Hello Lev

I finally got time for reading the paper (� what is a structural
representation in chemistry � you referred to in a discussion 2 or 3 months
ago.
You are perfectly right to notice that existing methods suffer from their
stiffness at changing distance measure, but you do not suggest attacking
this problem which could be a general classification problem. Instead, you
attack a peculiar � if interesting sub problem � the one a block building
domain, where the building rules are  fairly well-known.

I have two specific critiques you might answer.

1. Your definition 23 leaves little room for noisy data. The curse of
noise-brittleness is on others too, such as ILP, but many are attacking this
issue. The relations defined in ILP look strangely similar to your
structures.
2. The main problem, besides noise, is the one of the existence of both -
structure and numerical uncertain values. I would think that Bayesian
networks are more able to take them into account than your approach. The
relational net you propose could a base for such a BN.

Best cheers


Yves

Reply via email to