpected score changes happens. Because in
one position putting the knight in e5 does not do anything (because for example
the black king has castled queenside and is well defended), but in another
position knight to e5 starts a storm (because the black has castled short and is
now under attack by several pieces).
This is why Gambit Tiger has huge score changes from one ply depth to another.
Because in just one additional move (ply) maybe you can bring an additional
piece to attack the opponent's king. And if there were already 2 attackers, now
that makes 3 and the attack is now a serious one. And at the next move (ply) if
the attacked side can bring another defender, then the attack is now looking
less serious.
Many programs give very modest scores for king attacks, so they will in the same
positions display score changes of +/-0.30. You will not notice these changes
because they happen all the time even in more quiet positions.
But Gambit Tiger evaluates a serious king attack much more than a doubled pawn.
So in the case of Gambit the score changes will be in the range +/-1.00, or even
more.
These changes are also the reason why I believe that Gambit Tiger needs a little
bit more depth than Chess Tiger to achieve its full strength. At very shallow
ply depths, there is too much uncertainty for Gambit. It needs more depth to
find stable king attack plans.
>The second thing that attracted my atention was its use of time. In one of the
>games Gambit was thinking more than 15 minutes, probably a lot more -I did not
>take the time accurately; I was, as always, doing three things at the same time,
>including supping a glass of red wine and lessoning Frank Sinatra- and the
>position I produced with my move was somewhat intriguing BUT not that
>intriguing. It was as if an altoguether different code was putted in action to
>reasses the position from scratch.
The program does not start again from scratch. It's just because the time
management algorithm has been changed.
If the program sees that the previously used plan is not going to work anymore,
it takes its time to find another one. It can also happen if the program sees
that it is going to lose material. It takes its time to find the best way to
give up the material.
Christophe
This page took 0.01 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.