Building expert networks that really fly: computational issues

Expert networks have been proposed as a paradigm for combining rule-based expert systems with connectionist training algorithms. The rule-base furnishes the knowledge system with a coarse framework of logically dependent concepts; the training algorithm refines the knowledge by learning the strength of the interdependencies from data. Two learning algorithms, expert network backpropagation (ENBP) and goal-directed Monte Carlo search (GDMC), are considered. Issues which greatly impact the effectiveness of expert networks solutions to application problems include training sample generation, validation/generalization issues, introduction of fault tolerance via alternate path generation, and scaling up to real-sized networks. This paper addresses these computational issues with an eye to identifying the key elements which determine whether an expert network for a particular application will or will not work as expected