Re: [computer-go] Statistical significance (was: SlugGo vs Many Faces,newest data)

["chrilly" <chrilly@xxxxxxxx>]

I agree fully with the points Don and Mark have mentioned. It is exactly my
experience. Actually it is even worse. In a program with search, the
resulting positions are from a human perspective bizarre. These positions
are never seen in a human game (and probably also never in the mind of a
human player). Nevertheless the decision at the root is based on the
evaluation of these positions. The evaluation function has therefore to be
able to score these weired positions.
I would also say that being able to evaluate these positions is a strength
of the programs in relation to humans. Human-chess was certainly influenced
by computer-chess in the last years. Humans have learned that much more
positions are playable than they thought before.

One very interesting point in Don comments is the dependence of the
evaluation on the opening-phase. One of the examples are in chess the so
called Piece-Square-Tables. Each piece-type gets a fixed evaluation-term
depending on the square it stands (independent of the other pieces). These
encodes practically some opening knowledge where to develop pieces. I have
recently removed these Piece-Square-Tables. This was probably the best idea
I had in the last 10 years.

Chrilly Donninger

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