Re: [gnugo-devel] Patch: removal of value()s from patterns.db

[Gunnar Farneback]

Inge wrote:
> ./regress.sh . strategy2.tst 
> 79 unexpected FAIL: Correct 'N3', got 'Q6'
> 
> Looking at strategy2:79, I don't think I agree that Q6 is the only
> viable move.

You read this backwards. Black Q6 is not viable at all. The correct
answer N3 truly is huge.

> I didn't touch the high handicap patterns, nor the reinforcement
> patterns, only the ordinary patterns.  I think the reason for the
> small regression delta is that these patterns are simply rests from
> the old engine, and that we have much better machinery for deciding
> values of moves today.

If you by old engine mean 2.6 that is not the case. Most of these
patterns appeared in the 2.7.x series to work around serious defects
in the then new move valuation scheme. It's correct that the machinery
is much better today but the main reason for the small regression
delta is probably that too few of these patterns have any
corresponding test cases. 

Dan wrote:
> I don't think the value(30) in ED45 should be taken out. It is
> valued slightly less than a J pattern in a situation where
> it is urgent that a move be made to prevent W from pushing
> through and cutting. In strategy2:79 it is possible that the
> hane at R4 is better than N3 but actually this is not so
> clear. In any case the move at Q6 is not good here. This
> pattern prevents a tenuki which is almost always a serious mistake.

Running with --experimental-connections we get

  N3: 9.77 - change in territory
  N3:   3.95 - O3 strategic attack/defend
  N3:   4.46 - Q3 strategic attack/defend
  N3:   4.54 - P4 strategic attack/defend
  N3: 4.46 - strategic effect on Q3
  N3: 3.95 - strategic effect on O3
  N3: 4.54 - strategic effect on P4
  N3:   8.36 - total followup value, added 8.36 as territorial followup
  N3: 4.18 - added due to followup (8.36) and reverse followup values (0.00)
  N3: 4.24 - shape (shape values +3.00(1) -0.00(0), shape factor 1.158)
  N3: 0.20 - connects strings (connect value 2, shape factor 0.040)
Move generation values N3 to 31.34

The ED45 fixed value doesn't even come into play. This valuation can
probably be questioned on many accounts but the point is that the move
valuation can understand the urgency of such a move without crude
fixed values. Also without --experimental-connections this could
probably be fixed with a B pattern in conn.db to inhibit the
connection between O3 and Q3.

Arend wrote:
> It may always be a problem that such patterns tend to be to general, and
> I'd prefer to think of all these high fixed value patterns as temporary
> workarounds. E.g. it would be nice if GNU Go could actually understand
> the problem in situations as in this handtalk game. With the new
> connection code it may actually be possible to tackle this. (Note that in
> the handtalk game, C17/C13/C10 are incorrectly amalgamated.)

I whole-heartedly agree about this.

/Gunnar