Re: [gnugo-devel] Twogtp matches

[Gunnar Farnebäck]

Dan wrote:
> The second was 3.5.9 (W) vs 3.4 (B) with 5.5 komi. The
> results were about even. W won 52 games and was ahead
> by an average of .89 points.
> 
> I'm attaching the worst game at the end.

I tested to replay white's moves with 3.4. The following moves differ
more than by random choice:

Move 48 (white): GNU Go plays L5 (11.75) - Game move N4 (11.14)

  The 3.5.9 move looks like an improvement, although the followup at
  move 50 is bad.

  
Move 52 (white): GNU Go plays E11 (10.90) - Game move S19 (10.48)

  Neither move is great but 3.4 looks better.

  
Move 62 (white): GNU Go plays B4 (37.63) - Game move F5

  The 3.4 move seems better. 3.5.9 probably has an owl reading problem.

  
Move 68 (white): GNU Go plays L4 (7.84) - Game move D2

  3.5.9 spends a ko threat pointlessly, might be a real problem.

  
Move 70 (white): GNU Go plays L4 (7.84) - Game move K5 (7.64)

  No big difference between the moves.

  
Move 78 (white): GNU Go plays S9 (10.05) - Game move P4 (7.47)

  Both moves okay.

  
Move 86 (white): GNU Go plays K7 (7.93) - Game move P10 (6.98)

  3.4 maybe somewhat better.

  
Move 104 (white): GNU Go plays N13 (31.04) - Game move K13 (30.23)

  3.4 looks clearly better here

  
Move 112 (white): GNU Go plays T2 (5.03) - Game move E16 (4.89)
Move 114 (white): GNU Go plays T2 (5.03) - Game move B10 (4.00)
Move 120 (white): GNU Go plays T2 (5.03) - Game move D10 (4.79)
Move 122 (white): GNU Go plays T2 (5.03) - Game move A14 (4.33)

  3.5.9 generally better. T2 overvalued by 3.4.

  
Move 126 (white): GNU Go plays E12 (9.16) - Game move B15 (9.36)

  Both moves kind of big, 3.4 maybe better.

  
Move 138 (white): GNU Go plays T2 (5.03) - Game move A18 (4.42)
Move 144 (white): GNU Go plays T2 (5.03) - Game move J7 (4.46)

  T2 still overvalued by 3.4. 3.5.9 better.

  
Move 158 (white): GNU Go plays K6 (5.81) - Game move K8 (5.05)

  3.5.9 better.

  
Move 172 (white): GNU Go plays L11 (8.68) - Game move J9 (10.76)

  3.4 much better.

  
Move 174 (white): GNU Go plays M12 (48.73) - Game move N17

  3.5.9 much better, but probably only by luck since the followup is wrong.


The important differences, where 3.4 does better, are moves 62, 68,
104, and 172. My guess is that 62, 68, and 172 are owl problems which
have turned up more or less accidentally while doing owl tuning. 104
might be a different problem. All of these may be worth closer
investigation.

Ok, I looked into move 104. There are no big changes in the valuation
and the differences seem mostly to relate to accidental side effects
of influence tuning. (Unfortunately 3.4 doesn't have the GTP commands
needed for view.pike to display influence information nicely, so I
won't investigate this any further.)

/Gunnar