Information Technology Reference
In-Depth Information
kind of selection T cells undergo, when apoptosis is induced in all the cells while
the other cells survive to recognize the non-self elements as enemies.
Neural algorithms
share with evolutionary algorithms the notion of fitness. In
fact, they search for the best neural network connecting a given input/output corre-
spondence. Any neuron is an element which is stimulated by inputs and reacts to
them by producing outputs, generated by means of certain
reaction functions
,usu-
ally of sigmoid nature. Their outputs are propagated to neurons of a next level. The
solution that a neural network reaches is the function that provides the outputs cor-
responding to a given class of inputs. The reaction function assigned to each neuron
is chosen within a class of functions, but some parameters, usually called
weights
,
express the specific kind of reactivity for the given neuron. Formally, if we assume
n
neurons for each of
m
levels, the neural network is represented by a matrix
n
m
of
weights
for the neural connection of the neuron to the neurons of the next level
(zero weight means no connection).
In order to find the best neural network providing the best approximation to a
given input/output behavior, some trainings with input/output cases are selected and
the measure of the efficacy of the solution consists of the ability to provide the
right correspondence for all the cases covered by the problem. A different class of
bio-inspired algorithms are the
population-based cooperative algorithms
which
are designed in a perspective where individuals are agents searching for the best
solution to a given problem. In this case, solutions are searched through
cooper-
ative and competitive search
or
cooperative and distributed generation
pro-
cesses. In the first case, some agents move in the solution space by evaluating, at
any step, their current solutions, by comparing them with the current solutions of
other agents, and by using this comparison for developing a better current solution.
The different classes of bio-inspired algorithms, realizing cooperative-competitive
procedures, are characterized by different methods of cooperation-competition. The
following are the general phases of these algorithms:
×
1. Try
2. Compare
3. Modify
4. Communicate.
The
communicate
module refers to the way agents share their knowledge in order
to improve their current solution. Many schemata can be used for realizing this cru-
cial aspect of cooperation. One possibility is based on some common space where
periodically agents put their current solutions, and some specific mechanism, in this
space, for assigning a score to all the solutions. In this case, any agent puts his
solution in the common space if it is better than those which are published there.
Another possibility is realized by means of a network, possibly dynamically de-
fined, which ensures a communication between agents who ensure they update their
knowledge and the consequent evaluation of their solutions. The agent who first
reaches a solution with a given level of satisfaction publishes it and the process ter-
minates. Therefore, in this case the publication of the solution coincides with the end
of the process. Of course, many other intermediate possibilities can be conceived.
