Light but fruitful: enhanced fuzzy inference via weight-guided selection of rules with attribute weights

Inference using fuzzy rules enables decision-making that is supported with imprecise knowledge. Unlike conventional fuzzy reasoning approaches which directly perform pattern-matching in response to an input observation, recent techniques have integrated rule-firing-based and rule interpolating-based inference methods. This is in order to address challenging issues where observations are of different matching degrees to the rules within a given rule base, including unmatched ones. While applied generally, such a unified inference mechanism may become too complex to exploit the entire rule base for deriving a reasonable conclusion. In practice, only a small number of ‘appropriate’ rules are selected to accomplish the required inference. This paper presents an enhanced integrated fuzzy inference mechanism, which is fed with fewer rules returned by a weight-guided selection procedure. In particular, the weights of rule attributes are utilised in a dual manner: guiding the selection of appropriate rules for rule firing and determining the nearest neighbouring rules for rule interpolation. The resulting mechanism is applied to a real-world problem, empirically demonstrating its significant efficacy.