Abstract
Neural networks and fuzzy neural networks - fundamental constructs of Computational Intelligence are sources of knowledge. In real-word scenarios, such models become engaged in various pursuits of knowledge management: collaboration, consensus building, knowledge transfer and alike. A concept that is crucial to knowledge management is that of information granularity.
Information granules offer a great deal of flexibility and become necessary to realize mechanisms of knowledge management. The result of knowledge management is a granular neural network or granular fuzzy neural network, which become a more general model by emerging at the higher level of abstraction. Numeric connections are generalized to their granular counterparts, say intervals, fuzzy sets, rough sets, just to name a few commonly available alternatives.
The diversity of formal treatments of information granules implies a significant variety of realizations of granular neurocomputing coming in the form of interval fuzzy neural networks, fuzzy fuzzy (fuzzy2) neural networks, probabilistic neural networks, and rough neural networks.
We discuss main categories of mechanisms of knowledge management including knowledge transfer and two modes of collaboration (hierarchical and non-hierarchical). In all design pursuits of granular fuzzy models, information granularity (and ensuing information granules) is regarded as an important design asset whose proper allocation results in the optimal granular architecture. Both a single-criterion and multiobjective formulations of the optimization of granular neural networks are presented.
A number of essential protocols of granularity allocation are studied along with the detailed optimization schemes. Given the underlying nature of the optimization problem itself, we confine ourselves to the technology of Evolutionary and Swarm Computing.