Search Improvements in Multirelational Learning

In this thesis we lay the foundations to develop multirelational learning systems which can cope better with the challenges posed by structural and topological domains. Even though many interesting application domains contain structural or topological data, current multirelational systems have difficulties dealing with the complexity of the search space of theses domains, their indeterminacy, and the presence of non-discriminating relations. In this work, we describe macrooperators which are a formal method to reduce the search space explored in structural or topological domains and can also be used to alleviate the myopia of greedy systems. We also explore parallel search based on stochastically selected examples to reduce the instability of example-driven learning. As a third contribution, we present active inductive learning as an approach to improve the efficiency of the instance space exploration.