Static and dynamic channel assignment using neural networks

We examine the problem of assigning calls in a cellular mobile network to channels in the frequency domain. Such assignments must be made so that interference between calls is minimized, while demands for channels are satisfied. A new nonlinear integer programming representation of the static channel assignment (SCA) problem is formulated. We then propose two different neural networks for solving this problem. The first is an improved Hopfield (1982) neural network which resolves the issues of infeasibility and poor solution quality which have plagued the reputation of the Hopfield network. The second approach is a new self-organizing neural network which is able to solve the SCA problem and many other practical optimization problems due to its generalizing ability. A variety of test problems are used to compare the performance of the neural techniques against more traditional heuristic approaches. Finally, extensions to the dynamic channel assignment problem are considered     

Longitudinal wheel slip control using dynamic neural networks

The control of automotive braking systems performance and a wheel slip is a challenging problem due to nonlinear dynamics of a braking process and a tire–road interaction. When the wheel slip is not between the optimal limits during braking, the desired tire–road friction force cannot be achieved, which influences braking distance, the loss in steerability and maneuverability of the vehicle. In this paper, the new approach, based on dynamic neural networks, has been employed for improving of the longitudinal wheel slip control. This approach is based on dynamic adaptation of the brake actuation pressure, during a braking cycle, according to the identified maximum adhesion coefficient between the wheel and road. The brake actuated pressure was adjusted on the level which provides the optimal longitudinal wheel slip versus the brake actuated pressure selected by a driver, the current vehicle speed, load conditions, the brake interface temperature and the current value of the wheel slip. The dynamic neural network has been used for modeling of a nonlinear functional relationship between the brake actuation pressure and the longitudinal wheel slip during a braking cycle. It provided preconditions for control of the brake actuation pressure based on the wheel slip change.     

Application of generalized dynamic neural networks to biomedical data

This paper reviews the application of continuous recurrent neural networks with time-varying weights to pattern recognition tasks in medicine. A general learning algorithm based on Pontryagin's maximum principle is recapitulated, and possibilities of improving the generalization capabilities of these networks are given. The effectiveness of the methods is demonstrated by three different real-world examples taken from the fields of anesthesiology, orthopedics, and radiology.     

一种新的基于混沌神经网络的动态路由选择算法

针对通信网的路由选择问题,提出了一种动态路由选择的混沌神经网络实现方法。所提出的此方法具有许多优良特性,即暂态混沌特性和平稳收敛特性,能有效地避免传统Hopfield神经网络极易陷入局部极值的缺陷。它通过短暂的倒分叉过程,能很快进入稳定收敛状态。实验证明了本算法能实时、有效地实现通信网的路由选择,并且当通信网中的业务量发生变化时,算法能自动调整最短路径和负载平衡之间的关系。     

QoS provisioning dynamic connection-admission control for multimedia wireless networks using a Hopfield neural network

This paper presents a quality-of-service (QoS) provisioning dynamic connection-admission control (CAC) algorithm for multimedia wireless networks. A multimedia connection consists of several substreams (i.e., service classes), each of which presets a range of feasible QoS levels (e.g., data rates). The proposed algorithm is mainly devoted to finding the best possible QoS levels for all the connections (i.e., QoS vector) that maximize resource utilization by fairly distributing wireless resources among the connections while maximizing the statistical multiplexing gain (i.e., minimizing the blocking and dropping probabilities). In the case of congestion (overload), the algorithm uniformly degrades the QoS levels of the existing connections (but only slightly) in order to spare some resources for serving new or handoff connections, thereby naturally minimizing the blocking and dropping probabilities (it amounts to maximizing the statistical multiplexing gain). The algorithm employs a Hopfield neural network (HNN) for finding a QoS vector. The problem itself is formulated as a multi-objective optimization problem. Hardware-based HNN exhibits high (computational) speed that permits real time running of the CAC algorithm. Simulation results show that the algorithm can maximize resource utilization and maintain fairness in resource sharing, while maximizing the statistical multiplexing gain in providing acceptable service grades. Furthermore, the results are relatively insensitive to handoff rates.     

Wide-band dynamic modeling of power amplifiers using radial-basis function neural networks

A radial-basis function neural network (RBFNN) has been used for modeling the dynamic nonlinear behavior of an RF power amplifier for third generation. In the model, the signal's envelope is used. The model requires less training than a model using IQ data. Sampled input and output signals were used for identification and validation. Noise-like signals with bandwidths of 4 and 20 MHz were used. The RBFNN is compared to a parallel Hammerstein (PH) model. The two model types have similar performance when no memory is used. For the 4-MHz signal, the RBFNN has better in-band performance, whereas the PH is better out-of-band, when memory is used. For the 20-MHz signal, the models have similar performance in- and out-of-band. Used as a digital-predistortion algorithm, the best RBFNN with memory suppressed the lower (upper) adjacent channel power 7 dB (4 dB) compared to a memoryless nonlinear predistorter and 11 dB (13 dB) compared to the case of no predistortion for the same output power for a 4-MHz-wide signal.     

Identification and control of continuous-time nonlinear systems via dynamic neural networks

In this paper, we present an algorithm for the online identification and adaptive control of a class of continuous-time nonlinear systems via dynamic neural networks. The plant considered is an unknown multi-input/multi-output continuous-time higher order nonlinear system. The control scheme includes two parts: a dynamic neural network is employed to perform system identification and a controller based on the proposed dynamic neural network is developed to track a reference trajectory. Stability analysis for the identification and the tracking errors is performed by means of Lyapunov stability criterion. Finally, we illustrate the effectiveness of these methods by computer simulations of the Duffing chaotic system and one-link rigid robot manipulator. The simulation results demonstrate that the model-based dynamic neural network control scheme is appropriate for control of unknown continuous-time nonlinear systems with output disturbance noise.     

LOGICAL ANALYSIS FOR NEURAL NETWORKS

By comparison with constraint satisfaction networks, this paper presents an essential frame of the logical theory for continuous-state neural networks, and gives the quantitative analyzing method for contradiction. The analysis indicates that the basic reason for the alternation of the logical states of the neurons is the existence of superior contradiction inside the networks. The dynamic process for a neural network to find a solution corresponds to eliminating the superior contradiction.     

Reduced-complexity circuit for neural networks

The Letter demonstrates that a 10 bit reduced-complexity VLSI circuit can be used in place of a 32 bit floating-point processor to speed up some neural network applications, reducing circuit area and power consumption by 88% with a negligible increase in RMS error. Applications were executed on a radial basis function neurocomputer using the reduced-complexity circuit implemented with FPGA technology. One application produced better results than had been previously obtained for a NASA data set using either neural network or non-neural network approaches     

Combining static and dynamic features using neural networks and edge fusion for video object extraction

Semantic object representation is an important step for digital multimedia applications such as object-based coding, content-based access and manipulations. The authors propose an image sequence segmentation scheme which provides region information for the semantic object representation of those applications. The objective is to develop a hardware-friendly segmentation algorithm by combining static and dynamic features simultaneously in one scheme. In the initial stage, a multiple feature space is transformed to one-dimensional label space by using self-organising feature map (SOFM) neural networks. The next stage is an edge fusion process in which edge information is incorporated into the neural network outputs to generate more precisely located boundaries of segmentation. The proposed algorithm differs from existing methods as follows: it can segment textured images with low-dimensional features; leads to more meaningful segmentation region boundaries; and is easier to map into hardware than existing methods. Experimental results are compared with an existing segmentation method using evaluation metrics to clarify the advantages of the proposed algorithm objectively.     

Scalable VLSI implementations for neural networks

This paper discusses research on scalable VLSI implementations of feed-forward and recurrent neural networks. These two families of networks are useful in a wide variety of important applications—classification tasks for feed-forward nets and optimization problems for recurrent nets—but their differences affect the way they should be built. We find that analog computation with digitally programmable weights works best for feed-forward networks, while stochastic processing takes advantage of the integrative nature of recurrent networks. We have shown early prototypes of these networks which compute at rates of 1–2 billion connections per second. These general-purpose neural building blocks can be coupled with an overall data transmission framework that is electronically reconfigured in a local manner to produce arbitrarily large, fault-tolerant networks.     

Double-exponential sigmoidal functions for neural networks

A computationally efficient sigmoidal activation function is presented, called a double-exponential signal function, and the properties are compared with other signal functions. The sigmoidal function is monotonously increasing, continuous in all derivaties, and its output is 0.5 for zero input. The weight multiplication can be replaced by an addition when the training of the network is performed offline. We also present an approximation of this signal function, called a polygonal signal function, reducing the computational effort solely to bit sets and shift operations.     

Digital systems for artificial neural networks

The current art of digital electronic implementation of neural networks is reviewed. Most of this work has taken place as digital simulations on general-purpose serial or parallel digital computers. Specialized neural network emulation systems have also been developed for more efficient learning and use. Dedicated digital VLSI integrated circuits offer the highest near-term future potential for this technology     

Improved neural networks for analog-to-digital conversion

In this paper we study the problem of designing a neural network that gives the correct binary representation of a given real number. Previously this problem has been studied by Tank and Hopfield. The network proposed by them exhibits hysteresis in the sense that the current vector of the network sometimes converges towards a binary vector that isnot the correct binary representation of the input current. The reason for this is that the network proposed by them has multiple asymptotically stable equilibria. In the present paper, we propose another neural network which has the property that it hasa single, globally attractive equilibrium for almost all values of the input current. Hence, irrespective of the initial conditions of the network, the current vector converges towards the correct binary representation of the input current.     

Four quadrant multiplier for neural networks

A four quadrant multiplier suitable for neural networks is presented. The product of two analogue voltages Vx and Vw is achieved by performing the difference of two currents, I/sub 1/ and I/sub 2/, that are nonlinearly related to Vx and Vw. Both currents are passed through the same transistor in two different time slices, eliminating, thereby, problems due to component mismatch. A test chip has been designed and fabricated in a 3 mu m CMOS technology.<>     

Projection learning for self-organizing neural networks

A new learning scheme, called projection learning (PL), for self-organizing neural networks is presented. By iteratively subtracting out the projection of the “twinning” neuron onto the null space of the input vector, the neuron is made more similar to the input. By subtracting the projection onto the null space as opposed to making the weight vector directly aligned to the input, we attempt to reduce the bias of the weight vectors. This reduced bias will improve the generalizing abilities of the network. Such a feature is important in problems where the in-class variance is very high, such as, traffic sign recognition problems. Comparisons of PL with standard Kohonen learning indicate that projection learning is faster. Projection learning is implemented on a new self-organizing neural network model called the reconfigurable neural network (RNN). The RNN is designed to incorporate new patterns online without retraining the network. The RNN is used to recognize traffic signs for a mobile robot navigation system     

Optical neural networks

Classical optical information processing and classical neural networks can be adapted and combined to create optical neural networks which offer significant and fundamental advantages over electronic neural networks in various well-defined cases. A systematic morphology of optical neural networks is presented. Special problems they create are discussed. The state of the art of their implementation is indicated, and some supportable speculations on their future are given     

STUDIES OF THE DYNAMIC BEHAVIORS OF A CLASS OF LEARNING ASSOCIATIVE NEURAL NETWORKS

This paper investigates exponential stability and trajectory bounds of motions of equilibria of a class of associative neural networks under structural variations as learning a new pattern. Some conditions for the possible maximum estimate of the domain of structural exponential stability are determined. The filtering ability of the associative neural networks contaminated by input noises is analyzed. Employing the obtained results as valuable guidelines, a systematic synthesis procedure for constructing a dynamical associative neural network that stores a given set of vectors as the stable equilibrium points as well as learns new patterns can be developed. Some new concepts defined here are expected to be the instruction for further studies of learning associative neural networks.     

Artificial neural networks

Examines the following questions associated with artificial neural networks: why people are interested in artificial neural networks; what artificial neural networks are, from the point of view of electronic circuits, and how they work; how they can be programmed and made to solve particular problems; and whether interesting problems can actually be put on such networks. The author then describes the current state of artificial neural network technology and the resulting challenges to people working on electronic devices