[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] How to teach gnugo wonderful tesujis?
From: |
Arend Bayer |
Subject: |
Re: [gnugo-devel] How to teach gnugo wonderful tesujis? |
Date: |
Sun, 4 May 2003 10:41:22 +0200 (CEST) |
Evan wrote:
> On Sun, 4 May 2003, Arend Bayer wrote:
>
> > Hmm. This is might not a majority opinion on this list, but I don't like
> > pattern driven approaches too much if there is an alternative.
>
> I tend to agree, but I think patterns have two big advantages:
>
> They're fast, as compared to complex algorithms. At least, they are if
> the reading constraints are simple.
I don't see a difference here in general. The move generation in
reading.c seems faster than a pattern based approach.
(Trevor's experimental reading tries to outweigh this additional cost by
trying to use much fewer nodes than the standard reading.)
> They're easy to understand. If a pattern based thing like the owl reader
> isn't seeing a move, the solution is simple and obvious: add a pattern,
> or change an existing one. If something else isn't seeing a move, it is
> harder to see the solution. Personally, I think ease of tuning and
> debugging is important.
Yes. But if you fix or add a pattern, you will modify gg's play in
typically s.th. like one out of 100 or 1000 games (yeah, ok, I am
exaggerating). Fixing an algorithm can do _much_ more than that.
I don't want to make a religious issue out of this. But I do get
skeptical with every hand-crafted pattern database larger than 100
patterns. Or a new database where it is obvious that 100 patterns would
not suffice at all.
Arend