Prediction Methods in System Modeling: An Environment of Granular Computing

Abstract

Prediction in complex systems is an ongoing challenge both as a methodological pursuit and applied endeavor exhibiting far-reaching implications. One of the visible trends in the current developments of prediction models has to be acknowledged: (i) there are no ideal prediction models, and (ii) quantifying quality of prediction becomes of paramount practical relevance. These two observations led to the emergence of a non-numeric constructs and mechanisms of evaluation of prediction results coming in the form of so-called prediction intervals. In this study, we cast the prediction problem in the framework of Granular Computing and take advantage of the well-established methodology and algorithms supporting a comprehensive of processing information granules. The concept of prediction intervals is generalized to prediction information granules. We also benefit from a variety of ways in which information granules are formally expressed as intervals, fuzzy sets, rough sets, etc. The ensuing algorithms producing granular prediction results are classified as those implied by optimized granular parameter space or optimized granular output space. We demonstrate that the proposed approach helps view prediction intervals as some special cases of prediction information granules. The concept of information granularity is also studied in the context of non-numeric (granular) prediction horizons, which can be conveniently formalized with the aid of information granules. Furthermore the concepts of information granules of type-2 and their role in the enhancements of prediction models are studied.