Catálogo de publicaciones - libros

The increased dependence on artificial neural network (ANN) models leads to a key question – will the ANN models provide accurate and reliable predictions? However, this important issue has received little systematic study. Thus this paper makes general researches on verification and validation (V&V) of ANN models. Basic problems for V&V of ANN models are explicitly presented, a new V&V approach for ANN models is developed, V&V methods for ANN models are deeply discussed, further research areas for V&V of ANN models are recommended, and an example is given.

Artificial neural networks (ANN) combined with PCA are being used in a growing number of applications. In this study, the fingerprint wavebands of apple were got through principal component analysis (PCA). The 2-dimensions plot was drawn with the scores of the first and the second principal components. It appeared to provide the best clustering of the varieties of apple. The several variables compressed by PCA were applied as inputs to a back propagation neural network with one hidden layer. This BP model had been used to predict the varieties of 15 unknown samples; the recognition rate of 100% was achieved. This model is reliable and practicable. So a PCA-BP model can be used to exactly distinguish the varieties of apple.

In this paper, an artificial neural networks identifier and a fuzzy logic controller for ITU Triga Mark-II Nuclear Research Reactor is presented. Three parted control function is used as a reference trajectory that the fuzzy logic controller tracks. The nonlinear behavior of the reactor is identified by using generalized neural networks. The validity of the proposed identification model is tested by comparing these results with the ones obtained by YAVCAN code. The effectiveness of the controller is demonstrated on the neural network model.

This paper presents the application of Differential Evolution (DE), an Evolutionary Computation method, for the optimization of adaptive FIR filter weights. This method is robust and easy to use and requires a few control variables. Since the algorithm uses differential property, it has a good convergence speed and also quite robust in the case of noise due to parallel structure. In the simulation study three well-known error functions are used to test the performance of proposed method in the Adaptive Linear Combiner (ALC) design.

In this paper, we introduce new architectures and comprehensive design methodologies of Evolutionally optimized Fuzzy Neural Networks (EoFNN). The proposed dynamic search-based GAs leads to rapidly optimal convergence over a limited region or a boundary condition. The proposed EoFNN is based on the Fuzzy Neural Networks (FNN) with the extended structure of fuzzy rules being formed within the networks. In the consequence part of the fuzzy rules, three different forms of the regression polynomials such as constant, linear and modified quadratic takes into consideration. The structure and parameters of the EoFNN are optimized by the dynamic search-based GAs.

Loop is an important structure in human written programs. However, it is seldom used in the evolved programs in genetic programming (GP). This paper describes an approach to the use of while-loop structure in GP for the factorial and the artificial ant problems. Two different forms of the while-loop structure, count-controlled loop and event-controlled loop, are investigated. The results suggest that both forms of the while-loop structure can be successfully evolved in GP, the system with the while-loop structure is more effective and more efficient than the standard GP system for the two problems, and the evolved genetic programs with the loop-structure are much easier to interpret.

XCS is currently considered as the state of the art Evolutionary Learning Classifier Systems (ELCS). XCS has not been tested on large datasets, particularly in the intrusion detection domain. This work investigates the performance of XCS on the 1999 KDD Cup intrusion detection dataset, a real world dataset approximately five million records, more than 40 fields and multiple classes with non-uniform distribution. We propose several modifications to XCS to improve its detection accuracy. The overall accuracy becomes equivalent to that of traditional machine learning algorithms, with the additional advantages of being evolutionary and with O ( n ) complexity learner.

This paper presents a design method of the automatic loop-shaping which couples up manual loop-shaping method to genetic algorithms (GAs) in quantitative feedback theory (QFT). The loop-shaping is currently performed in computer aided design environments manually, and moreover, it is usually a trial and error procedure. To solve this problem, an automatic loop-shaping method based on GAs and evolutionary computation is developed and a benchmark example is used to examine the performance of the proposed automatic loop-shaping compared with that of the manual loop-shaping and similar other research.

This paper reports on a study which compared the convergence of different-leveled structured Genetic Algorithms (sGAs) used to generate Neural Networks (NNs). Results suggest that sGAs are more effective at generating NNs compared to simple GAs. Using more than 2 sGA levels does not always yield a better error curve, as each added level provides a diminishing performance increase. The optimum number of sGA levels for NN generation is problem specific, though higher level sGAs tend to produce more efficient NNs. SGAs with more levels seem to perform better for difficult NN problems with complex features and large boundary conditions which create more redundancy. When the emphasis on complexity is increased in the fitness function, the error curve variations between different level configurations become more pronounced.

Directed mutation has been proposed for improving the convergence speed of GAs on problems involving real-valued alleles. This paper proposes a directed mutation approach based on the momentum term used in gradient descent training of neural networks. Mutation-with-momentum is compared against gaussian mutation and is shown to regularly result in improvements in performance during early generations. A hybrid of momentum and gaussian mutation is shown to outperform either individual approach to mutation.