Computational Intelligence Techniques - Computational Intelligence in Biomedical Engineering - page 125

a row, which results in k rules governing the fuzzy neural network and the
rules have the form

if A 1 then B 11

.

if A 1 then B 1 k

This is in fact conflicting because we now have several conclusions for a single
conclusion. To alleviate this, we can combine the rules to the following

if A 1 then B 11 OR B 12 ... OR B 1 k

With these ideas it is then further possible to construct AND/OR neurons,
neurons with inhibitory pathways and feedback fuzzy neurons.
Fuzzy neural systems have also been extended to other neural network
architectures such as the fuzzy multilayer perceptron (Figure 3.27), which
is the combination of fuzzy neurons similar to the standard MLP network.
A further example is the fuzzy SOM with linguistic interpretations attached.
The SOM can locate an area of neurons with the greatest weight to simulate
associative memory. This corresponds to a substantial amount of outputs as
it is in most unsupervised learning cases. For a fuzzy SOM, consider each
input variable as having a particular linguistic term defined as a fuzzy set in
the corresponding space, for example, A 1 , A 2 ,..., A n for X 1 . The activation
region indicates how intense or wide the coverage of this input is to the data
space and also implicitly its importance in the network. Higher activation
levels are interpreted as higher dominance of the linguistic variable or that it
is stressed more, for example, the temperature is lower, given X 1 represents
coldness. Temporal aspects of network interaction arise when different levels
of learning intensity occur.

Encoding Decoding

FIGURE 3.27
A general fuzzy MLP with fuzzy encoding and decoding.