[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [gnugo-devel] Floating point arithmetics
From: |
Portela Fernando |
Subject: |
RE: [gnugo-devel] Floating point arithmetics |
Date: |
Tue, 21 Sep 2004 15:54:11 +0200 |
A belated regression breakage analysis for my latest patches.
* break_in:100
Fairly simple. nando_7_1.1 causing an increase in connection nodes,
the break-in limit of 400 is reached, causing a failure. nando_7_1.2
gets it back under the bar and the test passes again.
* nicklas:2103
The connection reading is actually improved by nando_7_1.2 (the
engine "sees" more connection pairs), causing an overvaluation
of various moves in the area around the dead white strings.
Though, I think the failure isn't the fault of valuation or
connection reading. Among others, I found in the traces :
C6 : (dragon C6) white string of size 10 (10.17), (...)
(...)
- defendable at B5, defend code = 4
It seems to me that the tactical reading code could be able to find
out that this worm is actually defenseless. So I'd propose to add
a test in reading.tst :
#See also nicklas2:2103
loadsgf games/nicklas/nicklas21.sgf 74
202 defend C6
#? [0]
* nngs:1260
Bad. The connection reading for B16/C13 is worse and gets accidentally
the right conclusion.
* nngs:1280
Good. The weird territorial overvaluation of W H16 disappears (break-in
code)
* connect:70
Good, reading improved (new candidate move found)
* global:1
Accidental, break-in reading still fuzzy.
-- nando