[gnugo-devel] more break-in evaluation

[Arend Bayer]

I've tried one more go at that. I have replayed 67 games from NNGS with
--without-breakin, and it produced ~180 unique different move
suggestions (with identical random seed), so a little less than 3 per
game.

I manually went through maybe 30% of them, and my impression is as
follows:
- Maybe a third of the deviations are unimportant (e.g. would just
  lead to different order of moves in the actual game)
- Maybe a fifth or so seem to be due to changes in owl results. These
  are probably due to me replaying them not with an identical version
  to the one Gunnar used for NNGS.
- Most of the rest are small improvements by the break-in code, i.e. in
  the order of 1-5 pts.
- Very few, maybe once in 8-10 games, we get a crucial improvement, i.e.
  s.th. between 10 pts more territory up to saving/keeping dead a
  20stone group. (Sometimes by gnugo guessing right really nice tesujis.)

As 10 pts per games is about one stone difference in strength,
this seems to agree with the rating climb of about 0.25 stones that we
have observed on NNGS since the break-in code was turned on by default.

Conclusions:
- I will leave the decision whether we want to have the break-in code
  turned on for 3.4 to Gunnar and Dan, as they are more unbiased.
- I can recommend using it for Gifu, unless GNU Go will be run on a
  very slow computer there.

Arend