Databases Reference
In-Depth Information
is considered as a swarm and the mean square error (MSE) of each net is
considered for training the nets using PSO.
Further this discussion has been extended by incorporating Polynomial
Neural Network (PNN) to the fuzzy swarm net architecture to design the
architectures for Optimal Polynomial Fuzzy Swarm Net (OPFSN).
8
To
design this model at the first step different combinations of the set of input
features are taken to generate linear/quadratic/cubic polynomials called
Partial Descriptions (PDs). Least Square Estimation (LSE) technique is
used to determine the coecients of these PDs. The outputs of these PDs
are considered as the input to the fuzzy swarm net model to design the
classifier.
This chapter is organized as follows. In Sec. 9.2 the fuzzy net archi-
tecture is briefly discussed. Section 9.3 discusses the basics of particle swarm
optimization. Section 9.4 describes the fuzzy swarm net (FSN) classifier
design. The basic concept of PNN has been discussed in Sec. 9.5. Design
of classifier with Optimized Polynomial Fuzzy Swarm Net (OPFSN) has
been covered in Secs. 9.6. and 9.7 illustrates the experimental studies with
fuzzy swarm net and OPFSN models. We have concluded this chapter with
Sec. 9.8.
9.2. Fuzzy Net Architecture
The Multilayer Perceptron (MLP) architecture is used widely for many
practical applications, but it possesses certain drawbacks. A uniform or
standard model does not exist which will suit any type of application. For
each application a different model needs to be designed i.e., the user should
understand the complexity of the problem thoroughly and should decide
the number of hidden layers and the number of hidden nodes in each layer
to be taken. Understanding the complexity of a problem is not an easy
job for a new user, even an experienced user may face diculties in making
correct decision about the architecture design. Therefore very often the user
opts for a trial and error method to decide the MLP architecture. But it
is also not viable to explore all possible designs from the vast search space
to find the best one. As a result instead of searching for the best one we
compromise with an architecture that is acceptable.
The user very often remains in search of a net that is free from such
complexities and time consuming efforts. As an alternative approach certain
flat nets are suggested such as Functional Link Artificial Neural Network
