Advances in Neural Networks – ISNN 2013

In this paper, we aim to solve an infinite-time optimal tracking control problem for a class of discrete-time nonlinear systems using iterative adaptive dynamic programming (ADP) algorithm. When the iterative tracking control law and the iterative performance index function in each iteration cannot be accurately obtained, a new convergence analysis method is developed to obtain the convergence conditions of the iterative ADP algorithm according to the properties of the finite approximation errors. If the convergence conditions are satisfied, it is shown that the iterative performance index functions converge to a finite neighborhood of the greatest lower bound of all performance index functions under some mild assumptions. Neural networks are used to approximate the performance index function and compute the optimal tracking control policy, respectively, for facilitating the implementation of the iterative ADP algorithm. Finally, a simulation example is given to illustrate the performance of the present method.

This paper considers the distributed tracking control of uncertain nonlinear multi-agent systems in the presence of unmeasured states and unknown input of the leader over an undirected network. By approximating the uncertain nonlinear dynamics via neural network and constructing a local observer to estimate the unmeasured states, distributed output feedback tracking controllers, with static and dynamic coupling gains, respectively, are proposed, based on the relative observed states of neighboring agents. It is proved that with the developed controllers, the state of each agent synchronizes to that of the leader for any undirected connected graphs even when only a fraction of the agents in the network have access to the state information of the leader, and the distributed tracking errors are uniformly ultimately bounded. A sufficient condition to the existence of the distributed controllers is that each agent is stabilizable and detectable. Future works include an extension to the directed network topologies.

Quadrotor is a type of rotor craft that consists of four rotors and two pairs of counter-rotating, fixed-pitch blades located at the four corners of the body. The flight control parameters optimization is one of the key issues for quadrotor. Estimation of distribution algorithm is a new kind of evolutionary algorithm developed rapidly recently. However, low convergence speed and local optimum of the EDA are the main disadvantages that limit its further application. To overcome the disadvantages of EDA, a chaotic estimation of distribution algorithm is proposed in this paper. It is a combination of chaos theory and principles of estimation of distribution algorithm. Series of experimental comparison results are presented to show the feasibility, effectiveness and robustness of our proposed method. The results show that the proposed chaotic EDA can effectively improve both the global searching ability and the speed of convergence.

This technical note introduces stability analysis on pattern-based neural network (NN) control systems. Firstly, different control situations are defined as dynamical patterns and are identified via deterministic learning (DL). When the dynamical pattern is correctly classified, the corresponding NN learning controller with knowledge or experience is selected. Secondly, by adopting a class of switching signals with average dwell time (ADT) property , it is shown that the NN learning controller can achieve small tracking errors and fast convergence rate with small control gains. These results will guarantee not only stability of the closed-loop systems, but also better performance in the aspects of time saving or energy saving. Finally, the theoretical analysis is supported by simulations.

This paper investigates the observer-based

H

 ∞ 

control problem for a class of discrete-time Takagi-Sugeno fuzzy neural networks with both random communication packet losses. The random data losses are described by a Bernoulli distributed white sequence that obeys a conditional probability distribution. In the presence of random packet losses, sufficient conditions for the existence of an observer-based feedback controller are derived, such that the closed-loop control system is asymptotically mean-square stable and preserves a guaranteed

H

 ∞ 

performance. Finally, a numerical example is provided to illustrate the effectiveness of the developed theoretical results.

In this paper, a novel control design is proposed for wheeled inverted pendulum with input saturation. Based on Lyapunov synthesis method, backstepping design procedure and the Neural network (NN) approximation to the uncertainty of the system, the adaptive NN tracking controller is constructed by considering actuator saturation constraints. The stability analysis subject to the effect of input saturation constrains are conducted with the help of an auxiliary design system. The proposed controller guarantees uniformly ultimately bounded of all the signals in the closed-loop system, while the tracking error can be made arbitrarily small. Simulation studies are given to illustrate the effectiveness and the performance of the proposed scheme.

In this paper, we present an improved learning scheme for extracting T-S fuzzy rules from data samples, whereby a neuro-fuzzy architecture implements the T-S fuzzy system using ellipsoidal basis functions. The salient characteristics of this approach are as follows: 1) A novel structure learning algorithm incorporating a pruning strategy into new growth criteria is developed. 2) Compact fuzzy rules can be extracted from training data. 3) The linear least squares (LLS) method is employed to update consequent parameters, and thereby contributing to high approximation accuracy. Simulation studies and comprehensive comparisons with other well-known algorithms demonstrate the effective and superior performance of our proposed scheme in terms of compact structure and promising accuracy.

This paper deals with the functional equivalence between Generalized Ellipsoidal Basis Function based Neural Networks (GEBF-NN) and T-S fuzzy systems. Significant contributions are summarized as follows. 1) The GEBF-NN is equivalent to a T-S fuzzy system under the condition that the GEBF unit and the local model correspond to the premise and the consequence of the T-S fuzzy system. 2) The normalized (nonnormalized) GEBF-NN is equivalent to a normalized (nonnormalized) T-S fuzzy system using dissymmetrical Gaussian functions (DGF) as univariate membership functions and local models as consequent parts. 3) The equivalence between the normalized GEBF-NN and the nonnormalized T-S fuzzy system is established by employing GEBF units as multivariate membership functions of fuzzy rules. 4) These theoretical results would not only fertilize the learning schemes for fuzzy systems but also enhance the interpretability of neural networks, and thereby contributing to innovative neuro-fuzzy paradigms. Finally, numerical examples are conducted to illustrate the main results.

In this paper, an adaptive neural network (NN) control scheme is proposed for a class of strict-feedback discrete-time nonlinear systems with input saturation. which is designed via backstepping technology and the approximation property of the HONNs, aimed to solve the the input saturation constraint and system uncertainty in many practical applications. The closedloop system is proven to be uniformly ultimately bounded (UUB). At last, a simulation example is given to illustrate the effectiveness of the proposed algorithm.

In this brief, we consider the problem of tracking a desired trajectory for fully actuated autonomous underwater vehicle (AUV), in the presence of external disturbance and model errors. Based on the backstepping method and Lypunov stability theorem, we introduce the dynamic surface control (DSC) technique to tackle the problem of “explosion of complexity” which existing in the traditional backstepping algorithm. Furthermore, the norm of the ideal weighting vector in neural network (NN) systems is considered as the estimation parameter, such that only one parameter is adjusted. The proposed controller guarantees uniform ultimate bounded (UUB) of all the signals in the closed-loop system, while the tracking error converges to a small neighborhood of the origin. Finally, simulation studies are given to illustrate the effectiveness of the proposed algorithm.

This paper proposes to use a state feedback method to control the Hopf bifurcation for a fractional order Hindmarsh-Rose neuronal model. The order of the fractional order Hindmarsh-Rose neuronal model is chosen as the bifurcation parameter. The analysis shows that in the absences of the state feedback controller, the fractional order model loses stability via the Hopf bifurcation early, and can maintain the stability only in a certain domain of the gain parameter. When applying the state feedback controller to the model, the onset of the undesirable Hopf bifurcation is postponed. Thus, the stability domain is extended, and the model possesses the stability in a larger parameter range. Numerical simulations are given to justify the validity of the state feedback controller in bifurcation control.

This paper considers the problem of adaptive neural decentralized control for pure-feedback nonlinear interconnected large-scale systems. Radical basis function (RBF) neural networks are used to model packaged unknown nonlinearities and backstepping is used to construct decentralized controller. The proposed control scheme can guarantee that all the signals in the closed-loop system are semi-globally uniformly ultimately bounded. A numerical example is provided to illustrate the effectiveness of the suggested approach.

In this paper, the adaptive synchronization problem is investigated for a class of uncertain chaotic systems. By using the RBF networks to approximation unknown functions of the master system, an adaptive neural synchronization scheme is proposed with the combination of backstepping technique and dynamic surface control (DSC). This proposed method, similar to backstepping but with an important addition, can overcome the “explosion of complexity” of the traditional backstepping by introducing a first-order filtering. Thus, the closed-loop stability and asymptotic synchronization can be achieved. Finally, simulation results are presented to illustrate the effectiveness of the approach.

Based on the dynamic surface control (DSC), an adaptive neural network control approach is proposed for a class of stochastic nonlinear strictfeedback systems in this paper. This approach simplifies the backstepping design and overcomes the problem of ’explosion of complexity’ inherent in the backstepping method. The Lyapunov Stability analysis given in this paper shows that the control law can guarantee the solution of the closed-loop system uniformly ultimate boundedness (UUB) in probability. The simulation example is given to illustrate the effectiveness of the proposed control system.

In this paper, we focus on the cooperative tracking problem of multi-agent systems with nonlinear dynamics and communication time delays. Only a portion of agents can access the information of the desired trajectory and there are communication delays among the agents. Through designing an adaptive neural network based control law and constructing an appropriate Lyapunov-Krasovskii functional, it is proved that the tracking error of each agent converges to a neighborhood of zero. Simulation results are provided to show the effectiveness of the proposed algorithm.

This paper presents an excellent software phase-locked loop speed control system of permanent magnet synchronous motor (PMSM). A loop-gain adaptation scheme is developed using model reference adaptive system (MRAS) theory to suppress the torque disturbance which effect on motor speed. The following three points including accurate steady-state speed, fast transient response, and insensitivity to disturbance are especially important for speed control of permanent magnet synchronous motor. The software phase-locked loop (SPLL) technique has the significant ability to obtain precise speed regulation. When the feedback signal of the motor speed is synchronized with a reference signal, perfect speed regulation can be realized. The steady-state accuracy is about 0.02%~0.1% which is difficult to be obtained by conventional proportion integral differentiation (PID) speed control. But phase-locked loop system suffers from pool dynamics and limited lock range. The gain of SPLL has great effect on the performance of system. As the loop gain becomes larger, both the maximum speed error and load increase. If the loop gain varies according to the values of phase error and speed error, the SPLL system will be adaptive between the accuracy and sensitivity to load disturbance. A model reference adaptive system is designed to confine the transient phase error within the range of [-2

π

2

π

] at the present of torque disturbance. This means that the SPLL remains phase tracking. Also, in order to overcome the time varying and nonlinear of PMSM, and obtain the stable torque output, it is effective to utilize the neuron to seek the optimum controller parameter on line. Experiment results are presented to verify the validity of the proposed system.

In this paper, the global tracking control problem for a class of wheeled mobile robots is considered and a new adaptive position tracking control scheme is proposed where radial basis function (RBF) neural network (NN) is utilized to model the uncertainty. The feedback compensation scheme is obtained, where the information of reference position and real position of robot are both used as the NN input. Compered with the existing results, the main advantage is that the global stability of the closed-loop system can be ensured and the NN approximation domain can be determined based on the reference signal a prior. Finally, a simulation example is provided to demonstrate the effectiveness of the proposed control scheme.

In this paper, the adaptive control problem with input saturation is investigated for double inverted pendulums. Based on Lyapunov stability theory and backstepping technique, incorporating dynamic surface control (DSC) technique into neural network based adaptive control, an adaptive neural controller is developed by explicitly considering uncertainties, unknown disturbances and input saturation. An auxiliary system is presented to tackle input saturation, and the states of auxiliary design system are utilized to develop the tracking control. It is proved that all the signals in the closed-loop system are uniformly ultimately bounded (UUB) via Lyapunov analysis. Finally, simulation studies are given to demonstrate the effectiveness of the proposed method.

The control system of the permanent magnet synchronous motor (PMSM) has the characteristics of nonlinear and strong coupling. Therefore, In order to improve the control precision, the paper presents a novel approach of speed control for PMSM using adaptive BP (back-propagations)-PID neural network. The approach consists of two parts: on-line identification based on BP neural network and the adaptive PID controller. Lyapunov theory is used to prove the stability of the control scheme. Simulation results show that this control method can improve the dynamical performance and enhance the static precision of the speed system.

An adaptive control algorithm is applied to controlling a class of SISO continuous stirred tank reactor (CSTR) system in discrete-time. The considered systems belong to pure-feedback form where the unknown dead-zone and it is first to control this class of systems. Radial basis function neural networks (RBFNN) are used to approximate the unknown functions and the mean value theorem is exploited in the design. Based on the Lyapunov analysis method, it is proven that all the signals of the resulting closed-loop system are guaranteed to be semi-global uniformly ultimately bounded (SGUUB) and the tracking error can be reduced to a small compact set. A simulation example is studied to verify the effectiveness of the approach.

By employing Radial Basis Function (RBF) Neural Networks (NN) to approximate uncertain functions, an application-oriented adaptive neural networks design for ship linear-tracking control was brought in based on dynamic surface control (DSC) and minimal-learning-parameter (MLP) algorithm. With less learning parameters and reduced computation load, the proposed algorithm can avoid the possible controller singularity problem and the trouble caused by ”explosion of complexity” in traditional backstepping methods is removed, so it is convenient to be implemented in applications. In addition, the boundedness stability of the closed-loop system is guaranteed and the tracking error can be made arbitrarily small. Simulation results on ocean-going training ship ’YULONG’ are shown to validate the effectiveness and the performance of the proposed algorithm.

This paper studies the adaptive synchronization problem for a kind of stochastic Markovian jump neural networks with mode-dependent and unbounded distributed delays. By virtue of the Lyapunov stability theory and the stochastic analysis technique, a generalized LaSalle-type invariance principle for stochastic Markovian differential delay equations is utilized to investigate the globally almost surely asymptotical stability of the error dynamical system in the mean-square sense.

In this paper, an

H

 ∞ 

optimal tracking control scheme based on generalized Hamilton-Jacobi-Isaacs (GHJI) equation is developed for discrete-time (DT) affine nonlinear systems. First, via system transformation, the optimal tracking problem is transformed into an optimal regulation problem with respect to the state tracking error. Second, with regard to the converted regulation problem, in order to obtain the

H

 ∞ 

tracking control, the corresponding GHJI equation is formulated, and then the

L

2

-gain analysis of the closed-loop nonlinear system are employed. Third, an iterative algorithm based on the GHJI equation by using neural networks (NNs) is introduced to solve the optimal control. Finally, simulation results are presented to demonstrate the effectiveness of the proposed scheme.

In this paper, the adaptive neural attitude control is developed for near-space vehicles with the oblique wing (NSVOW) via using the sliding mode disturbance observer technique. The radial basis function neural network (RBFNN) is employed to approximate the unknown system uncertainty. Then, the sliding mode disturbance observer is designed to estimate the unknown external disturbance and the unknown neural network approximation error. Using outputs of the sliding mode disturbance observer and the RBFNN, the adaptive neural attitude control is proposed for NSVOWs. The stability of the closed-loop system is proved using the Lyapunov analysis. Finally, simulation results are presented to illustrate the effectiveness of the proposed adaptive neural attitude control scheme.

The paper presents the electromyogram (EMG)-based neural network control of an upper-limb power-assist exoskeleton robot, which is proposed to control the robot in accordance with the user’s motion intention. The upper limb rehabilitation exoskeleton is with high precision for co-manipulation tasks of human and robot because of its backdrivability, precise positioning capabilities, and zero backlash due to its harmonic drive transmission (HDT). The novelty of this work is the development of an adaptive neural network modeling and control approach to handle the unknown parameters of the harmonic drive transmission in the robot to facilitate motion control. We have conducted the experiments on human subject to identify the various parameters of the harmonic drive system combining sEMG information signals.

Based on the backstepping method and the neural networks (NNs) technique, a direct adaptive controller is proposed for a class of nonlinear fin stabilizer system in this paper. This approach overcomes the uncertainty in the nonlinear fin stabilizer system and solves the problems of mismatch and controller singularity. The stability analysis shows that all the signals of the closed-loop system are uniformly ultimate boundedness (UUB). A simulation example is given to illustrate the effectiveness of the proposed method.

In this paper, output feedback direct adaptive robust NN control is investigated for a class of nonlinear discrete-time systems in strict-feedback form. To construct output feedback control, the original strict-feedback system is transformed into a cascade form, which the output feedback of the nonlinear discrete-time system can be carried out. Then with employment of the inputs and outputs, the output feedback direct adaptive robust NN control is developed. The HONNs is exploited to approximate unknown function, and a stable adaptive NN controller is synthesized. The proposed algorithm improves the rubostness of the discrete-time nonlinear systems. It is proven that all the signals in closed-loop system are uniformly ultimately bounded (UUB). A simulation example is presented to illustrate the effectiveness of the proposed algorithm.

To reduce the roll of a surface ship, a robust fin controller based on functional-link neural networks is proposed. The plant consists of the ship roll dynamics and that of the fin actuators. Modeling errors and the environmental disturbance induced by waves are considered in the cascaded roll system, which are identified by the neural networks. Lyapunov function is employed in the controller design, which guarantees the stability of the fin stabilizer. Numerical simulation demonstrates the good performance of the roll reduction based on the controller proposed.

In this paper, by employing Radial Basis Function (RBF) Neural Networks (NN) to approximate uncertain functions, the robust adaptive neural networks design for a class of SISO systems was brought in based on dynamic surface control (DSC) and minimal-learning-parameter (MLP) algorithm. With less learning parameters and reduced computation load, the proposed algorithm can avoid the possible controller singularity problem and the trouble caused by "explosion of complexity" in traditional backstepping methods is removed, so it is convenient to be implemented in applications. In addition, it is proved that all the signals of the closed-loop system are uniformly ultimately bounded(UUB), and simulation results on ocean-going training ship ’YULONG’ are shown to validate the effectiveness and the performance of the proposed algorithm.

Based on the data-driven modeling theory, the integrated modeling and intelligent control method of the coke oven collector pressure is carried out in the paper. The system includes the regression predictive model of coke oven global collector pressure based on support vector machine (SVM), the subtractive clustering algorithm based operation pattern extraction and migration reconfiguration strategy and the self-tuning PID decoupling controller based on the improved glowworm swarm optimization (GSO) algorithm of the coke oven collector pressure. Simulation results and industrial application experiments clearly show the feasibility and effectiveness of control methods and satisfy the real-time control requirements of the coke oven collector pressure system.

The problem of ship linear path-keeping control is discussed. By employing radial based function neural network (RBF NN) to approximate uncertain nonlinear system functions, and by combining dynamic surface control (DSC) with backstepping technique and Nussbaum gain approach, the algorithm can not only overcome both the “explosion of complexity” problem inherent in the backstepping method and the possible “controller singularity” problem, but also reduce dramatically the number of on-line learning parameters, thus the algorithm can reduce the computation load of the algorithm correspondingly and make it easy in actual implementation. The stability analysis shows that all closed-loop signals will be semi-global uniformly ultimately bounded (SGUUB), when the tracking error converge to a small neighborhood around the origin through appropriately choosing design constants. Finally, simulation results are presented to show the effectiveness of the proposed algorithm.

In this paper, the problem of adaptive neural output feedback control is investigated for a class of uncertain nonlinear pure feedback systems with unknown backlash-like hysteresis. In the design, RBF neural networks are used to approximate the nonlinear functions of systems, and a neural state observer is designed to estimate the unmeasured states. By utilizing the neural state observer, and combining the backstepping technique with adaptive control design, an observer-based adaptive neural output feedback control approach is developed. It is proved that the proposed control approach can guarantee that all the signals in the closed-loop system are semi-globally uniformly ultimately bounded (SUUB), and both observer error and tracking error can converge to a small neighborhood of the origin.

To solve the path following problem of underactuated surface ships with internal dynamic uncertainties and external disturbances, an Active-Disturbance-Rejection Control (ADRC) controller is introduced to steer the ship to follow the desired path. Drift angle compensation is added to the controller by designing a coordinate transformation equation. The cross track static error caused by wind and current is overcome. Simulations were carried out on a fully nonlinear hydrodynamic model of a training ship to validate the stability and excellent robustness of the proposed controller.

Canonical correlation analysis (CCA) is a promising feature extraction technique of steady state visual evoked potential (SSVEP)-based brain computer interface (BCI). Many researches have showed that CCA performs significantly better than the traditional methods. In this paper, the neural network implementation of CCA is used for the frequency detection and classification in SSVEP-based BCI. Results showed that the neural network implementation of CCA can achieve higher classification accuracy than the method of power spectral density analysis (PSDA), minimum energy combination (MEC) and similar performance to the standard CCA method.

Common Spatial Pattern (CSP) and Support Vector Machine (SVM) are usually adopted for feature extraction and classification of two-class motor imagery. However, in a motor imagery based BCI-FES rehabilitation system, stroke patients usually are not able to conduct correct motor imagery like healthy people due to the injury of motor cortex. Therefore, motor imagery EEG of stroke patients lacks of specific discriminant features as appearances of healthy people, which significantly blocks CSP to seek the optimal projection subspace. In this paper, a method, which filters EEG into a variety of bands and improves performance through boosting principle based on a set of weak CSP-SVM classifiers, was proposed to solve the problem mentioned above and was evaluated on the EEG datasets of three stroke subjects. The proposed method outperformed the traditional CSP-SVM method in terms of classification accuracy. From data analysis, we observed that optimal spectral band for classification had been changing along with rehabilitation training, which may reveal mechanisms that dominant frequency band may be changed along with rehabilitation training and spectral power distribution may be changed in different stages of rehabilitation. In addition, this work also demonstrated the feasibility of our SJTU-BCMI BCI-FES rehabilitation training system.

From the viewpoint of image processing, a spectral feature-based TLS (Tikhonov-regularized least-squares) ensemble algorithm is proposed for tumor classification using gene expression data. In the TLS model, a test sample is represented as a linear combination of atoms of an overcomplete dictionary. Two types of dictionaries, spectral feature-based eigenassays and spectral feature-based metasamples, are proposed for the TLS model. Experimental results on standard databases demonstrate the feasibility and effectiveness of the proposed method.

P300 is a popular characteristic potential for electroencephalogram(EEG) based brain-computer interface(BCI). In P300-BCI, the extraction of P300 is a very crucial operation. Independent component analysis(ICA) technique is suitable for P300 extraction. In this paper, aiming at the current large volume of EEG data, the applications of three ICA algorithms were proposed for P300 extraction and were compared. The experiments ran on real EEG data respectively. PI and recognition accuracy were checked. The results show artificial fish swarm algorithm based ICA(AFSA_ICA) can extract P300 faster, reducing the computation time for BCI with PI remaining better.

A crucial problem for the overall performance of steady-state visual evoked potentials (SSVEP)-based brain computer interface (BCIs) is the right choice of the time-window length since a large window results in a higher accuracy but longer detection time, making the system impractical. This paper proposes an adaptive time window length to improve the system performance based on the subject’s online performance. However, since there is no known methods of assessing the online performance in real time, it is also proposed a feedback from the user, through a speller, for the system to know whether the output is correct or not and change or maintain the time-window length accordantly. The system was implemented fully online and tested in 4 subjects. The subjects have attained an average information transfer rate (ITR) of 62.09bit/min and standard deviation of 2.13bit/min with a mean accuracy of 99.00% and standard deviation of 1.15%, which represents an improvement of about 6.50% of the ITR to the fixed time-window length system.

In this paper, we propose a novel geodesic distance based clustering approach for delineating boundaries of touching cells. In specific, the Riemannian metric is firstly adopted to integrate the spatial distance and intensity variation. Then the distance between any two given pixels under this metric is computed as the geodesic distance in a propagational way, and the K-means-like algorithm is deployed in clustering based on the propagational distance. The proposed method was validated to segment the touching Madin-Darby Canine Kidney (MDCK) epithelial cell images for measuring their N-Ras protein expression patterns inside individual cells. The experimental results and comparisons demonstrate the advantages of the proposed method in massive cell segmentation and robustness to the initial seeds selection, varying intensity contrasts and high cell densities in microscopy images.

Considering the EEG signals are nonlinear and nonstationary, the nonlinear dynamical methods have been widely applied to analyze the EEG signals. Directly extracted the approximate entropy and sample entropy as features are efficient methods to analysis the EEG signals of epileptic parents. To detect the epilepsy seizure signals from epileptic EEG, choose an appropriate threshold value as the discrimination criteria is simplest. The experiment indicated the approximate entropy provide a higher accuracy in distinguishing the epileptic seizure signals from the EEG than sample entropy. To improve the accuracy of sample entropy, empirical mode decomposition (EMD) is used to decompose EEG into multiple frequency subbands, and then calculate sample entropy for each component. The results show that the accuracy is up to 91%, which could be used to discriminate epileptic seizure signals from epileptic EEG.

A major drawback of artificial neural network is long training time depending on a number of training data. Thus, the contribution of this work is to present the intelligent hybrid system for faster training on neural network. The concept of the proposed method is applying DBSCAN for removing noise and outliers then selecting the represented instances to form a smaller training set for further model training. The experimental results indicate that the proposed method can dramatically reduce a size of training set while the predictive performance of the classifiers are better or almost the same as models trained with original training sets.

Estimation of distribution algorithms (EDAs) are a class of evolutionary optimization algorithms based on probability distribution model. This article extends the basic EDAs for tackling multi-objective optimization problems by incorporating multivariate Gaussian copulas for constructing probability distribution model, and using the concept of preference order. In the algorithm, the multivariate Gaussian copula is used to construct probability distribution model in EDAs. By estimating Kendall’s

τ

and using the relationship of correlation matrix and Kendall’s

τ

, correlation matrix R in Gaussian copula are firstly estimated from the current population, and then is used to generate offsprings. Preference order is used to identify the best individuals in order to guide the search process. The population with the current population and current offsprings population is sorted based on preference order, and the best individuals are selected to form the next population. The algorithm is tested to compare with NSGA-II, GDE, MOEP and MOPSO based on convergence metric and diversity metric using a set of benchmark functions. The experimental results show that the algorithm is effective on the benchmark functions.

According to the dynamic, distribution and complexity of cloud computing, resource scheduling effectively with users’ QoS demand and achieving maximum benefit is the unprecedented challenge. To solve the above problem, we propose to use genetic algorithm: design for the crossover operator and build a cloud resource optimization scheduling model that promised to address user needs while optimizing resource allocation. With the experiments, this paper verifies the superiority of models made in this paper. The results show that the use of genetic algorithm to optimize cloud resource scheduling has the rationality and feasibility. Meanwhile, using the genetic algorithm is useful for effectively scheduling of cloud resource meeting the users’ QoS.

This paper investigates the effectiveness of the Genetic Algorithm (GA) and Simulated Annealing algorithm (SA) training artificial neural network weights and biases for rainfall forecasting, namely GAS–ANN. Firstly, a hybrid GA and SA method is used to train the begining connection weights and thresholds of ANN. Secondly, the back propagation algorithm is used to search around the global optimum. Finally, a numerical example of monthly rainfall data in a catchment located in a subtropical monsoon climate in Linzhou of China, is used to elucidate the forecasting performance of the proposed GASA–ANN model. The forecasting results indicate that the proposed model yields more accurate forecasting results than the autoregressive integrated moving average (ARIMA), back–propagation neural network (BP–NN) and pure Genetic Algorithm training Artificial Neural Network model (GA–ANN). Therefore, the GASA–ANN model is a promising alternative for rainfall forecasting.

The nonlinear constrained optimization problems have been widely used in many fields, such as engineering optimization and artificial intelligence. According to the deficiency of artificial fish swarm algorithm (AFSA), that the artificial fishes walk around aimlessly and randomly or gather in non-global optimal points, a hybrid algorithm-artificial fish swarm optimization algorithm based on mixed crossover strategy is presented. By improving the artificial fish’s behaviors, the genetic operation of mixed crossover strategy is used as a local search strategy of AFSA. So the efficiency of local convergence of AFSA is improved, and the algorithm’s running efficiency and solution quality are improved obviously. Based on test verification for typical functions, it is shown that the hybrid algorithm has some better performance such as fast convergence and high precision.

This paper presents a new network-based classification technique using limiting probabilities from random walk theory. Instead of using a traditional heuristic to classify data relying on physical features such as similarity or density distribution, it uses a concept called ease of access. By means of an underlying network, in which nodes represent states for the random walk process, unlabeled instances are classified with the label of the most easily reached class. The limiting probabilities are used as a measure for the ease of access by taking into account the biases provided by an unlabeled instance in a specific adjacency matrix weight composition. In this way, the technique allows data classification from a different viewpoint. Simulation results suggest that the proposed scheme is competitive with current and well-known classification algorithms.

In the next generation of recommender systems, multi- criteria recommendation could be regarded as one of the most important branches. Compared with traditional recommender systems with usually one single rating, multi-criteria recommender systems have several ratings from different aspects, and generally describe users’ interests more accurately. However, owing to the cost of ratings, multi-criteria recommender systems meet more severe data sparsity problem than traditional single criteria recommender systems.

In this paper, We design a new approach to compute the similarity between users, which tackles the challenge posed by data sparsity that one cannot obtain the similarity between users with no common rated items. With a new method of data preprocessing, the features of items are combined to eliminate the effect of noise and evaluation scale. We model the aggregation function using support vector regression which is more accurate and robust than linear regression. The experiments demonstrate that our method produces a better performance, while providing more powerful suitability on sparse and noisy datasets.

The semi-supervised learning paradigm allows that a large amount of unlabeled data be classified using just a few labeled data. To account for the minimal

a priori

label knowledge, the information provided by the unlabeled data is also used in the classification process. This paper describes a semi-supervised technique that uses random walk limiting probabilities to propagate label information. Each label is propagated through a network of unlabeled instances via a biased random walk. The probability of a vertex receiving a label is expressed in terms of the limiting conditions of the walk process. Simulations show that the proposed technique is competitive with benchmarked techniques.

The increasing on the human ability to gather data has led to an increasing effort on labeling them to be used in specific applications such as classification and regression. Therefore, automatic labeling methods such as semi-supervised transdutive learning algorithms are of a major concern on the machine learning and data mining community nowadays. This paper proposes a graph-based algorithm which uses the purity measure to help spreading the labels throughout the graph. The purity measure determines how intertwined are different subspaces of data regarding its classes. As high values of purity indicate low mixture among patterns of different classes, its maximization helps finding well-separated connected subgraphs; which facilitates the label spreading process. Results on benchmark data sets comparing to state-of-the-art methods show the potential of the proposed algorithm.

Traditional data stream classification techniques are not capable of recognizing new classes emerged in data stream. Recently, an ensemble classification framework focused on the new challenge. But the novel class detection technique is limited to the numeric data in the framework. And, both the lower process speed and the larger model size of base classifier trouble the framework. In this paper, a novel class instance detection technique is proposed to deal with mixed attribute data and the VFDTc is adopted as base classifier to speed up the process and reduce the model size. Experimental results showed that the algorithm outperformed the previous one in both classification accuracy and processing speed.

Spatial pyramid match scheme (SPM) is an important scheme in local feature based image classification which effectively adopts geometric structure information into image classification. Most previous approach formulized Spatial Pyramid in unsupervised manner by hierarchical splitting images into separate bins. We found that weak supervised information exists in this process totally unused. We cannot use information directly, because those information corresponding to the combination of all bins, thus we can use those weakly supervised information for bins selection. In this paper, we proposed to select those bins with better discriminative properties .The discriminative property can be well defined from neighborhood entropy. We incorporate local sensitive hash for fast neighborhood identification. We set those bins with higher neighborhood entropy weight zero. Analysis shows that our approach can down weight those non-discriminative bins, in contrast highlighting those discriminative bins. Experiments show that our approach can improve the performance of spatial pyramid match, especially for those categories with complex background. We also proof that under our scheme, result kernel matrix can still preserve positive semi-definite, which can guarantee that our algorithm will coverage.

This paper aims to verify the capability of fuzzy inference system in establishing time series model for ship manoeuvrability. The traditional modeling approaches are usually based on a unified framework. Due to the presence of outliers or noises in ship sailing records, it is difficult in achieving satisfactory performance directly from data. In this paper, we propose a combined time series modeling method by the use of data mining technique and fuzzy system theory. Data mining concepts are introduced to improve the fuzzy rule extraction algorithm to make the resulting fuzzy inference system more robust with respect to the noises or outliers. A ship 20°/20° zig-zag test is simulated. The data point records in time series are obtained from an actual manoeuvring test. With comprehensive robustness analysis, our fuzzy inference system using data mining technology is proved to be a robust and accurate tool for ship manoeuvring simulation.

In this work we utilize the social messages to construct an extensible event ontology model for learning the experiences and knowledge to cope with emerging real-world events. We develop a platform combining several text mining and social analysis algorithms to cooperate with our stream mining approach to detecting large-scale disastrous events from social messages, in order to achieve the aim of automatically constructing event ontology for emergency response First, we employ the developed event detection technique on Twitter social-messages to monitor the occurrence of emerging events, and record the development and evolution of detected events. Furthermore, we store the messages associated with the detected events in a repository. Through the developed algorithms for analyzing the content of social messages and ontology construction the event ontology can be established, allowing for developing relevant applications for prediction of possible evolution and impact evaluation of the events in the future immediately, in order to achieve the goals for early warning of disasters and risk management.

In this paper, we provide a revised inference for correlated topic model (CTM) [3]. CTM is proposed by Blei et al. for modeling correlations among latent topics more expressively than latent Dirichlet allocation (LDA) [2] and has been attracting attention of researchers. However, we have found that the variational inference of the original paper is unstable due to almost-singularity of the covariance matrix when the number of topics is large. This means that we may be reluctant to use CTM for analyzing a large document set, which may cover a rich diversity of topics. Therefore, we revise the inference and improve its quality. First, we modify the formula for updating the covariance matrix in a manner that enables us to recover the original inference by adjusting a parameter. Second, we regularize posterior parameters for reducing a side effect caused by the formula modification. While our method is based on a heuristic intuition, an experiment conducted on large document sets showed that it worked effectively in terms of perplexity.

Adaptive fault estimation problems of coupling connections for a class of complex networks have been studied in the concept of drive-response synchronization. By constructing a suitable response networks and designing the adaptive control laws and adaptive estimator of coupling connections, the coupling connections in drive networks can be estimated and correspondingly monitored online, which can be as indicators to judge whether a fault connection occurs or not. Meanwhile, when the actuator fault occurs, a passive fault tolerant controller is designed to guarantee the synchronization between drive and response networks.

This paper presents a model of Elman recurrent neural network (ERNN) for time series fault prediction in semiconductor etch equipment. ERNN maintains a copy of previous state of the input in its context units, as well as the current state of the input. Derivative dynamic time warping (DDTW) method is also discussed for the synchronization of time series data set acquired from plasma etcher. For each parameter of the data, the best ERNN structure was selected and trained using Levenberg Marquardt to generate one-step-ahead prediction for 10 experimental runs. The faulty experimental runs were successfully distinguished from healthy experimental runs with one missed alarm out of ten experimental runs.

Human behaviors monitoring by using wireless sensor networks has gained tremendous interest in recent years from researchers in many areas. To distinguish behaviors from on-body sensor signals, many classification methods have been tried, but most of them lack the relearning ability, which is quite important for long-term monitoring applications. In this paper, a relearning probabilistic neural network is proposed. The experimental results showed that the proposed method achieved good recognition performance, as well as the relearning ability.

Recently, by following Zhang et al.’s design method, a special class of recurrent neural network (RNN), termed Zhang neural network (ZNN), has been proposed, generalized and investigated for solving time-varying problems. In the design procedure of ZNN models, choosing a suitable kind of error function [i.e., the so-called Zhang function (ZF) used in the methodology] plays an important role, and different ZFs may lead to various ZNN models. Besides, differing from other error functions such as nonnegative energy functions associated with the conventional gradient-based neural network (GNN), the ZF can be positive, zero, negative, bounded, or unbounded even including lower-unbounded. In this paper, different newly-designed ZNN models are proposed, developed and investigated to solve the problem of time-varying underdetermined systems of linear equations (TVUSLE) based on different ZFs. Computer-simulation results (including the robotic application of the newly-designed ZNN models) show that the effectiveness of the proposed ZNN models is well verified for solving such time-varying problems.

In this paper we presents an observer based on artificial neuro-fuzzy networks approach to estimate the indicated torque of a diesel engine from crank shaft angular position and velocity measurements. These variables can be measured by low-cost sensors, since the indicated torque is an important signal for monitoring and/or control of a diesel engine; however, it is not practical to measure it, due to it is not easily measured and need expensive sensors. A model of average value of a diesel engine is used in the simulation to test the estimator of the indicated torque, these results are presented. This estimator may be useful in the implementation of control strategies or diagnostic where the indicated torque measurements are required.

For intermittent channel output signal, namely, the active periods followed by nonactive periods alternatively, the blind single-input single-output (SISO) system identification problem can be transformed into a blind multichannel identification problem. It is possible and feasible to blindly identify the channel using only second-order statistics from the channel output signal. A two-stage approach is proposed in this paper. At the first stage, two or more segments of channel input signal are estimated from the single channel observation; at the second stage, the channel impulse response is identified by exploiting the estimated channel input signal segments and their corresponding channel output signal segments. Simulations show that the proposed approach works well.

The determination of concrete mix ratio is known as the concrete mix design which involves many theories and practice knowledge and must satisfy some requirements. In order to get high performance concrete, the mix design should be tuned using optimization. However, a lot of concrete experiments are needed to correct models which are very time-consuming and expensive. In this paper, a neural network surrogate model based method is proposed to optimize concrete mix design. This approach focuses on the optimization of compressive strength. Experimental results manifest that the optimum design which achieves high compressive strength can be found by employing the novel approach.

The boom of Flickr, a photo-sharing social tagging system, leads to a dramatical increasing of online social interactions. For example, it offers millions of groups for users to join in order to share photos and keep relations. However, the rapidly increasing amount of groups hampers users’ participation, thus it is necessary to suggest groups according to users’ preferences. By analyzing user-generated tags, one can explore users’ potential interests, and discover the latent topics of the corresponding groups. Furthermore, users’ behaviors are affected by their friends. Based on these intuitions, we propose a topic-based group recommendation model to predict users’ potential interests and conduct group recommendations based on tags and social relations. The proposed model provides a way to fuse tag information and social network structure to predict users’ future interests accurately. The experimental results on a real-world dataset demonstrate the effectiveness of the proposed model.

For granular computing in discrete space, the inclusion relation between two granules is partially ordered. How to measure the fuzzy inclusion degree is one of the key issues. We proposed a fuzzy inclusion relation between two hyperbox granules using an inclusion measure function based on a linear positive valuation function induced by the longest diagonal a hyperbox granule. The fuzzy algebraic system is formed by the granule set and fuzzy inclusion relation between two granules, and used to guide the design of algorithms in granular computing.

Recently, Zhang dynamics (ZD) and gradient dynamics (GD) have been used frequently to solve various kinds of online problems. In this paper, the output tracking of time-varying linear (TVL) systems is considered. Then, for such a problem, three different types of tracking controllers (i.e., Z0G0, Z1G0 and Z1G1 controllers) are designed by exploiting the ZD and GD methods. Simulation results on different TVL systems show that such three types of controllers can be feasible and effective for the output-tracking problem solving. Especially, the Z1G1 controller is capable of conquering the control-singularity of systems.

As a successful and effective technique, recommendation systems have been widely studied. Recently, with the popularity of social networks, some researchers have proposed the social recommendation, which considers the social relations between users besides the rating data. However, in real world scenarios, both the social relations and ratings are very sparse, how to combine them together to improve the performance becomes a critical issue. To that end, in this paper, we propose a unified three-stage recommendation framework named

R

andom

W

alk

N

eighborhood-aware

M

atrix

F

actorization(RWNMF), which can effectively integrate the social and rating data together and alleviate the sparsity problem. Specifically, we first perform random walk on social graph to find potential neighbors of each user, then select behavioral neighbors based on the rating data. Lastly, both the social neighbors and behavioral neighbors can be incorporated into traditional SocialMF, leading to more accurate recommendations. Experimental results on Epinions and Flixster datasets demonstrate our approach outperforms the state-of-the-art algorithms.

In this paper we present Artificial Fish Swarm Algorithm (AFSA) applying to a two-dimensional non-guillotine cutting stock problem. Meanwhile, we use a converting approach which is similar to the Bottom Left (BL) algorithm to map the cutting pattern to the actual layout. Finally, we implement Artificial Fish Swarm Algorithm on several test problems. The simulated results show that the performance of Artificial Fish Swarm Algorithm is better than that of Particle Swarm Optimization Algorithm.

Job scheduling in hadoop is a hot topic, however, current research mainly focuses on the time optimization in scheduling. With the trend of providing hadoop as a service to the public or specified groups, more factors should be considered, such as time and cost. To solve this problem, we present a utility-driven share scheduling algorithm. Considering time and cost, algorithm offers a global optimization scheduling scheme according to the workload of the job. Furthermore, we present a model that can estimate job execute time by cost. Finally, we implement the algorithm and experiment it in a hadoop cluster.

Recent interest in line-tracking methods using UAV has been introduced in the research of pattern recognition and diagnosis of transmission system. A fault diagnosis method for transmission line based on Scale Invariant Feature Transform (SIFT) is proposed in this paper, which recognizes fault images by comparing aerial images with model images. The reliability and efficiency of the system is effectively improved by pro-calculating local scale-invariant features of models. The research can provide a new method for predictive maintenance of the transmission line.

In this paper, a pneumatic actuator fault diagnosis method based on GA optimized BP neural network and fuzzy logic is proposed. First of all, the Genetic algorithm is used to optimize the weights of BP neural network, overcoming the shortcoming of neural network including over learning and local optimum. Then the normal actuator model is trained by the GA optimized BP neural network using the health data of actuator. The residual is generated by comparing the output of the BP trained actuator model and the actual actuator, which is used to detect the fault. Finally, fuzzy logic reasoning is used to isolate the fault type of actuator. The simulation results based on DAMADICS valve model and Lublin Sugar Factory failure data indicate that the proposed method can detect and diagnosis fault of actuator fast and accurately.

We apply the radial basis functions (RBF) decoder adopting Gaussian function for the Turbo product codes (TPC). An extrinsic information extraction scheme based on RBF neural networks (NN) is suggested, and a novel RBF NNs decoding algorithm is proposed. The extrinsic information transfer (EXIT) charts have been used to analyze the convergence property of the TPCs. The EXIT chart analyses show that the proposed decoding algorithm could achieve convergence with about 5 iterations, and improve BER performance in low

E

b

/

N

0

regions. Simulation results show that the proposed algorithm achieves promising BER performance while decreasing decoding computation compared with the maximum a posterior (MAP) algorithm.

This paper proposed a new BP Neural Network (BPNN) based method for load harmonic current assessment where the nonlinearities of electricity loads have been modeled based on differential equations. With the trained BPNN, the load current due to fundamental voltage inputs can be well estimated and used to assess the harmonics components subsequently. The simulation results demonstrate that the proposed method can effectively estimate the total harmonic distortion of the load current when the supplied voltage is within the normal range of harmonic limits. The results also prove that the load harmonic current is nearly independent of load capacity and applied voltage, indicating its effectiveness to distinguish the responsibilities of harmonic pollution between the grid and load.

In the reconstructed phase space, based on the nonlinear time series local prediction method and the relevance vector machine model, the local relevance vector machine prediction method was proposed in this paper, which was applied to predict the small scale traffic measurements data. The experiment results show that the local relevance vector machine prediction method could effectively predict the small scale traffic measurements data, the prediction error mainly concentrated on the vicinity of zero, and the prediction accuracy of the local relevance vector machine regression model was superior to that of the feedforward neural network optimized by PSO.

The data gathering optimization of the large-scale, collaborative and concurrent multi-task in the sensing layer of internet of things is very important, especially in the environments where multiple geographically overlapping wireless sensor networks are deployed. In order to support large-scale, collaborative and concurrent multi-task monitoring, in this paper, we propose a massive sensor sampling data gathering optimization strategy in formed virtual sensor networks to meet various monitoring requirements from different kinds of application deployment and simplify the complexity of dealing with heterogeneous sensor nodes. Then, for the massive sensor sampling data gathering on the virtual sensor networks framework, the CH nodes set and update MinMax hierarchical thresholds to restrict the data transmission. Finally, the simulation results show that proposed strategy achieves more energy savings and increase the sensing layer lifetime of internet of things.

Neural Networks such as RBFN and BPNN have been widely studied in the area of network intrusion detection, with the purpose of detecting a variety of network anomalies (e.g., worms, malware). In real-world applications, however, the performance of these neural networks is dynamic regarding the use of different datasets. One of the reasons is that there are some redundant features for the dataset. To mitigate this issue, in this paper, we propose an approach of combining Neural Networks with Random Forest to improve the accuracy of detecting network intrusions. In particular, we design an intelligent anomaly detection system that uses the algorithm of Random Forest in the process of feature selection and selects an appropriate algorithm in an adaptive way. In the evaluation, we conducted two major experiments using the KDD1999 dataset and a real dataset respectively. The experimental results indicate that Random Forest can enhance the performance of Neural Networks by identifying important and closely related features and that our developed system can select a better algorithm intelligently.

In this paper, a novel computational intelligence scheme is proposed to forecast landslide based on functional networks. Two types functional networks, general functional networks with two variables basis function (GFN) and separable functional networks (SFN) are applied to predict a real-world example. In addition, the experiments reveal that the landslide prediction using functional networks is reasonable and effective, and GFN are consistently better than SFN in terms of the same measurements.

In the paper, three kinds of classifiers are fused they are BP network classifier, self-organizing feature map network classifier and RBF network classifier and the moment invariant features as well as roundness features as the inputs of the fused neural network. Given targets are recognized by the majority voting method and self-adapts weighted fusion algorithm of the fused classifier, and also by the three network classifiers respectively. The recognition results of single neural network and fusion algorithm are analyzed and compared. The results indicate that the recognition rate of multi-neural networks fusion algorithm is higher than any single neural network, and also show that the fusion algorithm has the significance for improving the accuracy of recognition.