Ticket #166 (new defect)

Opened 5 years ago

Last modified 3 years ago

Break-in weirdness

Reported by: gunnar Owned by: gnugo
Priority: normal Milestone: Future
Component: source Version:
Severity: normal Keywords:
Cc: patch: no

Description

In the attached game at move 15,











C5 is the obvious move. Now GNU Go tends to prefer E2 but ignoring that move there's an interesting problem with restricted_genmove B5 C5 and how the valuation of those moves varies with level.

0-1 2-9 10 11 12-15 16-20
C5 12.26 12.08 14.08 14.08 10.08 7.08
B5 10.59 10.42 12.42 10.42 10.42 10.42

Except for a 0.03 followup value all those scores are purely territorial values. What mainly differs above level 10 is the amount of "Erasing territory at ..." induced by the breakin code. Erased points shown in the table below:

10 11 12-15 16-20
C5 H7, C6 H7, C6 H7, C6, B4, G7, H8, J8 H7, C6, B4, G7, H8, J8, H8, J8, B6, C7
B5 H7 H7, G7, H8, B4, J8 H7, G7, H8, B4, J8 H7, G7, H8, B4, J8

(H8 and J8 are repeated in the debug output for C5 at levels 16-20, don't know whether that indicates anything abnormal.)

So, what's going on here?

Attachments

465598.sgf Download (799 bytes) - added by gunnar 5 years ago.
Problematic game record.

Regression Results

Attachment Rev. PASS FAIL Nodes Status

Change History

Changed 5 years ago by gunnar

Problematic game record.

comment:1 Changed 3 years ago by gunnar

  • Milestone changed from 3.8 to Future
Note: See TracTickets for help on using tickets.