Abstract
This study contributes to a development of granular models, which are realized through an aggregation of the outputs produced by numeric models formed at a lower level of hierarchy. The aim is to form a consensus through adjusting local sources of knowledge such that the granular outputs could reflect and quantify the diversity of local knowledge. The aggregation process is accomplished in an active mode by introducing a certain level of randomness to the numeric outputs of each individual model. The numeric outputs produced by numeric models are weighted by the analysis of the root mean squared error associated with the corresponding numeric constructs. Granular models are formed at a higher level of abstraction as the results of the aggregation mechanism. An augmented version of the principle of justifiable granularity is used as the underlying development vehicle for constructing granular outputs (intervals, fuzzy sets, etc.). The performance of the granular model is optimized through adjusting the random deviation added to the outputs of each local model and quantified by the coverage and specificity of the granular results. Experimental studies demonstrate that the proposed active aggregation mechanism can effectively improve the performance of the resulting granular models.
•The aim is to reflect and quantify the diversity of local knowledge.•Randomization mechanism is involved in the aggregation mechanism.•Weighted randomly biased outputs are used to build information granules.