gnugo-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [gnugo-devel] Opponent not passing


From: Trevor Morris
Subject: Re: [gnugo-devel] Opponent not passing
Date: Sun, 03 Feb 2002 15:32:08 -0500

At 02:29 AM 2/3/2002 +0100, Arend Bayer wrote:
>
>> > Certainly the choice of move that is generated by owl_prevent_threat
>> > has to be improved as well, see 13x13:73 for a very good example, where
>> > GNU Go originally played N12. (Note that this would be a situation where
>> > --owl-threats could really be useful.)
>> 
>> Any ideas for improving the move found?
>
>A vague idea, yes. What is usually most urgent with regards to dead dragons
>is to make safe connections on the outside, so that it cannot break through.
>So whenever a dead opponent dragon with aji has more than one neighbor of
>our color, we should try to find connecting moves between these.
>
>Weak connection (which I would define as "can be cut, but only at the
>cost of creating a new dragon") should be looked for, too.
>
>On a similar note, possible cuts between hostile neighbors should probably
>give some small safety bonus for weak dragons. That may be implemented
>once someone writes a better strategic effect evaluation.

I have some vague ideas on this front as well.

The owl code includes a number of different types of moves:
Owl attack moves:
 1) prevent escape
 2) trying to reduce/consolidate eye space
Owl defend moves: 
 1) escape
 2) make 2 eyes
 3) expand eye space

I agree with Arend that escaping is often the best owl or owl-threat
move, and I've tuned in this direction a lot.  I'd started tuning
the other way few months back, on the theory that making two eyes
fast would be the best way to reduce variations.  However, this
resulted in a lots of poor choices in games, where jumping out was
much preferred.

An abortive effort in the direction is still in the code in owl.c:
    /* If not too deep, evaluate shape moves (like escaping) first. */
    /* This experiment not too successful - but may return to it; tm */

I keep wondering if it may make sense to categorize the owl_*pat 
patterns as eye-attack/defend vs. escape-attack/defend patterns
(or in some cases both).  Then, for example, if the opponent has
just played an escape move, give blocking moves higher priority.
Likewise, with eye-making eye-destruction moves.






reply via email to

[Prev in Thread] Current Thread [Next in Thread]