Halis Altun

Treatment of multi-dimensional data to enhance neural network estimators in regression problems

Abstract This paper proposes and explains a data treatment technique to improve the accuracy of a neural network estimator in regression problems, where multi-dimensional input data set is highly skewed and non-normally distributed. The proposed treatment modifies the distribution characteristics of the data set. The prediction of the suspended sediment, which is an important problem in river engineering applications, will be considered as a case study. Conventional approaches lack in providing high accuracy due to the inherently employed simplicity in order to obtain empirical formulae. On the other hand, artificial neural networks are able to model the non-linear characteristics of the mechanism of the sediment transport and have a growing body of applications in diverse applications in civil engineering. It will be shown that a significant enhancement and superior score in accuracy, compared with the classical approaches, are obtainable when the proposed treatment is employed. The proposed technique is an extension to the understanding of the practical aspects of neural computing applications. Therefore the outcome of the present study is important as it is applicable to any scenario where neural network approaches are involved.

Boosting selection of speech related features to improve performance of multi-class SVMs in emotion detection

This paper deals with the strategies for feature selection and multi-class classification in the emotion detection problem. The aim is two-fold: to increase the effectiveness of four feature selection algorithms and to improve accuracy of multi-class classifiers for emotion detection problem under different frameworks and strategies. Although, a large amount of research has been conducted to determine the most informative features in emotion detection, it is still an open problem to identify reliably discriminating features. As it is believed that highly informative features are more critical factor than classifier itself, recent studies have been focused on identifying the features that contribute more to the classification problem. In this paper, in order to improve the performance of multi-class SVMs in emotion detection, 58 features extracted from recorded speech samples are processed in two new frameworks to boost the feature selection algorithms. Evaluation of the final feature sets validates that the frameworks are able to select more informative subset of the features in terms of class-separability. Also it is found that among four feature selection algorithms, a recently proposed one, LSBOUND, significantly outperforms the others. The accuracy rate obtained in the proposed framework is the highest achievement reported so far in the literature for the same dataset.

Implementing soft computing techniques to solve economic dispatch problem in power systems

Soft computing is the state-of-the-art approach to artificial intelligence and it has showed an excellent performance in solving the combined optimization problems. In this paper, issues related to the implementation of the soft computing techniques are highlighted for a successful application to solve economic dispatch (ED) problem, which is a constrained optimization problem in power systems. First of all, a survey covering the basics of the techniques is presented and then implementation of the techniques in the ED problem is discussed. The soft computing techniques, namely tabu search (TS), genetic algorithm (GA), Hopfield neural network (HNN) and multi-layered perceptron (MLP) are applied to solve the ED problem. The techniques are tested on power systems consisting of 6 and 20 generating units and the results are compared to highlight the performance of the soft computing techniques. Future directions and open-ended problems in implementation of soft computing techniques for constrained optimization problems in power system are indicated. Suggestions are presented to improve soft computing techniques.

New frameworks to boost feature selection algorithms in emotion detection for improved human-computer interaction

One of the primary aims in human-computer interaction research is to develop an ability to recognize affective state of the user. Such ability is indispensable to have a more human-like nature in human-computer interaction. However, the researches in this direction are not mature and intensive efforts have only been witnessed recently. This work envisages the possibility of enhancing feature selection phase of emotion detection task to obtain robust parameters which will be determined from verbal information to achieve an improved affective human-computer interaction. As highly informative feature selection is believed to be a more critical factor than classifier itself, recent studies have increasingly focussed on determining features that contribute more to the classification problem. Two new frameworks for multi-class emotion detection problem are proposed in this paper, so as to boost the feature selection algorithms in a way that the selected features will be more informative in terms of class-separability. Evaluation of the selected final features is accomplished by multi-class classifiers. Results show that the proposed frameworks are successful in terms of attaining lower average cross-validation error.

Treatment of skewed multi-dimensional training data to facilitate the task of engineering neural models

Successful application of neural network models relies heavily on problem-dependent internal parameters. As the theory does not facilitate the choice of the optimal parameters of neural models, these can solely be obtained through a tedious trial-and-error process. The process requires performing multiple training simulations with various network parameters, until satisfactory performance criteria of a neural model are met. In literature, it has been shown that neural models are not consistently good in prediction under highly skewed data. Consequently, the cost of engineering neural models rises in such circumstance to seek for appropriate internal parameters. In this paper the aim is to show that a recently proposed treatment of highly skewed data eases the task of practitioners in engineering neural network models to meet satisfactory performance criteria. As the applications of neural models grows dramatically in diverse engineering domains, the understanding of the treatment show indispensable practical values.

Constrained economic dispatch with prohibited operating zones: a Hopfield neural network approach

The work explores the use of Hopfield neural network for a load dispatch problem in a power system. It is a constrained economic dispatch problem with prohibited operating zones. Yalcinoz and Short (1997) discussed a special methodology to improve the performance of Hopfield networks for solving the unconstrained economic dispatch problem. In this paper the improved Hopfield network is applied to the constrained economic dispatch problem. A new mapping process has been used and a computational method for obtaining the weights and biases is described using a slack variable technique for handling inequality constraints. Applying the proposed approach is demonstrated by a 15 unit system with 4 units having prohibited zones.

Implementation of HOG algorithm for real time object recognition applications on FPGA based embedded system

Recent years HOG algorithm has been used to recognize objects in images, with complex content, with a very high success rate. Hardware implementation of this algorithm is very important because of the fact that it can be used in many object recognition applications. In this work HOG algorithm is implemented on FPGA to recognize different geometrical figures with a very high success rate. Objects vertical and horizontal edges have been sharpened using edge detection algorithms to calculate magnitude and angle of the local gradients. Obtained result are used to construct the histograms of gradient orientation. It is observed that each constructed histogram have distinctive features for every object. Rule based classifiers has been used to implement a successful real time object recognition approach on embedded system.

Improved license plate detection using HOG-based features and genetic algorithm

In this paper, a new improved plate detection method which uses genetic algorithm (GA) is proposed. GA randomly scans an input image using a fixed detection window repeatedly, until a region with the highest evaluation score is obtained. The performance of the genetic algorithm is evaluated based on the area coverage of pixels in an input image. It was found that the GA can cover up to 90% of the input image in just less than an average of 50 iterations using 30×130 detection window size, with 20 population members per iteration. Furthermore, the algorithm was tested on a database that contains 1537 car images. Out of these images, more than 98% of the plates were successfully detected.

New feature selection frameworks in emotion recognition to evaluate the informative power of speech related features

In this paper, we propose two new frameworks, so as to boost the feature selection algorithms in a way that the selected features will be more informative in terms of class-separability. In the first framework, features that are more informative in discriminating an emotional class from the rest of the classes are favoured for selection by the feature selection algorithms. In the second framework features that more informative in terms of separating an emotional class from another one are favoured for selection. Then, final feature subsets are constructed from the subsets of selected features using intersection and unification operators. It will be shown that the proposed frameworks fulfill the objectives by considerably reducing average cross-validation error.

The effect of genetic algorithm parameters on the solution of plate location detection

In this study, a new method based on genetic algorithm and neural networks for determining licence plate location is proposed. The effect of genetic algorithm parameters on the quality of solutions is investigated. The method is able to successfully locate a licence plate in avearge 40 msn, on the image of 768x288 size. This score is 200 times quicker compared to sequential search method. Futhermore the method is able to find multiple plates in an image.