STEPS-FIRST: A systematic inference system with facts and rules stored as activation patterns

The ability to apply a rule to a set of known facts is a common task in both natural and artificial intelligence. Many symbol processing system perform this task well; however, many previous connectionist inference systems require hard-coded rules and are limited in expressive power. This thesis presents the connectionist-inspired system STEPS-FIRST. It is a systematic tensor product system with facts and inference rules stored as activation patterns. The system has a central configurable inference engine which applies a single rule at a time to a set of facts. Both the rules and facts are stored as tensor products. The system binds values to variables in a powerful and systematic manner. This thesis describes the construction of the network and a detailed set of tests to highlight its abilities. Additionally, certain variants of the system are analyzed with respect to enhancing its performance.