Ticket #126 (new defect)
Blunder : a bamboo joint is cut and lose the game.
| Reported by: | alain | Owned by: | gnugo |
|---|---|---|---|
| Priority: | normal | Milestone: | 3.9.x |
| Component: | regressions | Version: | |
| Severity: | major | Keywords: | |
| Cc: | patch: | no |
Description (last modified by alain) (diff)
gnugo 3.7.8 and 3.7.9 play the same terrible move 26 and let the bamboo joint being cut. This lose the game.
- gnugo 3.7.7 connects.
- Default fourlib-depth is 7. Increasing it to 8 is enough for finding the good move (for 3.7.8 and 3.7.9).
$ gnugo-3.7.7-ref -l twogtp047.sgf -L 26 white (O) move E5 $ gnugo-3.7.9-ref -l twogtp047.sgf -L 26 white (O) move J6 $ gnugo-3.7.9-ref --fourlib-depth 8 -l twogtp047.sgf -L 26 white (O) move E5
Attachments
Regression Results
| Attachment | Rev. | PASS | FAIL | Nodes | Status |
Change History
comment:1 Changed 6 years ago by alain
- Description modified (diff)
- Severity changed from critical to major
comment:2 Changed 6 years ago by alain
The twin #104 changes the move and pass the test. This also mean that the connection has no move reason SEMEAI, STRATEGIC_ATTAC or STRATEGIC_DEFEND, as in those case the twin warns and is not used.
comment:3 Changed 6 years ago by alain
Some (weird) investigations :
- fourlib depth = 0 or 1 => C8 no.
- fourlib depth = 2, 3, 4, 5, 6 => E5 good
- fourlib depth = 7 => J6 no.
- fourlib depth = 8, 9, 10, 11 => E5 good
comment:4 Changed 6 years ago by alain
reverting 2270 and 2254 does fix this, but also lose their benefit, so it is probably not a good idead ;-)
comment:5 Changed 4 years ago by josephpiche
I confirm that in current CVS (as of today), this is still an issue. However:
$ interface/gnugo -l twogtp047.sgf -L 26 white (O) move J6
$ interface/gnugo -l twogtp047.sgf -L 26 --fourlib-depth 8 white (O) move E5
$ interface/gnugo -l twogtp047.sgf -L 26 --monte-carlo white (O) move E5
The MC carlo code finds E5. Maybe it should be enabled by default on 9x9 and smaller.

