Genetic algorithms (GAs) are stochastic methods that are widely used in search and optimization. The breeding process is the main driving mechanism for GAs that leads the way to find the global optimum. And the initial phase of the breeding process starts with parent selection. The selection utilized in a GA is effective on the convergence speed of the algorithm. A GA can use different selection mechanisms for choosing parents from the population and in many applications the process generally depends on the fitness values of the individuals. Artificial neural networks (ANNs) are used to decide the appropriate parents by the new hybrid algorithm proposed in this study. And the use of neural networks aims to produce better offspring during the GA search. The neural network utilized in this algorithm tries to learn the structural patterns and correlations that enable two parents to produce high-fit offspring. In the breeding process, the first parent is selected based on the fitness value as usual. Then it is the neural network that decides the appropriate mate for the first parent chosen. Hence, the selection mechanism is not solely dependent on the fitness values in this study. The algorithm is tested with seven benchmark functions. It is observed from results of these tests that the new selection method leads genetic algorithm to converge faster.
Underwater robotic vehicles have become an important tool for various underwater tasks because they háve greater speed, endurance, depth capability, and safety than human divers. The problem of controlling a remotely operated underwater vehicle in 6 degrees of freedom (DOF) is addressed in this paper, as an example of a system containing severe non-linearities. Neural networks are been used in a closed-loop to approximate the nonlinear vehicle dynamics. No prior off-line training phase and no explicit knowledge of the structure of the vehicle are required, and the proposed scheme exploits the advantages of both neural network control and adaptive control. A control law and a stable on-line adaptive law are derived using the Lyapunov theory, and the convergence of the tracking error to zero and the bounded-ness of signals are guaranteed by applying Barbalaťs Lyapunov-like lemma. In this páper, a neural network architecture based on radial basis functions has been ušed to evaluate the performance of the proposed adaptive controller for the motion of the Norwegian Experimental Remotely Operated Vehicle (NEROV).
The multilayer perceptron model has been suggested as an alternative to conventional approaches, and can accurately forecast time series. Additionally, several novel artificial neural network models have been proposed as alternatives to the multilayer perceptron model, which have used (for example) the generalized- mean, geometric mean, and multiplicative neuron models. Although all of these artificial neural network models can produce successful forecasts, their aggregation functions mean that they are negatively affected by outliers. In this study, we propose a new multilayer, feed forward neural network model, which is a robust model that uses the trimmed mean neuron model. Its aggregation function does not depend on outliers. We trained this multilayer, feed forward neural network using modied particle swarm optimization. We applied the proposed method to three well-known time series, and our results suggest that it produces superior forecasts when compared with similar methods.
Neural networks have shown good results for detecting a certain pattern in a given image. In this paper, faster neural networks for pattern detection are presented. Such processors are designed based on cross correlation in the frequency domain between the input matrix and the input weights of neural networks. This approach is developed to reduce the computation steps required by these faster neural networks for the searching process. The principle of divide and conquer strategy is applied through matrix decomposition. Each matrix is divided into submatrices small in size, and then each one is tested separately by using a single faster neural processor. Furthermore, faster pattern detection is obtained by using parallel processing techniques to test the resulting submatrices at the same time, employing the same number of faster neural networks. In contrast to faster neural networks, the speed-up ratio is increased with the size of the input matrix when using faster neural networks and matrix decomposition. Moreover, the problem of local submatrix normalization in the frequency domain is solved. The effect of matrix normalization on the speed-up ratio of pattern detection is discussed. Simulation results show that local submatrix normalization through weight normalization is faster than submatrix normalization in the spatial domain. The overall speed-up ratio of the detection process is increased as the normalization of weights is done offline.
This paper presents a segmentation technique to handwritten word recognition. This technique implements an algorithm based on an analytical approach. It uses a letter sweeping procedure with a step equal to the Euclidean distance between an established reference index and the entity (the alphabet letter). Then a dissociation of this entity is achieved when this distance will reach a rate of 80%. Our experience about this segmentation technique gives a rate of 81.05% of recognition. A neural multi-layer perceptron classifier confirms the extracted segment. This procedure is successively repeated from the beginning until the end of the word. A concatenation technique is finally used to the word reconstitution.
A binocular model for the prenatal development of the visual nervous
systém is proposed. The model is able to reproduce some properties observed in mammals at the moment of birth such as retinotopy, oriented receptive fields, and ocular dominance. One of the outstanding features of the model architecture is the existence of dendrodendritic interaction within each layer. The spontaneous activity of the neurons of the input layer is modeled by spatially and temporally decorrelated activity. The evolution of a connection depends on the output activity of both connected neurons. Hebbian learning has been used for the afferent excitatory connections and anti-Hebbian learning for the lateral inhibitory connections. The model is reduced to a set of ordinary differential equations obtained from a statistical treatment of the dynamics that avoids its explicit dependence on the spontaneous activity.
Attention decrease and an eventual micro-sleep of an artificial system operator is very dangerous and its early detection can prevent great losses. This work deals with an early detection of micro-sleep based on analysis of an electroencefalographic activity of tlie brain. There are classic spectral methods - the Discrete Fourier Transform and parametric methods - autoregressive models used for signal processing here. An influence of a band pass filter characteristic on classification is investigated. For the detection of the micro-sleep multi-layer perceptron, radial basis function (RBF) and the learning vector quantization (LVQ) neural networks are used. The k-nearest neighbor as a representative of non-parametric methods is examined. The last method used here is based on the Bayesian theory and its coefficients are found using the maximum likelihood estimation.
In this paper, a mechanism of adaptive width adjustment based on immunological vaccination is proposed for the evolutionary training of RBF neural networks. Inspired by the vaccination process of the natural immune system, the algorithm implements an individual-orientated adaptation of the width in training stages to optimize the potential solutions, therefore reinforces the evolutionary capability and efficiency. A two-layer genotype-coding scheme, which enables a simultaneous evolution of network structure and parameters, is presented to achieve a compact and consistent-in-form solution. The proposed learning strategy is tested on several benchmark problems and results demonstrate promise.
A new method to detect damages on crates of beverages is investigated. It is based on a pattern-recognition-system by an artificial neural network (ANN) with a feedforward multilayer-perceptron topology. The sorting criterion is obtained by mechanical vibration analysis which provides characteristic frequency spectra for all possible damage cases and crate models. To support the network training, a large number of numerical data-sets is calculated by the finite-elementmethod (FEM). The combination of artificial neural networks with methods of numerical simulation is a powerful instrument to cover the broad range of possible damages. First results are discussed with respect to the influence of modelling inaccuracies of the finite-element-model and the support of the ANN by training-data obtained from numerical simulation. Also the feasibility of neuro-numerical ANN training will be dwelled on.
Accuracy alone can be deceptive when evaluating the performance of a classifier, especially if the problem involves a high number of classes. This paper proposes an approach used for dealing with multi-class problems, which tries to avoid this issue. The approach is based on the Extreme Learning Machine (ELM) classifier, which is trained by using a Differential Evolution (DE) algorithm. Two error measures (Accuracy, $C$, and Sensitivity, S) are combined and applied as a fitness function for the algorithm. The proposed approach is able to obtain multi-class classifiers with a high classification rate level in the global dataset with an acceptable level of accuracy for each class. This methodology is evaluated over seven benchmark classification problems and one real problem, obtaining promising results.