AI - a Guide to Intelligent Systems.pdf - Member of EEPIS

NEURAL EXPERT SYSTEMS
263
In rule-based expert systems, the precise matching is required. As a result, the
inference engine cannot cope with noisy or incomplete data.
Neural expert systems use a trained neural network in place of the knowledge
base. The neural network is capable of generalisation. In other words, the new
input data does not have to precisely match the data that was used in network
training. This allows neural expert systems to deal with noisy and incomplete
data. This ability is called approximate reasoning.
The rule extraction unit examines the neural knowledge base and produces
the rules implicitly ‘buried’ in the trained neural network.
The explanation facilities explain to the user how the neural expert system
arrives at a particular solution when working with the new input data.
The user interface provides the means of communication between the user
and the neural expert system.
How does a neural expert system extract rules that justify its inference?
Neurons in the network are connected by links, each of which has a numerical
weight attached to it. The weights in a trained neural network determine the
strength or importance of the associated neuron inputs; this characteristic is used
for extracting rules (Gallant, 1993; Nikolopoulos, 1997; Sestito and Dillon, 1991).
Let us consider a simple example to illustrate how a neural expert system
works. This example is an object classification problem. The object to be
classified belongs to either birds, planes or gliders. A neural network used for
this problem is shown in Figure 8.2. It is a three-layer network fully connected
between the first and the second layers. All neurons are labelled according to the
concepts they represent.
The first layer is the input layer. Neurons in the input layer simply transmit
external signals to the next layer. The second layer is the conjunction layer. The
neurons in this layer apply a sign activation function given by
Y sign þ1; if X 5 0
¼ ; ð8:1Þ
1; if X < 0
where X is the net weighted input to the neuron,
X ¼ Xn
i¼1
x i w i ;
x i and w i are the value of input i and its weight, respectively, and n is the number
of neuron inputs.
The third layer is the output layer. In our example, each output neuron
receives an input from a single conjunction neuron. The weights between the
second and the third layers are set to unity.
You might notice that IF-THEN rules are mapped quite naturally into a
three-layer neural network where the third (disjunction) layer represents the
consequent parts of the rules. Furthermore, the strength of a given rule, or its