[Top][All Lists]

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

Re: [gnugo-devel] Patch: improve separation of similarly-valued moves

From: aquarius
Subject: Re: [gnugo-devel] Patch: improve separation of similarly-valued moves
Date: Thu, 15 Apr 2004 12:33:34 +0200
User-agent: Opera7.22/Win32 M2 build 3221

Two thoughts on this issue - perhaps too simple-minded, but ...

Am Wed, 14 Apr 2004 23:08:08 +0200 (MEST) schrieb Inge Wallin <address@hidden>:

Here is a very simple patch that improves the value separation of moves
suggested by Joseki patterns.  The original code used the computed value
for a move to add a small bonus to the min value suggested by patterns.
But it only did this for non-joseki moves.  This patch removes that

Why not set this bonus randomly from an appropriate range each move?
A carefully chosen range would assure that the best move (e.g. the joseki)
would be played at least some times.

Regression testing could then be done with range=0.

Am Wed, 14 Apr 2004 21:02:26 -0400 schrieb Evan Daniel <address@hidden>:

Well, personally I think we want it to play more predicatbly in cases
where there are clear right answers, and less predictably when there are
legitimate choices.  But that's a difficult problem :)

How is the current scheme deciding on playing a 'second-best move' for the
sake of randomness in play? Could'nt my suggestion be adopted to all moves?

And then make the random range a command-line parameter defaulting to little
or medium randomness. This way GNU Go would gain another aspect of "personality"
ranging from "play as good as you possibly can, but boring" to "make terrible
blunders now and then regardless of level parameter".

Regression testing could then be done with randomness 0.

Happy Go - ing


reply via email to

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