Generalized Probabilistic Reasoning and Empirical Studies on Computational Efficiency and Scalability

Expert Systems are tools that can be very useful for diagnostic purposes, however current methods of storing and reasoning with knowledge have significant limitations. One set of limitations involves how to store and manipulate uncertain knowledge: much of the knowledge we are dealing with has some degree of uncertainty. These limitations include lack of complete information, not being able to model cyclic information and limitations on the size and complexity of the problems to be solved. If expert systems are ever going to be able to tackle significant real world problems then these deficiencies must be corrected. This paper describes a new method of reasoning with uncertain knowledge which improves the computational efficiency as well as scalability over current methods. The cornerstone of this method involves incorporating and exploiting information about the structure of the knowledge representation to reduce the problem size and complexity. Additionally, a new knowledge representation is discussed that will further increase the capability of expert systems to model a wider variety of real world problems. Finally, benchmarking studies of the new algorithm against the old have led to insights into the graph structure of very large knowledge bases