Re: [gnugo-devel] more strategic defense moves

[Arend Bayer]

Gunnar wrote:
> 
> I've had another thought on the weakness problem. Currently a dragon
> can be declared very strong, i.e. weakness 0, either by having many
> eyes, much surrounding moyo, or high escape value. It seems to me that
> a dragon which only has a high escape value but is short on eyes and
> moyo very often is indeed quite weak and risks being isolated and cut
> off a few moves later.

Speaking of escape measurements in general, I am still working on a
patch to completely revise it. I have posted a current version on the
web, linked from patches.html.

What does it do differently:
1. It makes a separate escape influence computation for each dragon, in 
which the stones of the dragon in question have no influence strength at
all. This makes a couple of heuristics avoiding to "escape inwards"
differently.

2. It uses a continuos escape value estimation instead of the current
territory/moyo/area heuristic. It's a bit tough to get this tuned right,
but I hope this will in the end make the escape estimation considerably
more precise.

3. The other dragons get awarded a strength that depends on their own
weakness (computed in this case only from moyo size and eye space). This
prevents overoptimistic escape measurement when the only thing we can
reach is another eyeless dragon.

4. Instead of the current algorithm in dragon_escape(), it reuses the
distance computations from readconnect.c (with very small
modifications). This is not much of a benefit yet, but I hope to reuse
it later to generate escape moves in owl reading. The advantage is that
this method can also exhibit fields which we cannot reach yet, but which
we _could_ reach. I.e. towards which it is worth making an escape try.

5. There is a heuristic that should help reduce the problem Gunnar
mentions above: Basically, if a dragon only gets escape influence from
one direction (via one liberty), then it does not get awarded any escape
value. (See use of "highest_contribution" in dragon_escape().)

Unlike my patch posted quite some time ago it does not touch the owl
escape move generation (yet).

It currently has s.th. like 85 PASS and 85 FAIL...

Arend