Censor Updation during Dynamic Clustering of Hierarchical Censored Production Rules (HCPRs)

An incremental learning algorithm takes a new piece of information at each learning cycle and tries to revise the theory using the new data. In this paper a cumulative learning methodology, analogous to incremental learning, is suggested for appropriate modification of censor conditions during dynamic clustering of hierarchical censored production rules (HCPRs). HCPR system is capable of handling trade-off between the precision of an inference and its computational efficiency leading to trade-off between the certainty of a conclusion and its specificity. An HCPR has the form: Decision If <pre-conditions> Unless <censor conditions> Generality <general info> Specificity<specific info>, where censors (exceptions) are assumed to be low-likelihood assertions. Under tight resources, censors can be ignored and decision is true with high likelihood. However, decisions need to be revised if censors are later found to be true. The proposed algorithm appropriately modifies censor conditions at different level of hierarchy thereby removing redundancy and maintains consistency. The resulting knowledge base so obtained is used for the next learning cycle. Examples are given to demonstrate the behaviour of the proposed scheme.