Learning mutually dependent relations

摘要

So far, the task of learning relations has been concerned with the acquisition of intensional descriptions of unrelated concepts. However, in many real domains concepts are strictly related to each other and the instances of one of them cannot possibly be recognized without previous recognition of other objects as instances of related concepts. A typical case is the problem of labeling parts of a scene in order to interprete it. This paper extends in several ways the learning relations paradigm; in particular, a new methodology, allowing a recursive theory to be inferred from a set of examples, is presented. Each example may contain instances of several relations at the same time. The learning algorithm works bottom up, creating first an acyclic graph that classifies all the instances in the training set. Afterward, a recursive theory is synthesized from the graph. A prototype implementation of the algorithm has been realized by extending the program ML-SMART. The results obtained for an artificial labeling problem are described and discussed.