Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 177

network structure, reducing the effect of initial values of parameters, deciding on
the optimal learning parameters, and improving classification accuracy. However,
we may face the following problem in a typical neural network learning
environment. If we use only good examples without noise and missing data (i.e.,
some data items are missing), the trained network reasonably classifies target
patterns but may not have the ability to classify many random patterns into
optimal categories. On the contrary, if we train the network to classify random
patterns, the classification capability of target patterns may not be very good.
To avoid such a disadvantage in the learning, we propose the learning
method of immune multiagent neural networks (IMANNs) [1]. The IMANN has
macrophage agents, T-cell agents, and B-cell agents. The macrophage and T-cell
agents employ the planar lattice neural networks (PLNN) with a neuron
generation/annihilation algorithm [2]. The PLNN consists of hidden neurons in the
lattice and works a similar way to self-organized map (SOM) [3]. B-cell agents
employ Darwinian neural networks (DNN) [4], which have a structural learning
algorithm based on Darwin's theory of evolution.
On the other hand, we use genetic programming (GP) combined with
multiagent systems for rule extraction. We assume that some agents extract
general rules from frequently observed data and other agents extract exceptional
rules from rarely observed data. With cooperation between multiple agents, the
extracted rules can cover all the data. We propose an improved GP method,
automatically defined groups (ADG). The ADG aims to realize effective
cooperative behaviors among heterogeneous agents [5], [6]. This method can
optimize both the group structure of agents and the action rule of each group.
Moreover, there is a learning method for knowledge-based artificial neural
networks (KBANN) with structure-level adaptation of the NN. The KBANN
represents the knowledge structure of experts in the initial network structure and
generates or annihilates hidden neurons to reach a suitable structure during
learning. Extracting rules from KBANN may be easier than from usual neural
networks, because the initial network structure is transformed from If-Then rules
given by experts.
Sections 7.2, 7.3, 7.4, and 7.5 describe KBANN with structure-level
adaptation, ADG, and IMANN, respectively. To verify the effectiveness of our
proposed methods, we apply them to real medical databases and prove their high
classification capability. We also introduce techniques for extracting If-Then rules
from the trained networks.
7.2 KBANN with Structure Level Adaptation
We propose a method of knowledge based artificial neural network with
structure-level adaptation of the NN. This method can determine network structure
to represent the knowledge structure of medical experts. The network structure
translates the intermediate assumption that medical doctors usually set up before
their final diagnosis. This method can generate new neurons or annihilate
redundant ones. The neuron generation/annihilation algorithm, that is,