Re: [computer-go] Scalability Odd-Even Effect

["chrilly" <chrilly@xxxxxxxx>]

With my evaluation function there was a strong effect of odd/even search
depth. With odd depth the program would play unreasonable moves (because
it could always play some unreasonable attack as the last move).

Interesting. With a quiesence search it is the other way round. Usually the
moves at odd depths are more reasonable. At least when the search is not too
deep and the search tree is relatively balanced. At very deep searches (e.g.
Hydra searches typically to depth 17-18) the effect disappears. There are
many extensions and selective pruning and the end of the line is in many
case not at the nominal search-depth.

I think that finding reasonable and efficient quiesence rules is a very
critical topic for using a chess-like search in Go. In Chess it is relative
trivial.  But the chess programm Bebe had no quiesence at all. Bebe searched
in 2 ply steps.
I experimented also in Hydra to skip the quiescence search. It does not work
too badly. The evaluation is relative sophisticated and the search was
surprisingly stable. But it was less efficient than with a quiescence.
Typically the search with quiescence recogized tactical wins/losses 1-2
Plies earlier. The overhead for the quiesence is typical a factor of 2. 1-2
Plies additional lookahead is worth appr. a factor 5.

The stability of the search is probably also related with the deep search.
Forcing lines are even extended further and a lot of tactical instabilities
are therefore already resolved within the search tree.

Chrilly


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/