Donald Bren School of Information and Computer Sciences

This thesis develops the Binary Predicate Execution Model; a distributed, massively-parallel system for semantic networks and knowledge bases that is built on a subset of first-order predicate logic. The use of logic gives the model an easily-understood programming paradigm and a well-defined semantics of execution. When expressed in binary predicates, a simple graphical interpretation can be used. All program facts are represented in an assertion graph. Each vertex is associated with a term appearing in a fact and the edges are labeled with the predicate names. Similar graphs are also associated with each rule body and the query. Finding all possible solutions corresponds to finding all possible matches between the query graph and the assertion graph. Invoking a rule corresponds to substituting the graph of its body constrained by the dependencies between its arguments. This can be implemented in a parallel, message-passing fashion where the assertion graph vertices are active processing elements which asynchronously exchange messages identifying different parts of the query that remain to be matched and containing any binding information from previous matching required to accomplish this. The model is data-driven since every message can be immediately processed without the need for any centralized control or centralized memory. By restricting how functional terms can occur, distributed data structures and remote data look-ups for unification are eliminated. Thus, the model's performance on increasingly larger problems scales-up given increasingly larger machines in most cases. Architectural support for the model is investigated and simulation results of a relatively simple software implementation are reported. This suggests performance on the order of 10^5 logical inferences per second for 256 processing elements in an n-cube configuration. Further research directions, including that of increasing efficiency, are discussed.