Abstract

We present a relational and evidence-based approach to building systems which can learn various identification, location and planning tasks in spatial and temporal domains. This machine learning problem is a difficult one because it involves, in addition to database operations such as indexing, the abilitytogeneralize over training samples from continuous and relational data types. Relational evidence theory integrates methods from inductive logic programming with those from evidence theory and evaluates the symbolic representations formed. Generalization methods are combined with causal modeling and dynamic constraint satisfaction to optimize both the representation bias and search strategy used during learning. The approach is tested and compared with other machine learning techniques over several different supervised identification and dynamic learning tasks in the spatial and temporal domain. 1 1 Introduction The ability to efficiently predict the existence of patterns is an imp...

Citations