Much of the current research in machine learning has focused on the incorporation of multi-strategy learning paradigms. The motivation for such hybrid systems is clear; current systems based on unimodal paradigms tend to be too domain specific. The concepts such systems are capable of deriving, while demonstrating salient features, are overly restrictive and do not give rise to real world applications. This inability to "scale up" is the center of much research in the AI community.;The system presented in this paper, the Hybrid Boolean Learning Algorithm (HYBAL), is a hybrid learning algorithm employing both analytical and empirical methods. It will be demonstrated that HYBAL is capable of accurately solving problems where the hypothesis space grows exponentially. HYBAL implements analytical learning via Version Graphs and empirical learning via Genetic Algorithms; incorporation of these two methods was originally proposed in Reynolds' VGA. HYBAL extends the work on the VGA in 2 directions; it provides substantial computational speedup, and it is capable of learning a larger class of boolean concepts.;In solving problems of exponential complexity, it will be demonstrated that static version graphs significantly restrict the concepts that can be learned. This restriction is due to the specification requirements of static graphs; the implementer is required to enumerate, a priori, a list of candidate hypotheses. HYBAL will circumvent such bias by dynamically factoring the hypothesis space and incrementally building a version graph. As such, HYBAL may be viewed as a "scaling up" extension of Hirsh's Incremental Merging Algorithm, which in turn extended the original Candidate Elimination Algorithm.
展开▼
