Re: [gnugo-devel] thrashing dragons again

[Arend Bayer]

Gunnar wrote:

> Arend wrote:
> > - We use the thrashing dragon heuristic much more often: Do it even if we
> > are behind (up to 30 pts), and don't bother to make it dependant on the
> 
> The code comments are out of sync with this change.
Will fix this.
> 
> > move value without the thrashing dragon heuristic. This means we don't
> > have to run review_move_reasons() twice in the case of a thrashing
> > dragon. The rational is that we still won't miss an urgent move somewhere
> > else, as the thrashing dragon moves would have to have higher value
> > that this urgent moves.
> 
> I'm not sure I agree with this. Consider the case where the game is
> pretty much over and GNU Go is about to lose by half a point. Then the
> opponent (stupid computer or confused human) inexplicably adds a stone
> to a dead dragon without a working followup. I would be very annoyed
> if GNU Go answered with a restraining move inside own territory, not
> taking the gift.

There are 2 changes:
1. Dont make it dependent on the size of the move value without
thrashing dragon heuristic.
2. Don't make it dependent on being ahead.

I suppose you only disagree with 2., and it is certainly debatable. My
change was motivated by the fact that when a thrashing dragon move
occurs on kgs, I almost always want gnugo to answer. So my bet is that
it is a win more often than not.

For the record: If I change 30.0 to 0.0, there are 3 PASSes that go away:
nngs:440, gunnar:46, 9x9:250. I will try to think about this more.

> > nngs3:580       FAIL A4 [H3|J5|K3|J3]
> > Very bad. A4 should not get a strategic attack move reason vs C6.
> 
> Any suggestion how to characterize the A4 move so that it can be
> avoided? We already have a reduction in the attack thrashing dragons
> bonus for connections to inessential stones, but sometimes those moves
> are good ideas.

I don't know. But it would be sufficient if one of the correct moves
would get a strategic attack reason against C6; since their territorial
value is much higher than A4, they will then be played for sure.
(That's why I gave the strategic attacks exactly the same bonus as the
connection moves. I think this is actually the most successful part of
my patch.)


> A different solution for this testcase would be to add a c (weak
> connection) pattern for H3. Strictly speaking it does connect but the
> cutting stones get too many liberties for the connection reading to
> realize this.

Arend