Ticket #158 (closed defect: fixed)
GNU Go doesn't get a basic endgame tesuji right
| Reported by: | rillig | Owned by: | gnugo |
|---|---|---|---|
| Priority: | normal | Milestone: | 3.7.11 |
| Component: | source | Version: | |
| Severity: | normal | Keywords: | |
| Cc: | patch: | yes |
Description (last modified by rillig) (diff)
loadsgf tesuji1.sgf
1 reg_genmove black
#? [J2]
Attachments
Regression Results
| Attachment | Rev. | PASS | FAIL | Nodes | Status | |
| endgame_7_11.2.diff | never tested | |||||
| endgame_7_11.2a.diff | never tested |
Change History
Changed 5 years ago by rillig
-
attachment
tesuji1.sgf
added
Changed 5 years ago by nando
-
attachment
endgame_7_11.2.diff
added
endgame tuning (introducing the notion of "known safe move")
comment:2 Changed 5 years ago by nando
- patch set
- Milestone changed from 3.8 to 3.7.11
Hi, the proposed patch introduces the notion of "known safe move" as discussed in endgame:980. Impact on performance is completely negligible (just a few more nodes over the whole regression suite) and results in the single (and expected) PASS in endgame:980
The newly proposed pattern EE201b would solve the case presented in this ticket alone, without that "known safe move" mechanism, but I wanted to kill two birds in a single shot and make the engine understand positions like http://senseis.xmp.net/?YoseTesuji
comment:3 Changed 5 years ago by gunnar
- Status changed from new to closed
- Resolution set to fixed
The patch looks fine and has been added to the repository. I also included the proposed testcase after modifying it a bit so that there is an alternative 2 point gote move and balanced it so that the right move is the only move winning the game:


















































































The situation in which J2 is better than J3.