Integrating Human Knowledge into a Relational Learning System

The real world can be successfully modeled as consisting of sets of objects, whose associated properties change over time. Relationships among these objects are critical to understanding the world, whether in classifying a current situation or in predicting its future behavior. An intelligent system, then, should be able to incorporate these types of relationships into its reasoning about the world. This thesis presents an initial implementation of the Spatiotemporal Multidimensional Relational Framework (SMRF), rst discussed by Bodenhamer et al. (2009). This probability estimation tree model has two primary contributions: it explicitly uses multidimensional relationships, whereas most relational learning approaches use only single-dimensional relations, and all of its relationships are dynamically learned from the data, rather than being de ned a priori as in traditional relational learning. The great exibility of the SMRF model, however, necessitates expensive computations. To make learning more feasible for large and complex problems, we turn to advice from the human overseeing the learning task, who probably has some idea of what is important in the problem. We describe an early model of using this advice, supporting both advice that speci es objects as being important exemplars in the learning process and cues giving guesses about what kinds of properties are likely to be important to representing the problem. The two types of cues may also be combined. We validate the e ectiveness of both the SMRF model and its cue-based modi cations in both twoand three-dimensional domains created in several di erent ways.