Computational Intelligence and Its Applications

The following sections are included:

• Preface

• Contents

This chapter compares three machine learning methods to solve the pose estimation problem. The methods were based on support vector machines (SVMs), AdaBoost and soft margin AdaBoost (SMA) algorithms. Experiment results show that both the SVM-based method and SMA-based method are able to reliably classify frontal and pose images better than the original AdaBoost-based method. This observation leads us to compare the generalization performance of these algorithms based on their margin distribution graphs.

For a classification problem, features selection is the first step, and selecting better features results in better classification performance. The feature selection method described in this chapter is easy and efficient. Instead of resizing the whole facial image to a subwindow (as in the common pose estimation process), the prior knowledge in this method is only the eye location, hence simplifying its application. In addition, the method performs extremely well even when some facial features such as the nose or the mouth become partially or wholly occluded.

Experiment results show that the algorithm performs very well, with a correct recognition rate around 98%, regardless of the scale, lighting or illumination conditions associated with the face.

In this chapter, a particle swarm optimization (PSO) is proposed for polynomial modeling in a dynamic environment. The basic operations of the proposed PSO are identical to the ones of the original PSO except that elements of particles represent arithmetic operations and polynomial variables of polynomial models. The performance of the proposed PSO is evaluated by polynomial modeling based on a set of dynamic benchmark functions in which their optima are dynamically moved. Results show that the proposed PSO can find significantly better polynomial models than genetic programming (GP) which is a commonly used method for polynomial modeling.

Restoration of color-quantized images is rarely addressed in the literature, especially when the images are color-quantized with half-toning. Many existing restoration algorithms are inadequate to deal with this problem because they were proposed for restoring noisy blurred images only. In this chapter, a restoration algorithm based on Particle Swarm Optimization with multi-wavelet mutation (MWPSO) is proposed to solve the problem. This algorithm makes good use of the available color palette and the mechanism of a half-toning process to derive useful a priori information for the restoration. Simulation results show that it can improve the quality of a half-toned color-quantized image remarkably in terms of both signal-to-noise ratio improvement and convergence rate. The subjective quality of the restored images can also be improved.

Insulin-dependent diabetes mellitus is classified as Type 1 diabetes and it can be further classified as immune-mediated or idiopathic. Hypoglycemia is a common and serious side effect of insulin therapy in patients with Type 1 diabetes. In this chapter, we measure physiological parameters continuously to provide a non-invasive hypoglycemia detection for Type 1 diabetes mellitus (T1DM) patients. Based on the physiological parameters of the electrocardiogram (ECG) signal, such as heart rate, corrected QT interval, change of heart rate and change of corrected QT interval, an evolved fuzzy inference model is developed for classification of hypoglycemia. To optimize the rules and membership functions of the fuzzy system, a hybrid particle swarm optimization with wavelet mutation (HPSOWM) is introduced. For the clinical study, 15 children with Type 1 diabetes are volunteered overnight. All the real data sets are collected from the Department of Health, Government of Western Australia and are randomly organized into a training set (10 patients) and testing set (5 patients). The results show that the evolved fuzzy inference system approach performs well in terms of sensitivity and specificity.

In this chapter, limit cycle behavior of weights of perceptron is discussed. First, the weights of a perceptron are bounded for all initial weights if there exists a nonempty set of initial weights that the weights of the perceptron are bounded. Hence, the boundedness condition of the weights of the perceptron is independent of the initial weights. Second, a necessary and sufficient condition for the weights of the perceptron exhibiting a limit cycle behavior is discussed. The range of the number of updates for the weights of the perceptron required to reach the limit cycle is estimated. Finally, it is suggested that the perceptron exhibiting the limit cycle behavior can be employed for solving a recognition problem when downsampled sets of bounded training feature vectors are linearly separable. Numerical computer simulation results show that the perceptron exhibiting the limit cycle behavior can achieve a better recognition performance compared to a multi-layer perceptron.

Artificial neural network (ANN) is well known for its strong capability to handle nonlinear dynamical systems, and this modeling technique has been widely applied to the nonlinear time series prediction problem. However, this application needs more cautions as overfitting is the serious problem endemic to neural networks. The conventional method of avoiding overfitting is to avoid fitting the data too precisely while it cannot determine the exact model size directly. In this chapter, we employ an alternative information theoretic criterion (minimum description length) to determine the optimal architecture of neural networks according to the equilibrium between the model parameters and model errors. When applied to various time series, we find that the model with the optimal architecture both generalizes well and accurately captures the underlying dynamics. To further confirm the dynamical character of the model residual of the optimal neural network, the surrogate data method from the regime of nonlinear dynamics is then described to analyze such model residual so as to determine whether there is significant deterministic structure not captured by the optimal neural network. Finally, a diploid model is proposed to improve the prediction precision under the condition that the prediction error is considered to be deterministic. The systematic framework composed of the preceding modules is validated in sequence, and illustrated with several computational and experimental examples.

How to efficiently solve eigen-problems of matrices is always a significant issue in engineering. Neural networks run in an asynchronous manner, and thus applying neural networks to address these problems can attain high performance. In this chapter, several recurrent neural network models are proposed to handle eigen-problems of matrices. Each model is expressed as an individual differential equation, with its analytic solution being derived. Subsequently, the convergence properties of the neural network models are fully discussed based on the solutions to these differential equations. Finally, the computation steps are designed toward solving the eigen-problems, with numerical simulations being provided to evaluate the effectiveness of each model.

This chapter consists of three major parts, with each approach in these three parts being in the form of neural networks. Section 7.1 presents how to solve the eigen-problems of real symmetric matrices; Sections 7.2 and 7.3 are devoted to addressing the eigen-problems of anti-symmetric matrices; Section 7.4 aims at solving eigen-problems of general matrices, which are neither symmetric nor anti-symmetric.

Finally, conclusions are made in Section 7.5 to summarize the whole chapter.

Threaded fastenings are a widely used industrial method to clamp components together. The method is successfully employed in many industries including car manufacturing, toy manufacturing and electronic component assembly - the method is popular because it is a low-cost assembly process and permits easy disassembly for maintenance, repair, relocation and recycling. Particularly, the usage of self-tapping screws has gained wide-spread recognition, because of the cost savings that can be achieved when preparing for the fastening process - the components to be clamped together need to be only equipped with holes and as such the more complicated and costly placing of threaded holes as is required for non-self-tapping bolts can be avoided. However, the process of inserting self-tapping screws is more complicated and typically carried out manually. This chapter discusses the study of methods of automation for the insertion of self-tapping screws.

Over the past decade or so, support vector machines have established themselves as a very effective means of tackling many practical classification and regression problems. This chapter relates the theory behind both support vector classification and regression, including an example of each applied to real-world problems. Specifically, a classifier is developed which can accurately estimate the risk of developing heart disease simply from the signal derived from a finger-based pulse oximeter. The regression example shows how SVMs can be used to rapidly and effectively recognize hand-written characters particularly designed for the so-called graffiti character set.

In order to accurately regulate cardiovascular response to exercise for an individual exerciser, this study proposed a control oriented modeling approach to depict nonlinear behavior of heart rate response at both the onset and offset of treadmill exercise. With the aim of capturing nonlinear dynamic behaviors, a well designed exercise protocol has been applied for a healthy male subject. Non-invasively measured variables, such as ECG,, body movements and oxygen saturation (S_pO₂), have been reliably monitored and recorded. Based on several sets of experimental data, both steady state gain and time constant of heart rate response are identified. The nonlinear models relating to steady state gain and time constant (vs. walking speed) were built up based on support vector machine regression (SVR). The nonlinear behaviors at both onset and offset of exercise have been well described by using the established SVR models. The model provides the fundamentals for the optimization of exercise efforts by using model-based optimal control approaches, which is the following step of this study.

This study explores control methodologies to handle time variant behavior for heart rate dynamics at onset and offset of exercise. To achieve this goal, a novel switching model predictive control (MPC) algorithm is presented to optimize the exercise effects at both onset and offset of exercise. Specifically, dynamic matrix control (DMC), one of the most popular MPC control algorithms, has been employed as the essential of the optimization of process regulation while switching strategy has been adopted during the transfer between onset and offset of exercise. The parameters of the DMC/MPC controller have been well tuned based on a previously established SVM-based regression model relating to both onset and offset of treadmill walking exercises. The effectiveness of the proposed modeling and control approach has been shown from the regulation of dynamical heart rate response to exercise through simulation using Matlab.

Heating, Ventilation and Air Conditioning (HVAC) systems are often one of the largest energy consuming parts in modern buildings. Two focused issues of HVAC systems are energy saving and its safety. Regular checking and maintenance are usually the keys to tackle these problems. Due to the high cost of maintenance, preventive maintenance plays an important role. One cost-effective strategy is the development of analytic fault detection and isolation (FDI) modules by online monitoring of the key variables of HVAC systems. This chapter investigates real-time FDI for HVAC systems by using online support vector machine (SVM), by which we are able to train an FDI system with manageable complexity under real-time working conditions. It also proposes a new approach which allows us to detect unknown faults and update the classifier by using these previously unknown faults. Based on the proposed approach, a semi-unsupervised fault detection methodology has been developed for HVAC systems. This chapter also identifies the variables which are the indications of the particular faults we are interested in. Simulation studies are given to show the effectiveness of the proposed online FDI approach.

The following sections are included:

• Index