Re: [gnugo-devel] experimental code

[Gunnar Farneback]

I wrote:
> But even disregarding effects on speed, it's still true that the
> experimental connection code is not ready for prime time. It improves
> some aspects of the engine but also introduces a couple of serious new
> problems that need to be fixed. 

The regression delta is pretty large, 55 unexpected passes for
owl: 18, 47
connection: 1, 5, 6, 15, 17, 65, 70, 75, 78, 83, 84, 85, 86, 90
trevora: 130, 570
viking: 3
trevorb: 220, 230, 360, 650, 900
strategy2: 51, 80, 91
tactics: 5
nngs: 380, 420, 490, 630, 1280
trevorc: 350, 600, 820, 860, 870, 950, 1200, 1510, 1540
strategy3: 122
global: 10
vie: 50
13x13: 43
trevord: 180, 360, 570, 670
handtalk: 13
nngs2: 40, 100
century2002: 35
auto02: 6

and 23 unexpected failures for
ld_owl: 302
optics: 1801
connection: 2, 19
nngs1: 24, 39
strategy: 35
viking: 6
lazarus: 14
trevorb: 670
strategy2: 70
nicklas1: 811, 2301
nngs: 690, 1070, 1260
trevorc: 1150
strategy3: 101, 111
global: 9
13x13: 7, 67
trevord: 260

Strictly looking at the regressions things look fairly good, but as
usual these only cover earlier noticed weaknesses - not newly
introduced ones.

The main new weakness is caused by the experimental connection code
doing away with most of the B patterns in conn.db. This is not a bad
thing by itself but has the side effect of affecting the segmentation
of moyo regions (see the use of INHIBIT_CONNECTION in
whose_moyo_restricted()). In practice this means that the surrounding
moyo size may become severely overestimated in certain positions,
causing critical or possibly even dead dragons to be considered
trivially alive without any need for owl reading. This is obviously
very bad and needs to be solved somehow.

/Gunnar