- author = "Hong Guo",
- title = "Feature generation and dimensionality reduction using genetic programming",
- school = "Department of Electrical Engineering and Electronics, University of Liverpool",
- year = "2009",
- address = "UK",
- keywords = "genetic algorithms, genetic programming, KPCA, KGDA, ANN, ASK2, ASK4, BPSK, QPSK, FSK2, FSK4, QAM16, V29, V32, QAM64,",
-
URL = "
http://library.liv.ac.uk:2082/search~S8?/lTHESIS+20960.JI/lthesis+20960+ji/-3%2C-1%2C0%2CE/frameset&FF=lthesis+20960+guo&1%2C1%2C",
-
URL = "http://ethos.bl.uk/OrderDetails.do?did=1&uin=uk.bl.ethos.511054",
-
URL = "
https://livrepository.liverpool.ac.uk/3174551/1/511054.pdf",
- size = "198 pages",
-
abstract = "Generally speaking, a standard pattern recognition
system consists of two components, feature extraction
and pattern classification. During the feature
extraction process, information relevant to the pattern
classification is expected to be extracted from the
data and prepared as features to the inputs of the
classifier. Mathematically, this process is viewed as a
transformation of the original parameter vector into a
smaller feature vector with useful information for the
classification task ahead. Usually the original
parameter vector has fairly large number of dimensions,
in the order of hundreds (in some cases, even
thousands), for the coverage all the possible
mathematical relations to the problem. However, this
poses a significant computation demand to the
classifier. The feature space associated with the
original vector expands with the increase of
dimensions. As a result of that, the complexity of the
classifier might increase exponentially. This is
famously called the curse of dimensionality. Feature
extraction is introduced and designed to tackle this
problem, by reducing the dimensionality in the data. In
addition, it could possibly improve the classification
efficiency. In modern pattern recognition systems, the
functionality of feature extraction has become very
important for its ability to extract discrimination
information from the data and filter out unwanted
interference. The success of classification is often
linked to the quality of the extracted features.
In this research work, endeavours have been made to address the feature extraction issue using the strength of evolutionary computation. A genetic programming based feature extraction framework is proposed for the problem. The idea is inspired by the biological evolution process, in which the stronger features survive and weaker ones are washed out. It is believed that for features under-going such evolution, the discrimination ability will have been optimised under a certain criterion. Therefore, the classification benefits from such optimisation of features.
Compared to conventional feature extraction methods, such as Principal Components Analysis (PCA), Fisher Linear Discriminant Analysis (FLDA), Kernel Principal Components Analysis (KPCA) and Generalised Discriminant Analysis (GDA), this approach undertakes a genetic search in the feature space. The objective of the search is to get closer to the core of the problem. Rather than over-fitting, the genuine characteristic of each pattern is more likely to be identified. Practically, certain termination criterion arc set as the job-done points, such as a threshold on the fitness value, the maximum number of generations, etc. They are empirical values set to achieve satisfactory results for specific applications.
In order to cover different scenarios and thoroughly examine the capability of the proposed method, three systems are designed, including a multi-feature extraction system for multiclass problems, a single-feature extraction system for dual-class problems, and a single feature extraction system for multi-class problems. Real data are used for the evaluation of the performance of the systems developed. A series of experiments are conducted to evaluate and compare the results obtained using the combination of different feature extraction methods and classification methods. The classifiers involved range from a simple Minimal Distance classifier (MDC) to a sophisticated Multi-Layer Perceptron (MLP) neural network.
The results demonstrate that the proposed approach is superior to conventional methods for feature extraction for selected applications. The genetic programming system outperforms other systems in terms of classification success. The experimental results are promising. It is believed that this design can be implemented and applied in practical pattern recognition problems as a remarkably cost-effective solution.",
- notes = "LOCATION Harold Cohen Library THESIS 20960.GUO uk.bl.ethos.511054 Supervisor: Asoke K. Nandi",
