Dynamic Fuzzy Rule Interpolation

Abstract

Designers of effective and efficient fuzzy systems have long recognised the value of inferential hybridity in the implementation of sparse fuzzy rule based systems. Which is to say: such systems should have recourse to fuzzy rule interpolation (FRI) only when no rule matches a given observation; otherwise, when an observation partially or exactly matches at least one of the rules of the sparse rule base, a compositional rule of inference (CRI) should be used in order to avoid the computational overheads of interpolation. Sparse fuzzy rule bases are constructed by experts or derived from data and may support FRI reasoning in long run. However, two potential problems arise: (1) a system’s requirements may change over time leading to rule redundancy; and (2) the system may cease in the long run to provide precise and pertinent results. The need to maintain the concurrency and accuracy of a sparse fuzzy rule base, in order that it generates the most precise and relevant results possible, motivates consideration of a dynamic (real-time) fuzzy rule base. This thesis therefore presents a framework of dynamic fuzzy rule interpolation (D-FRI), integrated with general fuzzy inference (CRI), which uses the FRI result set itself for the selection, combination and promotion of informative, frequently used intermediate rules into the existing rule base. Here two versions of the D-FRI approach are presented: k-means-based and GA-aided. Integration uses the concept of α-cut overlapping between fuzzy sets to decide an exact or partial matching between rules and observation so that CRI can be utilised for reasoning. Otherwise, the best closest rules are selected for FRI by exploiting the centre of gravity (COG), Hausdorff distance (HD) and earth mover’s distance (EMD) metrics. Testing seeks to show that dynamically-promoted rules generate results of greater accuracy and robustness than would be achievable through conventional FRI tout court, and to support the claim that the D-FRI approach results in a more effective interpolative reasoning system. To this end, an implementation of D-FRI is applied to the problem domain of intrusion detection systems (IDS), by integrating it with Snort in order to improve port-scanning detection and increase the level of accuracy of alert predictions.

Date of Award	13 Jan 2015
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Qiang Shen (Supervisor) & Neal Snooke (Supervisor)