[gnugo-devel] --experimental-semeai

[Arend Bayer]

I just looked throught the --experimental-semai FAILs, and would like to
cast a vote for making --experimental-semeai default.

Dan wrote:
> In other news, here's the current results with experimental semeai
> enabled.

> ./regress.sh . lazarus.tst
> 15 unexpected PASS!
> 16 unexpected FAIL: Correct '!L17|J5|K5|K6', got 'L17'

L17 only looses a few points.

> ./regress.sh . trevorb.tst
> 430 unexpected PASS!
> ./regress.sh . strategy2.tst
> 72 unexpected PASS!
> 73 unexpected PASS!
> 75 unexpected FAIL: Correct 'Q11', got 'F7'
> 76 unexpected FAIL: Correct 'Q11', got 'F7'

Q11 is only correct because it is a owl threat (and with --owl-threats,
GNU Go gets it right again -- IOW, plain GNU Go makes two mistakes
here, GNU GO --experimental-semeai only one).

> 93 unexpected FAIL: Correct 'D11', got 'C11'
This is indeed a mistake and should be fixed.

> ./regress.sh . nicklas1.tst
> 301 unexpected FAIL: Correct 'H3|H6', got 'J8'
Same here (I think this shouldn't be too hard to fix).

> ./regress.sh . trevor.tst
> 461 unexpected PASS!
> ./regress.sh . nngs.tst
> 290 unexpected PASS!
> 2040 unexpected PASS!
> ./regress.sh . strategy3.tst
> 113 unexpected FAIL: Correct 'P1', got 'M15'
Bad mistake, indeed very surprising this can happen.

> 135 unexpected PASS!
> ./regress.sh . global.tst
> 46 unexpected PASS!
> ./regress.sh . 13x13.tst
> 39 unexpected PASS!
> ./regress.sh . strategy4.tst
> 188 unexpected PASS!
> ./regress.sh . nngs2.tst
> 50 unexpected FAIL: Correct 'K2', got 'N1'
Also bad mistake. (I don't quite understand what is going wrong here.)
> 60 unexpected PASS!

So there remain 4 mistakes that need fixing.

One reason not to make the new semeai code default was that it didn't
help much in the GNU Go vs GNU Go GTP matches. This might just indicate
that there are very few semeais in these matches (I don't think I've
ever seen one).

But a more important reason why I'd like to see it as default is: If
GNU Go makes a semeai mistake (that can often cost a game) with
--experimental-semeai, then we at least get s.th. to fix, instead of
"hmm guess this wouldn't have happened with --experimental-semeai"...

Also, the strategy valuation is partly a guess work towards reducing the
impact of owl failures (that will always happen until GNU Go is 6D).
I'd rather tune this towards including the --experimental-semeai
failures if we want to include it

Arend