Author: Christophe Theron
Date: 10:31:04 10/25/98
Go up one level in this thread
On October 25, 1998 at 11:48:13, John Timm wrote:
> In general, when Fritz and Hiarcs are set to display any given number of
> "best" moves, evaluations and variations, each program will, on average,
>increase the length (depth) of the displayed variation with each increase in ply
>search depth, but, on some occasions, not increase, or even decrease, the length
>of one or more displayed variations when displaying the result of the latest
>search. In other words, if, in a given position, a display shows a variation
>six ply deep at a search depth of ten ply, the display at search depth eleven is
>likely to show six or seven ply, but may show five (or even one!) ply.
>
> I
>don't know why this occurs, but a reasonable guess is that the programs tend to
>stop the displayed analysis at a point where two (or more?) continuations result
>in approximately the same (how close?) evaluations.
>
> Is this interpretation
>correct? If not, why do programs sometimes decrease the length of displayed
>"best" variations at increased search depths? Does anyone know the actual
>parameters that are used (for example, the analysis is stopped at a given point
>if there are three or more continuations which evaluations within .10 of each
>other)?
>
> If my guess is correct then the
>occasional "early" truncation of the analysis seems to be a useful feature. In
>effect, the displayed analysis "reveals" a very human-like thought ("there are a
>number of roughly equal continuations here, so I should try to evaluate
>precisely this position rather than waste time for little gain by analysing
>deeper").
None of your guesses are correct.
It is very hard to explain why the programs truncate their lines without
explaining in details how alpha-beta, null move and hash tables works.
Tiger has exactly the same behaviour, which is mainly caused by the use of PVS
and hash tables concurrently.
The programs search as deep as they say (10 plies when they are in the 10th
iteration), and the fact that the lines are truncated has nothing to do with a
"human-like thought".
You should not deduce anything from the length of the best variation.
I give you a simple example: some programs just try to demonstrate that a move
is the best by showing that no opponent move can refute this best move. Once
they have done this, they have a main line of length 1 (the best move), but they
have not computed which of the opponent move is the best (they have just seen
that no opponent move refutes this best move). Such a program can, more than
often, stops its computation and show a 1 ply length best line.
You should not deduce from this that it has not computed deeper. It is just a
side effect of a very "inhuman-like" way of thinking (choose a move because it
cannot be refuted without even trying to see what is the opponent best reply).
In this case, even the score of the best line can be wrong. It does not matter.
The only real goal of a chess program is to find the best move. The rest of the
line and the exact score does not matter much.
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.