Aiming at the training problem of time-varying input-output process neural networks(PNN),a learning algorithm based on chaos genetic algorithm(CGA) combined with particle swarm optimization(PSO) whose inertial factor is dynamic is proposed in the paper.With the application of the experience memory and sharing information of PSO algorithm,and chaos track traverse searching of CGA,the hybrid evolutionary optimization mechanism of CGA and PSO algorithm is built based on the PNN's training objective function.The adaptive switching of two algorithms is implemented through estimating the fitness and optimization efficiency,and the global optimal solution is obtained in feasible solution space.Experimental results show that the algorithm considerably improves the training efficiency of PNN.

In recent years, dynamic time series analysis with the concept drift has become an important and challenging task for a wide range of applications including stock price forecasting, target sales, etc. In this paper, a recentness biased learning method is proposed for dynamic time series analysis by introducing a drift factor. First of all, the recentness biased learning method is derived by minimizing the forecasting risk based on a priori probabilistic model where the latest sample is weighted most. Secondly, the recentness biased learning method is implemented with an autoregressive process and the multi-layer feed-forward neural networks. The experimental results have been discussed and analyzed in detail for two typical databases. It is concluded that the proposed model has a high accuracy in time series forecasting.

A new linear classifier based on minimum circum circle (CMCC) is proposed in this paper. It first calculates the minimum circum circle of samples for each class. Then the distributions of the samples can be described by these circles. The linear separating hyperplane will intersect the connecting line of the centers of circles. Consequently, the perpendicular to the connecting line of each two centers is defined as the classifier of these two classes. Moreover, some improved classifiers are proposed when the separating hyperplane is not perpendicular to the connecting line or when there are outliers in the samples. The combined classifier based on subclasses is also discussed. In the experiments, the CMCC and its improved algorithms are compared with some other classifiers such as support vector machine, linear discriminant analysis, etc. The experimental results show that the CMCC gives a relatively good performance on both classification accuracy and time cost.

Aiming at the control problem of nonlinear dynamic system,a control signal solving model and algorithm based on Process Neural Network(PNN) was proposed.First,a system forward identification model based on PNN was set up by using nonlinear transform mechanism and self-adaptive learning ability of PNN to time-varying input-output signals of dynamic system,then according to the established model,the system control structure and the expected output signals,a control signal solving model and algorithm which satisfies system dynamic signal transform mechanism and transfer constraint relation was constructed.The information processing mechanism based on PNN control model was analyzed and the control signal optimize method based on GA coupled with LMS was given.The experiment results veri-fy the feasibility of the model and algorithm.

For dynamic information processing problems with process fuzzy information and dynamic domain rules,a fuzzy reasoning process neural network（FRPNN） is proposed in the paper.FRPNN combines fuzzy process reasoning rules with dynamic information processing mechanism of numerical PNN,representing reasoning rules as process neurons,and implements self-adaptive processing to the process quantitative and qualitative mixed information using learning mechanism of PNN.The information processing mechanism of FRPNN is analyzed,and the learning algorithm is given.Taking pumping unit balance diagnosis as example,application results show the effectiveness of the model and the algorithm.

Existing anti-virus methods make use of signatures to detect malicious codes.They are inefficient to detect various forms of computer viruses,especially new variants and unknown viruses.Inspired by biologic immune system,a novel artificial immune based signature extraction method is proposed.This method automatically identifies bit patterns that correlate with viruses using instruction frequency and file frequency,and then identifies higher-level genes that are associated with viruses,generating a detecting virus gene library using the negative selection algorithm which leads to a fairly low false positive rate compared with the traditional signature-based methods.The advantages of our proposed method are described as follows.In the feature extraction phase,the detecting virus gene library stores virus samples with variable number of variable length genes at individual level,and uses multiple genes coexistence in one virus to avoid the possible loss of information considerably,fully taking the advantages of relevance between viral instructions within a virus program;in the classification phase,suspicious programs are analyzed at individual level in contrast to the existing gene matching technique.Experimental results indicate that the proposed method yields high detection rates for obfuscated viruses with an averaged recognition rate of 94% in real-world conditions,the false positive rate can be maintained below 2%.The method has a good generalization ability,and is able to effectively and efficiently detect new variants of known virus and unknown viruses.

Along with the evolution of computer viruses, the number of file samples that need to be analyzed has constantly increased. An automatic and robust tool is needed to classify the file samples quickly and efficiently. Inspired by the human immune system, we developed a local concentration based virus detection method, which connects a certain number of two-element local concentration vectors as a feature vector. In contrast to the existing data mining techniques, the new method does not remember exact file content for virus detection, but uses a non-signature paradigm, such that it can detect some previously unknown viruses and overcome the techniques like obfuscation to bypass signatures. This model first extracts the viral tendency of each fragment and identifies a set of statical structural detectors, and then uses an information-theoretic preprocessing to remove redundancy in the detectors' set to generate 'self' and 'nonself' detector libraries. Finally, 'self' and 'nonself' local concentrations are constructed by using the libraries, to form a vector with an array of two elements of local concentrations for detecting viruses efficiently. Several standard data mining classifiers, including K-nearest neighbor (KNN), radial basis function (RBF) neural networks, and support vector machine (SVM), are leveraged to classify the local concentration vector as the feature of a benign or malicious program and to verify the effectiveness and robustness of this approach. Experimental results show that the proposed approach not only has a much faster speed, but also gives around 98% of accuracy.

The channel assignment problem can heavily impact the performance of multi-radio multi-channel wireless mesh networks. Many channel assignment algorithms are proposed,whereas most of them need the whole network topology or the flow model,which is hard to obtain in the distributed networks. Based on the above analysis,in this paper,we propose the local information based channel assignment (LICA) strategy,which means,by using the heuristic information of local topology and channel usage of all the neighborhood nodes,it allocates the channel resources on each node dynamically. The result shows that algorithm LICA can significantly improve the end-to-end throughput and channel utilization in lower time complexity and also has better expansibility.

In this paper, we investigate how to extract the lowest frequency features from an image. A novel Laplacian smoothing transform (LST) is proposed to transform an image into a sequence, by which low frequency features of an image can be easily extracted for a discriminant learning method for face recognition. Generally, the LST is able to be an efficient dimensionality reduction method for face recognition problems. Extensive experimental results show that the LST method performs better than other pre-processing methods, such as discrete cosine transform (DCT), principal component analysis (PCA) and discrete wavelet transform (DWT), on ORL, Yale and PIE face databases. Under the leave one out strategy, the best performance on the ORL and Yale face databases is 99.75% and 99.4%; however, in this paper, we improve both to 100% with a fast linear feature extraction method for the first time.

This study proposes an adaptive staged particle swarm optimization (ASPSO) algorithm based on analyses of particles’ search capabilities. First, the search processes of the standard PSO (SPSO) and the linear decreasing inertia weight PSO (LDWPSO) are analyzed based on our previous definition of exploitation. Second, three stages of the search process in PSO are defined. Each stage has its own search preference, which is represented by the exploitation capability of swarm. Third, the mapping between inertia weight, learning factor (w-c) and the exploitation capability is given. At last, the ASPSO is proposed. By setting different values of w-c in three stages, one can make swarm search the space with particular strategy in each stage, and the particles can be directed to find the solution more effectively. The experimental results show that the proposed ASPSO has better performance than SPSO and LDWPSO on most of test functions.

As a new kind of vehicles with low fuel cost and low emission, hybrid electric vehicle (HEV) has been given more and more attention in recent years. The key technique in the HEV is the optimal control strategy for the best performance. This paper proposed a new torque control strategy with charge buffer (TCSCB) to control the two power sources of the HEV. The TCSCB is based on the control of engine torque which make the control strategy easily distribute the output power to the engine and motor. In this control strategy, the real time optimization based on the engine efficiency map increases engine efficiency observably. The charge buffer reduces the dramatic fluctuation of the engine torque to improve the fuel economy. The prediction engine torque based on the neural network improves the control performance by the future information greatly. The simulation results showed the TCSCB could reach a higher fuel economy and lower emission compared to the current control strategies. In order to optimize the control performances, the parameters in the TCSCB were also discussed in details.

"Process Neural Network: Theory and Applications" proposes the concept and model of a process neural network for the first time, showing how it expands the mapping relationship between the input and o

Aiming at the training problem of process neural networks,a training algorithm based on numerical integration was proposed.In proposed algorithm,the numerical integration was directly applied to deal with the weighted aggregation of dynamic samples and weight functions in time-domain,and the gradient descent method was used to adjust the weight function characteristic parameters and network property parameters.Three kinds of numerical integration methods of Trapezoidal,Simpson,and Cotes were designed.Taking the prediction of sunspot data as an example,the simulation results show that the training algorithms based on numerical integration are efficient,and the approximation performance of Simpson integration is optimal.

In this paper, we show how support vector machine (SVM) can be employed as a powerful tool for k-nearest neighbor (kNN) classifier. A novel multi-class dimensionality reduction approach, discriminant analysis via support vectors (SVDA), is proposed. First, the SVM is employed to compute an optimal direction to discriminant each two classes. Then, the criteria of class separability is constructed. At last, the projection matrix is computed. The kernel mapping idea is used to derive the non-linear version, kernel discriminant via support vectors (SVKD). In SVDA, only support vectors are involved to compute the transformation matrix. Thus, the computational complexity can be greatly reduced for kernel based feature extraction. Experiments carried out on several standard databases show a clear improvement on LDA-based recognition.

This paper proposes a new learning method for process neural networks (PNNs) based on the Gaussian mixture functions and particle swarm optimization (PSO), called PSO-LM. First, the weight functions of the PNNs are specified as the generalized Gaussian mixture functions (GGMFs). Second, a PSO algorithm is used to optimize the parameters, such as the order of GGMFs, the number of hidden neurons, the coefficients, means and variances of Gaussian functions and the thresholds in PNNs. In PSO-LM, the parameter space is transformed from the function space to real number space by using GGMFs. PSO can give a global search in the real number parameter space by avoiding the premature and gradient calculations in back propagation method. According to our analysis and several experiments. PSO-LM can outperform current basis function expansion based learning method (BFE-LM) for PNNs and the classic back propagation neural networks (BPNNs).

Aiming at the multidimensional dynamic information processing and nonlinear system modeling problem,a theory and method based on process neural networks is proposed in this paper.The inputs/ outputs and connection weights of process neural networks are all multivariate time functions,its aggregation operation includes both weighted space aggregation to multiple input functions and accumulation to multivariate process effect.It can reflect common influence and accumulation result of process effect to multiple multivariate input signals in multidimensional space at the same time.In the paper,some process neural network models are given aiming at typical multidimensional dynamic information processing problems,and its application adaptability in aspects of dynamic pattern recognition,nonlinear system data modeling,system identification and process control,dynamic system simulation are analyzed from information processing mechanism.

A probabilistic process neural network has been proposed in order to provide integration of a priori knowledge with dynamic information classification. In this model,Bayesian classification was combined with the dynamic information processing of process neural networks. Dynamic information classification based on Bayesian rules was realized by adding a pattern neuron layer and a summing neuron layer to a feed forward process neural network and applying the normalized exponential activation function to the hidden layer. Classification equivalence between probabilistic process neural networks and Bayesian rules was analyzed and a concrete learning algorithm presented. Experimental results showed the effectiveness of the proposed model and algorithm.

Both the input and link weights of process neural network can be all time-various functions, an aggregation operator on time is added to the process neuron, which provides the neural network with the capability of handling simultaneously two dimension information of time and space. In consideration of the complexity of the aggregation operation of time in process neural networks, a new learning algorithm based on function orthogonal basis expansion is proposed. Firstly a group of proper function orthogonal bases in the input function space of the neural network is selected, and then the input functions and the network weight functions are represented as expansion of the same orthogonal basis. With orthogonality of basis functions, the aggregation operation of process neurons to time is simplified. The application shows that the algorithms simplify the computing complexity of process neural networks, and raise the efficiency of the network learning and the adaptability to real problem resolving. The effectiveness of the algorithm has been proved in the rotation machinery fault diagnosis and the simulation in oil field development process.