gnugo-devel
[Top][All Lists]
Advanced

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

RE: [gnugo-devel] Improved stabilisation in get_next_move_from_li st


From: Portela Fernand
Subject: RE: [gnugo-devel] Improved stabilisation in get_next_move_from_li st
Date: Fri, 25 Oct 2002 20:39:44 +0200

Teun wrote:

> > The problem I see with your patch, Teun, is that it gives up 2)

> Arend, I don't get this point. Why does comparing bdist(move1) <
> bdist(move2) instead of comparing move1 < move2 give this up? 
> The only difference I see is that bdist maps some moves on the
> same distance.

Then it will generate noise. No way around this (yet).

> This can either be a bug or a fundamental property of the owl code.

I think it is (as of now, see below)

> Reaching 1) can only be achieved in steps, such as the one I propose
> in my patch. In the same way the orientation depencendies in reading.tst
> have been removed step by step. In the end btw these all turned out to
> be due to bugs.

The differences between reading.c and owl.c is mainly the use of patterns
associated to various cut-offs and depth limits.

The most relevant limits in regard to orientation independency are
MAX_MOVES and owl_node_limit. Whichever way we choose to sort the list,
if 2 moves are generated by the same pattern and unfortunately are 3rd
and 4th in the list, and unfortunately one of them is the solution and
and the other is not, then we have a problem. Also, let's imagine these
moves come at 1st and 2nd places, but one of them needs 1400 nodes and
doesn't succeed while the other one only needs 200 nodes to WIN. So,
you need to do something about this limit too.

My point is following: if we are to address orientation independency,
then we have to work first on these problematic limits. Because then,
whichever sorting algorythm we will be using becomes automatically
pointless.

In the meantime, I prefer to have 100% noiseless regression results,
even with an ugly and unsatisfactory (100% agreed), but known and
documented trick.

/nando




reply via email to

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