Information Technology Reference
In-Depth Information
the stability-plasticity dilemma and so are stable enough to incorporate new
information without destroying the memory of previous learning. They have
a stable memory structure even with fast on-line learning that was capable of
adapting to new data input, even forming totally new category distinctions.
The most advanced model of the ART family is Fuzzy ARTMAP which
was developed for supervised slow learning. Unlike traditional MLP neural
networks the architecture of FAM is selforganising according, as previously
mentioned, to the plasticity-stability dilemma: the network is able to retain
learned patterns (stability) while remaining able to learn new ones (plastic-
ity). In a standard MLP network used for pattern classification an output
neuron is assigned to each class of objects that the network is expected to
learn, and must be trained off-line. In FAM the network dynamically assesses
the assignment of output neurons to categories by competitive learning. Two
ART modules are interconnected by an associative memory and internal con-
trol structures; the first module handles input patterns while the second one
handles the class patterns. The network is able to perform real time learn-
ing without losing previously learnt patterns by using an incremental weight
update procedure known as slow recoding. The main drawback of FAM is
its architecture complexity; this limit is partially overcome by the Simplified
Fuzzy ARTMAP (SFAM) network that reduces the complexity of the net-
work architectures. It must be noticed that, removing the redundancies in
FAM, SFAM are able to train much faster.
8 Finding the Predictive Model by a Genetic Algorithm
The choice of the most suitable pattern analysis algorithms is not trivial. The
decision is strongly dependent from the specific problem domain and available
data, in terms of quantity and distribution; moreover the performance of a spe-
cific classifier is affected to the previous feature extraction, selection and pro-
jection operations, that is, different classifiers can have different best feature
subset and projection. Thus, each single operation of the pattern analysis mod-
ule should not be considered as a separate step, independent from the others.
If there is no specific reason to choose one approach w.r.t. others, as in the ol-
factory signal analysis usually happens, one should, in theory, evaluate all the
possible combinations of feature selection, projection and classifier. Of course
a similar approach is unfeasible. A possible solution, adopted in this work, is
provided by Genetic Algorithms (GAs).
GAs are heuristic search algorithms based on the mechanics of natural selec-
tion; as optimization technique, they mimic the evolutionary process of survival
of the best taking inspiration by the Darwinian process of natural selection.
They perform a global random search on population of potential solutions (chro-
mosomes), which allow the technique to be massively parallel in operation. In
particular, at each generation the best potential solutions are selected with a
certain probability and used to generate new solutions. These latter are pro-
duced mixing the parents information (crossover) and introducing few random
Search WWH ::




Custom Search