Re: [gnugo-devel] Owl-eyes and Ko

[Gunnar Farneback]

Trevor wrote:
> The following problems all indicate a basic problems with the
> Owl code and Ko.

The problem is primarily with the optics code, i.e. the individual
eyespace evaluations.

> Having rooted around in the code a bit, I feel a bit tied up in
> knots. Is there a short-term fix for these?

Presumably they can be solved with explicit owl patterns. Possibly
they need to be of the vital variety.

> Is there a preferred medium or longer term solution?

The long term solution is to revise the optics code. The current one
is starting to show its age. When it was written it was a huge step
forward in understanding of eyespaces, but there are a number of
(relatively speaking) minor shortcomings that are starting to get
important. Another thing to take into account is that when it was
written it was intended for an entirely static life and death
analysis. Now that we have the owl code, the conditions are somewhat
different.

> trevor:350, trevor:360, trevor:470, trevor:471

Dan wrote:
> The first two of these look like the same bug and yes, we are 
> supposed to get problems like this right. Maybe something is broke.

No, these show a problem we've always had. The basic problem is that
the optics code was written before any concept of ko was introduced.
In particular this goes for the topological eye analysis. In test
case 350, M9 is a half eye if the ko is won and a false eye if the ko
is lost. There are no provisions in the code to keep track of both
these possibilities. It might be possible to patch in handling of
this, but I don't think it's easy and it may be better to wait for a
more serious revision of the optics code.

/Gunnar