Environmental Engineering Reference
In-Depth Information
evidence. This makes Bayesian model generation and analysis much simpler.
Unfortunately, the independence rule is often violated in real domains and
distorts the final results. This effect has to be carefully controlled.
5.4 Learning Technologies
Learning techniques allow systems to adapt to changing circumstances such as
new environments or failures in hardware. They also allow systems to respond
more rapidly to situations they have previously explored and to which they
have found good solutions. While many techniques exist, this section will focus
on artificial neural networks and genetic algorithms.
5.4.1 Artificial Neural Networks
Artificial neural networks are a learning technique based loosely on biological
neural networks. Figure
5.9
is a pictorial representation of a neural net. Each
“neuron” is a simple mathematical model that maps its inputs to its outputs.
It is shown in Fig.
5.9
as a circle. The input situation is represented as a
series of numbers called the input vector that is connected to the first layer of
neurons. This layer is usually connected to additional layers finally connecting
to a layer that produces the output vector. The output vector represents the
solution to the problem described by the input vector.
Neural networks are trained using a set of input vectors matched to out-
put vectors. By iterating through this training set, the network is “taught”
how to map this and similar input vectors into appropriate output vectors.
During this training, the network determines what features of the input set
are important and which can be ignored. If the training set is complete and
the network can be trained, the resulting system will be robust.
Input
Vector
Output
Vector
.3
.1
.1
.9
.7
.1
Fig. 5.9.
Artificial neural network
