Information Technology Reference
In-Depth Information
for direct encoding represents the
connection matrix
that exactly specifies the
architecture of the network to be evolved. For direct encoding the following
approaches have been recommended:
x
connectionist encoding
x
node-based encoding
x
layer-based encoding
x
S-expressions based encoding.
Indirect encoding needs much more work in styling the phenotypes adequately,
because here, in encoding of phenotypes, the
rewrite rules
and
construction rules
are also applied recursively. For indirect encoding the following approaches are
recommended:
x
matrix re-writing
x
edge encoding
x
cellular encoding
x
growth encoding
6
5
10
3
4
0100
1010
0110
0101
0011
Figure 8.1.
Binary representation of parameter values
Figure 8.1 shows the transparency and the simplicity of binary representation of
a neural network, whose architecture is given and whose connection weights are
represented as a 4-bits binary chain. Binary representation enables a direct acting
of crossover and mutation operators on the coding structure. But still, the serious
drawback of binary coding is that the total length of the
concatenated strings
grows steadily with the number of interconnections to be considered. This,
increasingly slows down the computational speed of the genetic algorithm. The
total length of concatenated strings grows even more if the higher computational
accuracy is required, because in this case more bits need to be represented in
binary. This can be mastered by using the real numbers for connection weight
representation, so that each individual in the evolving population becomes a real
vector. However, new circumstances are faced here, since it is difficult to use
directly the binary-encoded crossover and mutation operators. A better way to
Search WWH ::
Custom Search