 |
A.A. Zhilenkov, S.G. Chernyi Extracting information using neural network architectures as networks of information granule associations |
 |
|
Abstract.
The article proposes an approach to building granular models directly on the basis of information granules expressed both in the input and output spaces. By linking these information granules, the constructed granular models are represented within neural networks in a system that includes three levels: input granules, output schema, and output granules. An extended principle of reasonable granularity is applied to construct information granules in the input space. This principle creates information granules not only by striking a reasonable balance between the two criteria, coverage and specificity, but also by optimizing these information granules based on their homogeneity, assessed with respect to the data localized in the output space. An inference scheme is assumed by analyzing the location of the input data in relation to the already formed information granules in the input space. The calculated ratio can be quantified as degrees of membership, resulting in aggregation results that include information granules in the output space. The high efficiency of the proposed granular model is ensured by the mechanisms of granular computations and the principle of justified granularity. Experimental studies have been carried out on synthetic and public data and some benchmarking has been done using rule-based models.
Keywords:
information, neural network structures, modeling, network architecture, optimization, information granules.
PP. 81-90.
DOI: 10.14357/20790279220308
|
