Gifu testcases

[gunnar]

The Gifu 2005 tournament games is a good source for new tests. The games can be found in source:regression/games/gifu2005.

Most important test case targets are the two losses against Katsunari and KCC. Also interesting are the games against Many Faces of Go, Go Intellect, Aya, and GoStar. The games against Daigosan (everything died!), Nanashi, and Igoppi are of no or very limited interest.

These should go into a new gifu2005.tst file, but we can start collecting testcases in the comments to this ticket.

[arend]

Testcases for the Katsunari game.

[gunnar]

Two more suggestions for the Katsunari game:

• Move 216: Can black live in the corner if white cuts at D15?

• Move 224: S18 makes a 7 points sente seki, doesn't it? It should maybe have been played even earlier.

[arend]

(Answers and more suggestions.)

• Move 216: I don't think White can cut, Black could capture the cutting stones with A16.
• Move 224: Agree.
• Move 238: Might be worth a testcase to keep this move.
• Move 252: This should be at T14! Tricky semeai combination. Too tough right now anyway?

[gunnar]

You're right about move 216. I just imagined that A16 was ineffective.

Move 238 is definitely worth a testcase. Currently it's way undervalued at 2.29 points and only barely beats G9 at 2.13.

Why T14 at 252? S18 can still be played and I don't see much value in T14 anyway. Am I missing something?

[arend]

Incorporates discussions on katsunari game below. (Checked-in version).

[gunnar]

# Two points in sente.
loadsgf games/gifu2005/gnugo-katsu.sgf 238
1140 reg_genmove white
#? [G8]

Isn't this three points in sente? But it needs to be a restricted_genmove test since S18 still is the biggest move on the board.

loadsgf games/gifu2005/gnugo-katsu.sgf 254
1150 reg_genmove white
#? [T14]

Um, I'm afraid S18 remains the bigger move even now.

[arend]

Fixes 1140, 1150

[arend]

Fixes 1140, 1150

[gunnar]

Thoughts on the Many Faces game:

• Move 25: Q11 mandatory?
• Move 81: It looks important to connect at S15.
• Move 195: No point in playing B18. The problem is that the tactical reading finds an imaginary attack on C19.
• Move 204: This is Many Faces' move but GNU Go also misses the importance of E17, so this is worth regressing over.
• Move 205: Same thing for the other color. GNU Go happened to play the move but without understanding the full implications, so it's also worth a testcase.
• Move 259: M8 should be at M6.

[gunnar]

Testcases for the Many Faces game.

[gunnar]

Testcases for the Many Faces game have been attached in gunnar_7_11.1. After that has been added it's pointless to continue looking at these games; they are too old now.