State of the Art and Development Trend - Computational Intelligence in Time Series Forecasting - page 340

higher hierarchical level, so that they can be seen as partial open systems
themselves. This interaction creates new, more complex open systems having a
“higher order” intelligent behaviour, which is analogous with the capabilities of
biological modules building higher level systems (multi-cellular organisms) out of
lower level modules (cells) that, within the higher level system, behave as partially
bounded open systems with mutual interaction.
The core issue, however, is: How should the modules interact mutually? This is
the issue that was irrelevant for general systems theory. Furthermore, the question
also arises as to what internal models should be embedded in individual modules.
At least now, contemporary intelligent technology, particularly neuro-technology,
is called for help. For instance, in analogy with the modules of biological systems,
modules made up of neural networks should be structured as kinds of nested
networks made up of networks that themselves build the individual modules
capable of mutual communication. This indicates that the overall hierarchically
organized modular system should have some fractal structure.
The operational principle of fractally configured neural networks is as follows.
The modules at the lowest hierarchical level primarily have a sensing function.
While interacting with the environment, the basic function is to collect the input
data and to learn their characteristic features. The modules thereafter interact with
modules of the next higher hierarchical level by sending the results of learning to
them. The higher level modules receive from more than one lower level module the
information learnt and perform a “higher level abstraction” that is forwarded to
higher level modules, etc . This procedure is repeated until the central module of
the system receives the combined information needed for final recognition and
interpretation of the environment situation.
Following this operational principle, the entire neural network to be built
becomes fractally configured. The problem now is what types of neural network
should be used for system implementation. Because the modules should transfer
the learning results towards to higher level modules, the feed-forward networks
could be appropriate for this function. These types of network, however, do not
have the storage capacity that, for example, the recurrent networks have. They can
also perform self-organized learning, but, again, cannot be easily organized
hierarchically. For this purpose, Morita (1993) proposed using what he called non-
monotone neural networks , capable of “abstracting” the input signals and of
building the associative memory .
Finally, the structure of the hierarchically organized modular neural network
was worked out as shown in Figure 10.7, in which the sensory level , recognition
level , abstraction (generalization) level , and the final interpretation and decision
level are chained hierarchically. This depicts the cerebral cortex hierarchy made
up of sensory cortices , association cortices , frontal association cortices , and the
central motor cortex on the top of the hierarchy. From the figure it is evident that
the fractal neural networks are tree-structured neural networks made up of
hierarchically distributed sub-network clusters.
All the modules presented in Figure 10.7 are made up of non-monotone neural
networks, the simplified structure of which is shown in Figure 10.8. In fact, the
internal neural networks of modules consist of non-monotone networks,
represented as circles. The non-monotone networks themselves consist of a number