Unsupervised credit assignment in knowledge-based sensor fusion systems

Validating knowledge for sensor fusion systems is intrinsically problematic. Knowledge evaluation in this context is viewed as an unsupervised credit assignment problem. An approach to dynamic knowledge evaluation based on a credit assignment algorithm called the bucket brigade is explored. The performance of individual elements of knowledge in rule-based systems is incrementally assessed (credit is assigned) in the conflict resolution phase of the system´s operation. A formulation is developed that avoids the requirement for supervision, leading to dynamic knowledge evaluation that is not dependent on externally derived system critiques. Finally, the mechanism is extended to apportion credit to input sources as well as elements in the system´s knowledge base