Author: Frank Schneider
Date: 11:32:58 06/22/98
Go up one level in this thread
On June 21, 1998 at 21:23:22, Don Dailey wrote:
<snip>
>>>Has anybody tried this before? Any comments?
>>>
>>
>>
>>How does your rule apply to the Spanish ? e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 O-O Be7.
>>Seems that both white and black play illogical sequences where no move is
>>related to the other.
>>
>>I think the definition of "logical" is too simple. Possibly, with effort, it can
>>be improved, but may not work anyway because of this: Extensions and forward
>>pruning are not about following "logical" lines but about getting a reliable
>>eval for a line. I don't see an obvious reason why a "logical" line should be
>>searched more deeply while an "illogical" line less deeply. Conceivably, you can
>>stop the search of an excellent move after two plies, because the score is
>>reliable enough, while you would need to search a terrible line to 12 plies to
>>reach the conclusion that it is terrible.
>>
>>You may have discovered that depth reduction is in many cases a useful thing to
>>do, and the justification you found for it is perhaps irrelevant. Evaluating
>>such a heuristic with a test-suite is not a very sensitive or appropriate way.
>>
>>Amir
>
>
>Hi Amir,
>
>One of us did not understand the heuristic. I think he should
>explain it a little better because I am unclear about the details.
>(FRANK, if you are listening ...)
>
>But, by what I (thought) I understood then:
>
>1. e4 e5
>2. Nf3 (attacks the square black just moved to)
>2. ... Nc6 (defends the square the last move attacked)
>3. Bb5 (attacked the last piece that moved)
>4. ... a6 (attacks the last piece moved)
>5. Ba4 (moves piece just attacked)
>5. ... Nf6 (-- NO DIRECT RELATION but attacks d6 which the
> bishop attacks through the knight)
>... etc.
>
Hi Don,
your explanation of the line (e4 e5, Nf3 Nc6,...) makes more sense than the
current implementation does. Things like 'defends the square the last move
attacked' are not recognized (but could be). Here is an explanation of what
the heuristic currently does:
1. e4 e5
2. Nf3 (Pe2 is no longer defending f3 and Qd1 is now attacking f3,
therefore Nf3 is 'logical')
Nc6 (illogical)
3. Bb5 (after 2.Nf3 f1 is attacked by the rook -> OK)
a6 (no longer attacked after 2. Nc6 -> OK)
4. Ba4 (Bb5 attacked by 3. a6 -> OK)
5. Nf6 (illogical)
6. O-O (illogical, BUT castling could be considered logical in any case
and after Bb5-a4 f1 is no longer attacked by the bishop)
I admit that some moves were 'logical' for the wrong reason and that Nc6 and Nf6
should not be called illogical. However, the line Amir gave is
a) a good line. The heuristic tries to detect bad lines. The heuristic would
search Nc6 and Nf6 with reduced depth BUT would re-search them to normal depth
if the shallow search returned a value > alpha.
b) an opening line with many undeveloped pieces which just begin to 'play
together'.
It is easy to find position where the heuristic currently fails, but there are
many possible improvements/alternatives:
- improve the 'illogical-test' (see Don's ideas)
- do only test computermoves and always search opponentmoves
(kind of 'defensive' pruning)
- do something like part-ply-pruning
>
>You mentioned that the real goal was to get the most reliable
>evaluation which I agree with. I believe (in the spirit of
>what Frank is trying to do) this is quite consistant with that
>goal. Whether it works or can be made to work is an open
>question (at least to me) but it may be that your intuition
>on this is stronger. I cannot dismiss it yet without taking
>a closer look.
>
>This stuff is VERY similar to some stuff one of our team members
>is experimenting with. We get a nice node reduction for free,
>but the implementation reduces our nodes per second drastically.
>Our experiences with this gives me reason to believe something
>like what Frank is experimenting with has a chance to pay off.
Thats why I posted it. I wanted to know what more experienced people think about
it. I did not mean to post an elaborate algorithm that is ready to use.
It was very interesting to read that one of Don's teammembers is working on
similar ideas.
>
>I definitely agree with you on the test suites. My experience
>has been that they have no value at all for evaluating how good
>the chess program is, but I use them quite extensively
>because they have a lot of value for evaluating the behavior
>of various algorithms. When it comes time to really find out
>if some change really improves the program, other methods must
>be used. I have this idea that a few of the programmers use
>problem sets to evaluate their programs since it is by far the
>simplest thing to do and I believe this may be a serious mistake.
I agree that testsuites do not prove that a heuristic works, but they could show
that it doesn't. I simply had no time (just preparing my exams) to try different
methods. I used the WAC and BK positions because (so far) I don't have a good
positional testsuite (where can I find a (good, large) one?).
>
>- Don
Frank
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.