Search

Recently, graph embedding-based methods have drawn increasing attention for dimensionality reduction (DR) of hyperspectral image (HSI) classification. Graph construction is a critical step for thos...

The domain adaptation uses labeled source domain data to train a classifier to be used in the target domain with no or small amount of labeled data. Usually there exists discrepancy in terms of marginal and conditional distributions for both source and target domains, which is of critical importance to minimize the distribution discrepancy between domains. As a classical model in deep learning, the autoencoder is capable of realizing distribution matching and enhancing classification accuracy by extracting more abstract and effiective features from data. A Domain adaptation network based on autoencoder (DANA) is proposed. The DANA structure consists of a couple of encoding layers: a feature extraction layer and a classification layer. For the feature extraction layer, the marginal distributions of source and target domains are matched by using the nonparametric maximum mean discrepancy measurement. For the classification layer, the softmax regression model is applied to encode the label information of source domains meanwhile to match the conditional distribution. Experimental results on ImageNet, Corel and Leaves datasets have shown the enhanced classification accuracy by our proposed algorithm compared with the classical methods.

A reliable and precise recognition of the differentially expressed genes of tumor is crucial to treat the cancer effectively. The small number of differentially expressed genes in a huge gene expression dataset determines the important role of sparse methods, such as Penalty matrix decomposition (PMD), among the feature selection methods. The sparse methods always have the drawback: they do not take advantage of known class labels of gene expression data. A novel supervised-sparse method named as Supervised PMD (SPMD) is proposed by adding the class information into PMD via the total scatter matrix. The brief idea of our method used to select the differentially expressed genes is given as follows. The total scatter matrix is obtained according to the gene expression data with class label. The obtained total scatter matrix is decomposed by PMD to acquire the sparse vectors. The non-zero items in sparse vectors are selected as the differentially expressed genes. The Gene ontology (GO) enrichment of functional annotation of the selected genes is detected by ToppFun. Experiments on synthetic data and two real tumor gene expression datasets show that the proposed SPMD is quite promising to select the differentially expressed genes.

Identification of characteristic genes associated with specific biological processes of different cancers could provide insights into the underlying cancer genetics and cancer prognostic assessment. It is of critical importance to select such characteristic genes effectively. In this paper, a novel unsupervised characteristic gene selection method based on sample learning and sparse filtering, Sample Learning based on Deep Sparse Filtering (SLDSF), is proposed. With sample learning, the proposed SLDSF can better represent the gene expression level by the transformed sample space. Most unsupervised characteristic gene selection methods did not consider deep structures, while a multilayer structure may learn more meaningful representations than a single layer, therefore deep sparse filtering is investigated here to implement sample learning in the proposed SLDSF. Experimental studies on several microarray and RNA-Seq datasets demonstrate that the proposed SLDSF is more effective than several representative characteristic gene selection methods (e.g., RGNMF, GNMF, RPCA and PMD) for selecting cancer characteristic genes.

Deep convolutional network demonstrates that the classification accuracy can be remarkably improved by increasing the number of network layers, however, increases the accuracy by 1% of costs nearly doubling the number of layers. Meanwhile, gradient dispersion will occur in the training process, which leads to performance degradation. In order to solve the problem of training difficulty with the increased number of layers, we focus on network architecture and propose a deep convolutional network based on the pyramid structure. In the network architecture, as the number of layers increased, the feature map dimensions (i.e., the number of channels) are gradually increased at each layer to distribute the burden concentrated at locations of structural units affected by downsampling, such that all units are equally distributed. By exploring the sequence between the stacked elements inside the structural unit, we design a pyramidal building block, as its shape gradually widens from the top downwards, which is called the deep pyramid convolutional network (DPCNet). Experimental results on CIFAR-10 and CIFAR-100 datasets have shown that DPCNet has the superior generalization capability and can effectively improve the image classification accuracy.

Since tumor is seriously harmful to human health, effective diagnosis measures are in urgent need for tumor therapy. Early detection of tumor is particularly important for better treatment of patients. A notable issue is how to effectively discriminate tumor samples from normal ones. Many classification methods, such as Support Vector Machines (SVMs), have been proposed for tumor classification. Recently, deep learning has achieved satisfactory performance in the classification task of many areas. However, the application of deep learning is rare in tumor classification due to insufficient training samples of gene expression data. In this paper, a Sample Expansion method is proposed to address the problem. Inspired by the idea of Denoising Autoencoder (DAE), a large number of samples are obtained by randomly cleaning partially corrupted input many times. The expanded samples can not only maintain the merits of corrupted data in DAE but also deal with the problem of insufficient training samples of gene expression data to a certain extent. Since Stacked Autoencoder (SAE) and Convolutional Neural Network (CNN) models show excellent performance in classification task, the applicability of SAE and 1-dimensional CNN (1DCNN) on gene expression data is analyzed. Finally, two deep learning models, Sample Expansion-Based SAE (SESAE) and Sample Expansion-Based 1DCNN (SE1DCNN), are designed to carry out tumor gene expression data classification by using the expanded samples. Experimental studies indicate that SESAE and SE1DCNN are very effective in tumor classification.

To achieve high classification accuracy of hyperspectral data, a dimensionality reduction algorithm called Sample-dependent repulsion graph regularized auto-encoder (SRGAE) is proposed. Based on the sample-dependent graph, by applying the repulsion force to the samples from different classes but nearby, a sampledependent repulsion graph is built to make the samples from the same class will be projected to samples that are close-by and the samples from different classes will be projected to samples that are far away. The sampledependent repulsion graph can avoid the neighborhood parameter selection problem existing in the nearest neighborhood graph. By integrating advantages of deep learning and graph regularization technique, the SRGAE can maintain the learned deep features are consistent with the inherent manifold structure of the original hyperspectral data. Experimental results on two real hyperspectral data show that, when compared with some popular dimensionality reduction algorithms, the proposed SRGAE can yield higher classification accuracy.

With the semi-conductor technology gradually approaching its physical and heat limits, recent supercomputers have adopted major architectural changes to continue increasing the performance through more power-efficient heterogeneous many-core systems. Examples include Sunway TaihuLight that has four Management Processing Element (MPE) and 256 Computing Processing Element (CPE) inside one processor and Summit that has two central processing units (CPUs) and 6 graphics processing units (GPUs) inside one node. Meanwhile, current high-resolution Earth system models that desperately require more computing power, generally consist of millions of lines of legacy codes developed for traditional homogeneous multi-core processors and cannot automatically benefit from the advancement of supercomputer hardware. As a result, refactoring and optimizing the legacy models for new architectures become a key challenge along the road of taking advantage of greener and faster supercomputers, providing better support for the global climate research community and contributing to the long-lasting society task of addressing long-term climate change. This article reports the efforts of a large group in the International Laboratory for High-Resolution Earth System Prediction (iHESP) established by the cooperation of Qingdao Pilot National Laboratory for Marine Science and Technology (QNLM), Texas A & M University and the National Center for Atmospheric Research (NCAR), with the goal of enabling highly efficient simulations of the high-resolution (25-km atmosphere and 10-km ocean) Community Earth System Model (CESM-HR) on Sunway TaihuLight. The refactoring and optimizing efforts have improved the simulation speed of CESM-HR from 1 SYPD (simulation years per day) to 3.4 SYPD (with output disabled), and supported several hundred years of pre-industrial control simulations. With further strategies on deeper refactoring and optimizing for a few remaining computing hot spots, we expect an equivalent or even better efficiency than homogeneous CPU platforms. The refactoring and optimizing processes detailed in this paper on the Sunway system should have implications to similar efforts on other heterogeneous many-core systems such as GPU-based high-performance computing (HPC) systems.

Canned electrical machines as hydraulic pump drives are widely used. Such a machine has quite a difference in electromagnetic properties due to the can shield in air gap, causing essential can effect, which is investigated in this paper. A canned 12–10 permanent magnet (PM) electrical machine with tooth concentrated coils is studied. An analytical model with a concentric-layer structure is developed. The air-gap flux variation due to the can effect and then ohmic losses from cans (can loss) due to PMs and armature windings at typical operating condition are studied, respectively. Loss and output characteristics with regard to can thickness and material are discussed.

The can shields in airgap characterize a canned electrical machine as a hydraulic pump drive. This paper studies modeling and characterization of the can effect. To analyze electromagnetic field on can shields, followed by can loss, an analytical model is proposed. Such a model is of concentric cylindrical rotor layer structure with the slotting effect and magnetic saturation. This work is performed by comparing an ordinary induction machine with a canned version. Based on differences of airgap flux distribution, iron loss due to the can effect is analyzed. Then, the can loss and harmonics under typical operating condition are analyzed. The can loss calculation determined by magnetic saturation is discussed. Finite element method and measurement are used as reference.

This paper presents thermal characteristics of a novel canned Switched Reluctance Machine (SRM) as a hydraulic pump drive. Due to considerable ohmic loss from the can shield structure, thermal analysis is essential. A novel lumped parameter network model featured by using compensation elements is proposed. As a result, calculation accuracy is improved by removing traditional systematic error. The modeling process is described in detail, including thermal resistances and compensation elements. Accuracy of the model is validated by both finite element method (FE) and measurement.

In our study, the active learning and semisupervised learning methods are comprehensively used for label delivery of proteins with known functions in Proteinprotein interaction (PPI) network so as to predict the functions of unknown proteins. Because the real PPI network is generally observed with overlapping protein nodes with multiple functions, the mislabeling of overlapping protein may result in accumulation of prediction errors. For this reason, prior to executing the label delivery process of semi-supervised learning, the adjacency matrix is used to detect overlapping proteins. As the topological structure description of interactive relation between proteins, PPI network is observed with party hub protein nodes that play an important role, in co-expression with its neighborhood. Therefore, to reduce the manual labeling cost, party hub proteins most beneficial for improvement of prediction accuracy are selected for class labeling and the labeled party hub proteins are added into the labeled sample set for semisupervised learning later. As the experimental results of real yeast PPI network show, the proposed algorithm can achieve high prediction accuracy with few labeled samples.

In this paper, a novel magnetic equivalent circuit (MEC) is proposed to calculate magnetic flux of saturated electrical machines. The machine features double salient poles with high saturation. Emphases are made on modeling of air-gap permeance and magnetic saturation using a fitting method. The analyses are applied based on a 4-phase 8/6 switched reluctance machine (SRM) with high power density, which was designed for traction application. The air-gap reluctance in MEC that connects a tooth pair is studied, followed by distribution of magnetic saturation on teeth. The magnetic reluctance of circuit loop can be fitted as well as saturation quantity if the electrical steel is known. The proposed network offers a fast and accurate way of estimating flux characteristics of especially highly saturated machines with double slotting effect.

Concentrated coils have been gaining interest for lower copper loss and production cost [1–3]. As to SRMs, all turns are featured by being fixed onto a single tooth. To have a comprehensive analysis, including the use of cans, an analytical model is necessary, in which the magneto-motive force (MMF) distribution is one primary consideration [4, 5]. In this chapter an analytical MMF model is proposed.

Determination of flux linkage characteristics is a central step in electromagnetic modeling. The flux linkage, a group of curves varying with phase current levels and rotor positions, is required very accurate, on which main outputs such as torque, radial force and iron loss highly depend [1, 2]. However, modeling of flux linkage, a seemingly simple question, is still under study, especially for electrical machines with salient poles. This is because of the saliency that leads to rotor position dependent airgap paths together with magnetic saturation. Further in this study, flux linkage curves are complicated by adding can-shields in airgap. Due to electrical resistance of cans, the alternating airgap flux will induce eddy current circulation, which will in turn affect that airgap flux, making a coupled relation.

In this chapter, electromagnetic analysis of a canned SRM is studied. Finite element (FE) method is applied, with validation via measurement. Chapter organization is as follows: In Sect. 3.1, canned SRM and operation principle are briefly introduced, with the calculation method described. In Sect. 3.2, eddy current distribution at defined typical rotor positions by a single phase excitation is illustrated, followed by applying all phases. In Sect. 3.3, the can loss analysis is extended to one stroke period and further different control modes. In Sect. 3.4, the airgap flux and eddy current loss features due to the use of cans are shown. In Sect. 3.5, experimental validation is taken.

The can loss is the main constraint to high efficient and high power density operation of canned machines and most analyses are FE based or with resort to empirical equations. Detailed study of the can effect is necessary, which leads to detailed electromagnetic analysis. Electromagnetic field analysis plays the central role in predicting output characteristics. Besides the numerical methods [1, 2], analyses fall into semi-analytical [3, 4] or analytical methods. As to the analytical, a couple of alternatives are predominantly developed. One is the Maxwell’s theorem [5–11]. Specifically in [5, 6], the machine is divided into sub-domains (iron cores, airgap, slots and magnets). Our previous work has been applied on an induction, permanent magnet or switched reluctance machine [8, 12, 13]. Magnetic vector potential of each domain is calculated based on magneto-motive force (MMF) distribution as prerequisite. For simplicity a smooth airgap of constant radial length is assumed. By applying Carter’s factor [14] or airgap permeance function [8], the airgap is radially enlarged to account for slotting effect. Another alternative focuses on airgap flux density by studying respectively airgap permeance function and MMF distribution [9]. However when cans are concerned, both of these methods may cause numerical deviation.

A canned electrical machine is the core drive component in a hydraulic pump system, which has applications in nuclear power, coal or deep sea mining, as well as hydraulic energy conversion, etc. The pump body is connected with an electrical machine to form a drive system. The traditional structure is shown in Fig. 1.1a: The blade is driven by an ordinary machine and torque is delivered via a connecting mechanism that however leads to low efficiency, less reliability and complicated maintenance. Alternatively the improvement is shown in Fig. 1.1b that the machine is canned by adding a can-shield structure in airgap. The structure includes a stator can that is fixed onto the inner bore of the stator part and a rotor can onto the outer bore of the rotor part. The liquid being pumped is able to get into the airgap between cans. The liquid stays in the chamber walled by both cans and the axial end cap, and however cannot enter into rotor and stator slots. Configuration of the can shields is shown in Figs. 1.1c and d. Cans are resistant of high temperature, high pressure, erosion, and are zero leakage, which ensures durable operation for special liquid delivery under harsh environment.

The application of can shields makes electrical machines thoroughly different in electromagnetic performance. Thermal analysis is particularly necessary [1–4], as the can loss value is considerably higher than copper or iron loss. Identification of temperature rise on the cans plays the central role, including the overall rise and local thermal sensitive regions. A fast and accurate mathematic heat transfer model is essential and one classic is the lumped parameter network [5–7]. Derived from the machine geometry, heat sources, materials and cooling, this model takes all components and heat transfer mechanisms [8–10]. However, the heat flow is considered constant within the volume being estimated, whereas in fact it alters [11, 12], causing systematic mistake.

Tumor clustering is a powerful approach for cancer class discovery which is crucial to the effective treatment of cancer. Many traditional clustering methods such as NMF-based models, have been widely used to identify tumors. However, they cannot achieve satisfactory results. Recently, subspace clustering approaches have been proposed to improve the performance by dividing the original space into multiple low-dimensional subspaces. Among them, low rank representation is becoming a popular approach to attain subspace clustering. In this paper, we propose a novel Low Rank Subspace Clustering model via Discrete Constraint and Hypergraph Regularization (DHLRS). The proposed method learns the cluster indicators directly by using discrete constraint, which makes the clustering task simple. For each subspace, we adopt Schatten $p$ -norm to better approximate the low rank constraint. Moreover, Hypergraph Regularization is adopted to infer the complex relationship between genes and intrinsic geometrical structure of gene expression data in each subspace. Finally, the molecular pattern of tumor gene expression data sets is discovered according to the optimized cluster indicators. Experiments on both synthetic data and real tumor gene expression data sets prove the effectiveness of proposed DHLRS.

An identification algorithm of overlapping protein complexes is put forward by simultaneously considering the topological structural and biological functional information of Protein-protein interaction (PPI) network. Main works include: constructing the edge weight of weighted PPI network on the basis of structural and functional information of PPI network to more accurately describe the correlation between protein vertices; improving the Newman algorithm to make it applicable to weighted PPI network and thus to identify overlapping protein complexes; and providing the denoising criteria based on the structural and function information of PPI network: connections which have no contribution to the high aggregation of PPI network or which are among proteins of independent functions are judged to be false positive connections. The experimental results on the dataset of saccharomyces cerevisiae PPI network show that the proposed algorithm has higher identification accuracy and matching rate when compared with the current representative identification algorithms of protein complexes.

A quick convex hull building algorithm using grid and binary tree is proposed for the minimum convex buidling of planar point set. Grids are used to assess and eliminate those interior points wihtout any contribution to convex hull building and points are sought in the boundary grid only so as to enhance the efficiency of algorithm. The minimum convex bull is built by taking such advantages of binary tree as quick, convenient and applicable for various point sets with different distributions, so as to resolve the description problem of concave point. The time complexity of the algorithm is low because of grid pretreatment. As the results of comparative expriment of random point and actual picture show, the proposed algorithm can obtain the best profile of 2D planar picture with minimum time, which is applicable for describing the shape of irregular convex-concave objects.

The concept for generalα,m,r-convex functions, as a generalization of convex functions, is introduced. Then, Sandor type inequalities for the Sugeno integral based on this kind of function are established. Our work generalizes the previous results in the literature. Finally, some conclusions and problems for further investigations are given.

The learning ability and generalizing performance of the support vector machine (SVM) mainly relies on the reasonable selection of super-parameters. When the scale of the training sample set is large and the parameter space is huge, the existing popular super-parameter selection methods are impractical due to high computational complexity. In this paper, a novel super-parameter selection method for SVM with a Gaussian kernel is proposed, which can be divided into the following two stages. The first one is choosing the kernel parameter to ensure a sufficiently large number of potential support vectors retained in the training sample set. The second one is screening out outliers from the training sample set by assigning a special value to the penalty factor, and training out the optimal penalty factor from the remained training sample set without outliers. The whole process of super-parameter selection only needs two train-validate cycles. Therefore, the computational complexity of our method is low. The comparative experimental results concerning 8 benchmark datasets show that our method possesses high classification accuracy and desirable training time.

The effect of the copper content on properties of electroless Ni-Cu-P coating on heat exchanger surface was investigated, such as adhesion strength and surface characteristic, and anti-fouling property, which were considered to mitigate the accumulation of mineral fouling in the heat exchangers. The electroless ternary Ni-Cu-P coatings with different copper content were prepared on mild steel (1015) substrate surfaces by adjusting process parameters. Surface morphologies of coating and adhesion strength were investigated by using scanning electron microscopy (SEM) and MFT-4000 multifunctional material surface performance instrument, respectively. The results showed that the adhesion strength was improved with the addition of copper in the coating. With the increase of copper content, the deposition rate of ternary Ni-Cu-P coatings was increased, and the boundary of nodular became obvious. Moreover, the surface free energy of ternary Ni-Cu-P coatings was increased with the increase of copper content in the coatings and then decreased when enhancing the copper content further. The further fouling experiments indicated that all the ternary Ni-Cu-P coating surfaces with different copper content inhibited the adhesion of fouling compared with the stainless steel surface. The adhesion weight of fouling was approximately in proportion with the copper addition of ternary Ni-Cu-P coatings, but not the value of surface free energy. The work provides evidence that both adhesion strength and anti-fouling ability should be combined to use when applying surface modification in the field of heat exchanger.

The purpose of the present study was to investigate the effect of the electroless Ni–Cu–P-PTFE deposit surface on anti-fouling of heat exchangers, which was considered as a way to mitigate the accumulation of mineral fouling in the heat exchangers. Electroless Ni–Cu–P-PTFE deposits with various PTFE content were prepared on mild steel (1015) substrate surface by different process parameters. Surface morphology and microhardness were investigated by using SEM, MH-6 Vickers, respectively. The results showed that the addition of PTFE particles into the Ni–Cu–P matrix hardly affected the microstructure of the deposits. Microhardness was decreased with the addition of PTFE in the deposits. Moreover, the surface free energy of Ni–Cu–P-PTFE deposits was decreased with the increase of PTFE particles in the deposits. Further fouling experiments indicated that the surfaces of Ni–Cu–P-PTFE deposits with different PTFE content inhibited the adhesion of fouling compared with the mild steel surface of the heat exchangers. The adhesion weight of fouling was approximately in inverse proportion with the addition of PTFE particles in the deposits, but not the value of surface roughnesss. The anti-fouling property can not be improved ideally even considering the option of making Ni–Cu–P-PTFE coatings smooth.

The use of metallic cans in airgap forms an electromagnetic shielding phenomenon that characterizes a canned electrical motor. Imposed in alternating airgap flux field, cans induct strong and distributive eddy current, the fundamental feature that interacts original flux and determines output features. In particular that induction leads to additional can-shielding loss that is considerably much higher than traditional copper or iron losses. In this article, the can effect in terms of eddy current and loss is studied, based on a state-of-the-art permanent magnet motor. First, distribution characteristics of eddy induction and loss are analyzed via a combination method that each magnetic excitation is separately and then integrally taken. Then, the can loss by magnetic excitation, rotor speed, flux harmonics, as well as the change of traditional losses due to the use of cans, is analyzed. Measurement is taken as verification.

Coal mines require various kinds of machinery. The fault diagnosis of this equipment has a great impact on mine production. The problem of incorrect classification of noisy data by traditional support vector machines is addressed by a proposed Probability Least Squares Support Vector Classification Machine (PLSSVCM). Samples that cannot be definitely determined as belonging to one class will be assigned to a class by the PLSSVCM based on a probability value. This gives the classification results both a qualitative explanation and a quantitative evaluation. Simulation results of a fault diagnosis show that the correct rate of the PLSSVCM is 100%. Even though samples are noisy, the PLSSVCM still can effectively realize multi-class fault diagnosis of a roller bearing. The generalization property of the PLSSVCM is better than that of a neural network and a LSSVCM.

This paper presents a dynamic model and a design method for an accurate self-tuning pressure regulator for pneumatic-pressure–load systems that have some special characteristics such as being nonlinear and time-varying. A mathematical model is derived, which consists of a chamber continuity equation, an orifice flow equation and a force balance equation of the spool. Based on a theoretical analysis of the system dynamics, a three-order controlled auto-regressive moving average (CARMA) model is used to describe the practical pressure–load systems. Then a linear quadratic Gaussian self-tuning pressure regulator is designed to realize an adaptive control of pressure in the chamber. Because the system parameters are time-varying and the system states are difficult to detect, the recursive forgetting factor least-squares algorithm and the Kalman filtering method are adopted to estimate the system parameters and the system states. Experimental results show that the proposed self-tuning pressure regulator can be adapted to parameters which vary with such factors as the volume of the chamber and the setting pressure and that better dynamic and static performances can be obtained.

This paper presents a new multi-step predictive controller based on neural networks and researches the adaptability of the predictive controller for a pneumatic position servo system which has some typical characteristics of non-linearity and time-varying. A diagonal recurrent neural network (DRNN) is used to predict the system output of the multi-step ahead directly. According to the intrinsic defects of a back-propagation (BP) algorithm that cannot update network weights incrementally, a new hybrid learning algorithm combining the temporal differences (TD) method with the BP algorithm to train the DRNN is put forward. A three-layer feedforward BP neural network is used as a non-linear rolling optimal controller to realize the optimization of control input of the next step according to a single-value predictive control algorithm to simplify computation. Simulation and experimental results indicate that the proposed predictive controller is suitable for real-time control of a pneumatic position servo system because of its characteristics of a simple algorithm, fast calculation of the control input and good tracking effects.

In this paper, a Barnes-Godunova-Levin type inequality for the Sugeno integral based on an $( {\alpha,m} )$ -concave function is proved. Some examples are given to illustrate the validity of these inequalities. Finally, several important results, as special cases of an $( {\alpha,m} )$ -concave function, are also obtained.

A new Pareto-based differential evolution (PDE) algorithm for solving multi-objective optimization problems was proposed by applying the nondominated sorting and ranking selection procedure developed in NSGA-II to select nondominated individuals to constitute a nondominated solution set. The PDE algorithm was validated using eight benchmark cases. The experimental results show that PDE, compared with NSGA-II algorithm, can find many Pareto optimal solutions distributed onto the Pareto front uniformly, which is an effective method to solve multi-objective optimization problems.

In order to make full use of the advantages of both parametric and non-parametric models simultaneously, a kind of semi-parametric support vector machine (SVM) was proposed by combining a non-parametric SVM model and a parametric linear basis function model. The semi-parametric SVM was used to estimate the Q values of continuous-state-discontinuous-action pairs in an on-line manner so as to generalize a standard Q learning method to continuous state spaces. Simulation results concerning the balancing control problem of an inverted pendulum show that the proposed Q learning method has good adaptability for changes of system parameters and initial states, which provides a new approach to solve the generalization problem of continuous space of reinforcement learning.

One of the difficulties encountered in the application of reinforcement learning methods to real-world problem is the generalization of large-scale or continuous state space. In order to solve the curse of dimensionality problem caused by discretizing continuous state space, a kind of Q-learning method for continuous state space based on a dynamic Elman neural network was proposed in this paper. The inputs and the output of Elman network are the system state-action pair and the corresponding Q-value. That is, Elman network is used to estimate the Q-value of state-action pair on-line. Eligibility trace for connecting weights is introduced by borrowing ideas from the eligibility trace mechanism of state in temporal difference algorithm to improve the learning speed of neural network. Computer simulations on mountain car control illustrate the performance and applicability of the proposed Q-learning scheme.