Using causal knowledge to guide retrieval and adaptation in case-based reasoning about dynamic processes

Abstract

This paper considers case bases used for reasoning about processes where each case consists of a temporal sequence. In general, these temporal sequences include persistent and transitory (non-persistent) attributes. As these sequences tend to be long, it is unlikely to find a single case in the case base that closely matches a problem case. By utilizing causal knowledge in the form of a dynamic Bayesian network (DBN) and exploiting the independence implied by the structure of the network and known attributes, our system matches portions of the problem case to corresponding sub-cases from the case base. The division of a case into sub-cases relies mostly on independence relations extracted from the causal knowledge. The matching of sub-cases takes into account the persistence properties of attributes. The approach is applied to a process involving an automotive paint curing oven in which a vehicle moves through stages within the oven to satisfy some requirements in each stage. In addition, testing has been conducted using cases randomly generated from known causal networks.