W-Infer-polation: Approximate reasoning via integrating weighted fuzzy rule inference and interpolation

Abstract

Fuzzy rule-based inference systems facilitate approximate reasoning for decision-making under circumstances where knowledge is imprecisely characterised. Compositional rule of inference (CRI) and fuzzy rule interpolation (FRI) are two typical types of technique to implement fuzzy systems. Questions of when to apply which of these two fuzzy inference techniques is often addressed by checking whether there exist certain rules that can match given observations. Both techniques have been widely investigated to improve the performance of approximate reasoning, with increasingly more attention being paid to the development of systems where rule antecedent attributes are associated with measures of their relative significance or weights. Unfortunately, they are mostly done in a separate manner within their own fields, making it difficult to achieve accurate reasoning when both techniques are required simultaneously. This paper presents an innovative fuzzy inference mechanism, W-Infer-polation, which organically integrates both enhanced versions of CRI and FRI that are equipped with rule attribute weights. In particular, the individual weights of rule antecedents and consequent are generated using the given unweighted rules only, creating an attribute weighted rule base. From this, W-Infer-polation exploits the learned weights to guide the selection of weighted rules for rule firing in CRI and that of nearest weighted rules for FRI. A systematic comparative experimentation is also presented, demonstrating the efficacy of W-Infer-polation.