Lev, I agree completely that we need better method to express knowledge than simply rules and decision trees. First Order Logic is one step beyond that and AI has provided us with a whole subbranch devoted to the representation of knowledge. The issue is not only how to do it but how to find efficient ways of doing it. While decision trees are very efficient to do simple things they are not so great to discover and represent higher-level structured knowledge.
ETS has some advantages here but there are two problems that need to be addressed. First is the need for efficient implementation of the minimal transformation cost calculations and the second is an efficient implementation of the search for the best substitutes that should allow to discover class structures. We have once worked on algorithmic complexity (or Levin's complexity) that in principle could discover the simplest program solving a given task but in practice it couldn't do much being NP hard. Is there any progress in this direction? Wlodek Duch Dept. of Computer Methods, N. Copernicus University http://www.phys.uni.torun.pl/~duch