Eiji WATANABE Noboru NAKASAKO Yasuo MITANI
This paper proposes a prediction method for non-stationary time series data with time varying parameters. A modular structured type neural network is newly introduced for the purpose of grasping the changing property of time varying parameters. This modular structured neural network is constructed by the hierarchical combination of each neural network (NNT: Neural Network for Prediction of Time Series Data) and a neural network (NNW: Neural Network for Prediction of Weights). Next, we propose a reasonable method for determination of the length of the local stationary section by using the additive learning ability of neural networks. Finally, the validity and effectiveness of the proposed method are confirmed through simulation and actual experiments.
Nobuo FUNABIKI Junji KITAMICHI Seishi NISHIKAWA
A neural network approach called the "Gradual Neural Network (GNN)" for the time slot assignment problem in the TDM multicast switching system is presented in this paper. The goal of this NP-complete problem is to find an assignment of packet transmission requests into a minimum number of time slots. A packet can be transmitted from one source to several destinations simultaneously by its replication. A time slot represents a switching configuration of the system with unit time for each packet transmission through an I/O line. The GNN consists of the binary neural network and the gradual expansion scheme. The binary neural network satisfies the constraints imposed on the system by solving the motion equation, whereas the gradual expansion scheme minimizes the number of required time slots by gradually expanding activated neurons. The performance is evaluated through simulations in practical size systems, where the GNN finds far better solutions than the best existing altorithm.
Toshiyuki TANAKA Hideki KURIYAMA Yoshiko OCHIAI Masao TAKI
Neural networks can be used as associative memories which can learn problems of acquiring input-output relations presented by examples. The learning time problem addresses how long it takes for a neural network to learn a given problem by a learning algorithm. As a solvable model to this problem we analyze the learning dynamics of the linear associative memoty with the least-mean-square algorithm. Our result shows that the learning time τ of the linear associative memory diverges in τ (1-ρ)-2 as the memory rate ρ approaches 1. It also shows that the learning time exhibits the exponential dependence on ρ when ρ is small.
In this letter, we obtain the absolute exponential stability result of neural networks with globally Lipschitz continuous, increasing and bounded activation functions under a sufficient condition which can unify some relevant sufficient ones for absolute stability in the literature. The obtained absolute exponential stability result generalizes the existing ones about absolute stability of neural networks. Moreover, it is demonstrated, by a mathematically rigorous proof, that the network time constant is inversely proportional to the global exponential convergence rate of the network trajectories to the unique equilibrium. A numerical simulation example is also presented to illustrate the analysis results.
Hiroyuki YAMAMOTO Takeshi NAKAYAMA Hiroshi NINOMIYA Hideki ASAI
This paper describes a neuro-based optimization algorithm for three dimensional (3-D) cylindric puzzles which are problems to arrange the irregular-shaped slices so that they perfectly fit into a fixed three dimensional cylindric shape. First, the idea to expand the 2-dimensional tiling technique to 3-dimensional puzzles is described. Next, to energy function with the fitting function of each polyomino is introduced, which is available for 3-D cylindric puzzles. Furthermore our algorithm is applied to several examples using the analog neural array. Finally, it is shown that our algorithm is useful for solving 3-D cylindric puzzles.
Chang Su LEE Chong-Ho CHOI Young CHOI Se Ho CHOI
The defects in the cold rolled strips have textural characteristics, which are nonuniform due to its irregularities and deformities in geometrical appearance. In order to handle the textural characteristics of images with defects, this paper proposes a surface inspection method based on textural feature extraction using the wavelet transform. The wavelet transform is employed to extract local features from textural images with defects both in the frequency and in the spatial domain. To extract features effectively, an adaptive wavelet packet scheme is developed, in which the optimum number of features are produced automatically through subband coding gain. The energies for all subbands of the optimal quadtree of the adaptive wavelet packet algorithm and four entropy features in the level one LL subband, which correspond to the local features in the spatial domain, are extracted. A neural network is used to classify the defects of these features. Experiments with real image data show good training and generalization performances of the proposed method.
This paper compares signal classification performance of multilayer neural networks (MLNNs) and linear filters (LFs). The MLNNs are useful for arbitrary waveform signal classification. On the other hand, LFS are useful for the signals, which are specified with frequency components. In this paper, both methods are compared based on frequency selective performance. The signals to be classified contain several frequency components. Furthermore, effects of the number of the signal samples are investigated. In this case, the frequency information may be lost to some extent. This makes the classification problems difficult. From practical viewpoint, computational complexity is also limited to the same level in both methods.IIR and FIR filters are compared. FIR filters with a direct form can save computations, which is independent of the filter order. IIR filters, on the other hand, cannot provide good signal classification deu to their phase distortion, and require a large amount of computations due to their recursive structure. When the number of the input samples is strictly limited, the signal vectors are widely distributed in the multi-dimensional signal space. In this case, signal classification by the LF method cannot provide a good performance. Because, they are designed to extract the frequency components. On the other hand, the MLNN method can form class regions in the signal vector space with high degree of freedom.
Hsuen-Chyun SHYU Chin-Chi CHANG Yueh-Jyun LEE Ching-Hai LEE
A structure of neural network suitable for clustering and deinterleaving radar pulses is proposed. The proposed structure consists of two networks, one for intrinsic features of pluses and the other for PRIs (pulse repetition intervals). The unsupervised learning method which adjusts the number of nodes for clusters adaptively is adopted for these two networks to learn patterns. These two networks are connected by a set of links. According to the weights of these links, the clusters categorized by the network for features can be refined further by merging or partitioning. The main defect of the unsupervised network with an adaptive number of nodes for clusters is that the result of classification closely depends on the learning sequence of patterns. This defect can be improved by the proposed refinement algorithm. In addition to the proposed structure and learning algorithms, simulation results have also been discussed.
The asynchronous transfer mode (ATM) provides efficient switching capability for various kinds of communication services. To guarantee the minimum quality of services in the ATM networks, the bandwidth allocation setup procedure between the network nodes and users is very important. However, most of call admission control (CAC) methods which have been proposed so far are not fully appropriate to apply to real environments in terms of the complexity of the hardware implementation or the accuracy of assumptions about the cell-arrival processes. In addition, the success of broad bandwidth applications in the future multimedia environments will largely depend on the degree to which the efficiency in communication systems can be achieved, so that establishing high-speed CAC schemes in the ATM networks is an indispensable subject. This paper proposes a new cell-loss rate estimation method for the real time CAC in ATM networks. A neural network model using the Kalman filter algorithm was employed to improve the error minimizing process for the cell-loss estimation problem. In the process of optimizing the three-layer perceptron, the average, the variance, and the 3rd central moment of the number of cell arrivals were calculated, and cell-loss rate date based on the non-parametric method were adopted for outputs of the neural network. Evaluation results concerned with the convergence using the sum of square errors of outputs were also discussed in this paper. Using this algorithm, ATM cell-loss rates can be easily derived from the average and peak of cells rates coming from users. Results for the cell-loss estimation process suggest that the proposed method will be useful for high-speed ATM CAC in multimedia traffic environments.
Kazuhiko TAKAHASHI Minoru SASAKI
A method is presented for implementing a neural control system for controlling a piezopolymer bimorph flexible micro-actuator. Two neural controllers were constructed, both with an adaptive-type neural identifier and a learning-type direct or open-loop neural controller, focusing on the difference in learning speed between the adaptive and learning schemes. Simulated use of the proposed controllers to control a flexible micro-actuator showed that they can do so effectively. Experiments also demonstrated that a neural controller can be used to control a flexible micro-actuator.
In this paper an analog CMOS implementation of Approximate Identity Neural Networks is suggested. In particular a one-input one-output Neural Network with 6 neurons has been designed and fabricated with a 2µm CMOS technology. Due to the small area occupied the circuit proposed for the neuron is suited for the implementation of larger networks.
The purpose of this study is to present results of forecast of ranges for yen to US dollar exchange rate fluctuation in order to evaluate the performance of two algorithms: the original backpropagation (OBP), which is the most widely used algorithm, and the second algorithm (NBP), which is a proposed modification to the first one by the authors. The set of data consisted of economic and financial values that have already been calculated by the Bank of Japan and the Japanese Ministry of Planning and Finance. This data was available though the Nikkei Data Service and stretched from January, 1986, to the end of December, 1992. The results obtained show not only that NBP performs better than OBP since the former speeds up convergence time to a given error value, but also NBP shows a good generalization performance.
Recurrent neural networks have the potential of performing parallel computation for associative memory and optimization, which is realized by the electronic implementation of neural networks in VLSI technology. Since the time delays in real electronic implementation of neural networks are unavoidably encountered and they can cause systems to oscillate, it is thus practically important to investigate the qualitative properties of neural networks with time delays. In this paper, a class of sufficient conditions is obtained, under which neural networks are globally asymptotically stable independent of time delays.
Seiichiro MORO Yoshifumi NISHIO Shinsaku MORI
When N oscillators are coupled by one resistor, we can see N-phase oscillation, because the system tends to minimize the current through the coupling resistor. Moreover, when the hard oscillators are coupled, we can see N, N - 1, , 3, 2-phase oscillation and get much more phase states. In this study, the two types of coupled oscillators networks with third and fifth-power nonlinear characteristics are proposed. One network has two-dimensional hexagonal structure and the other has two-dimensional lattice structure. In the hexagonal circuit, adjacent three oscillators are coupled by one coupling resistor. On the other hand, in the lattice circuit, four oscillators are coupled by one coupling resistor. In this paper we confirm the phenomena seen in the proposed networks by circuit experiments and numerical calculations. In the system with third-power nonlinear characteristics, we can see the phase patterns based on 3-phase oscillation in the hexagonal circuit, and based on anti-phase oscillation in lattice circuit. In the system with fifth-power nonlinear characteristics, we can see the phase patterns based on 3-phase and anti-phase oscillation in both hexagonal and lattice circuits. In particular, in these networks, we can see not only the synchronization based on 3-phase and anti-phase oscillation but the synchronization which is not based on 3-phase and anti-phase oscillation.
Mikio HASEGAWA Tohru IKEGUCHI Takeshi MATOZAKI Kazuyuki AIHARA
We analyze additive effects of nonlinear dynamics for conbinatorial optimization. We apply chaotic time series as noise sequence to neural networks for 10-city and 20-city traveling salesman problems and compare the performance with stochastic processes, such as Gaussian random numbers, uniform random numbers, 1/fα noise and surrogate data sets which preserve several statistics of the original chaotic data. In result, it is shown that not only chaotic noise but also surrogates with similar autocorrelation as chaotic noise exhibit high solving abilities. It is also suggested that since temporal structure of chaotic noise characterized by autocorrelation affects abilities for combinatorial optimization problems, effects of chaotic sequence as additive noise for escaping from undesirable local minima in case of solving combinatorial optimization problems can be replaced by stochastic noise with similar autocorrelation.
A new learning algorithm is proposed to enhance fault tolerance ability of the feedforward neural networks. The algorithm focuses on the links (weights) that may cause errors at the output when they are open faults. The relevances of the synaptic weights to the output error (i.e. the sensitivity of the output error to the weight fault) are estimated in each training cycle of the standard backpropagation using the Taylor expansion of the output around fault-free weights. Then the weight giving the maximum relevance is decreased. The approach taken by the algorithm described in this paper is to prevent the weights from having large relevances. The simulation results indicate that the network trained with the proposed algorithm do have significantly better fault tolerance than the network trained with the standard backpropagation algorithm. The simulation results show that the fault tolerance and the generalization abilities are improved.
Hanzhong GU Haruhisa TAKAHASHI
In this paper, we apply the method of relating learning to hypothesis testing [6] to study average generalization performance of concept learning from noisy random training examples. A striking aspect of the method is that a learning problem with a so-called ill-disposed learning algorithm can equivalently be reduced to a simple one, and for this simple problem, even though a direct and exact calculation of the learning curves might still be impossible, a thorough empirical study can easily be performed. One of the main advantages of using the illdisposed algorithm is that it well models lower quality learning in real situations, and hence the result can provide useful implications as far as reliable generalization is concerned. We provide empirical formulas for the learning curves by simple functions of the noise rate and the sample size from a thorough empirical study, which smoothly incorporates the results from noise-free analysis and are quite accurate and adequate for practical applications when the noise rate is relatively small. The resulting learning curve bounds are directly related to the number of system weights and are not pessimistic in practice, and apply to learning settings not necessarily within the Bayesian framework.
Yoshiaki WATANABE Keiichi YOSHINO Tetsuro KAKESHITA
The Hopfield neural network for optimization problems often falls into local minima. To escape from the local minima, the neuron unit in the neural network is modified to become an oscillatory unit by adding a simple self-feedback circuit. By combining the oscillatory unit with an energy-value extraction circuit, an oscillatory neural network is constructed. The network can repeatedly extract solutions, and can simultaneously evaluate them. In this paper, the network is applied to four NP-complete problems to demonstrate its generality and efficiency. The network can solve each problem and can obtain better solutions than the original Hopfield neural network and simple algorithms.
This paper obtains some new results about the existence, uniqueness, and global asymptotic stability of the equilibrium of a nonlinear continuous neural network, under a sufficient condition weaker than ones presented in the literature. The avobe obtained results can also imply the existing ones about avsolute stability of nonlinear continuous neural networks
Teruyuki MIYAJIMA Takaaki HASEGAWA
In this paper, a multiuser receiver using a Hopfield network (Hopfield network receiver) for asynchronous codedivision multiple-access systems is proposed. We derive a novel likelihood function for the optimum demodulation of a data subsequence whose length is far shorter than that of the entire transmitted data sequence. It is shown that a novel Hopfield network receiver can be derived by exploiting the likelihood function, and the derived receiver leads to a low complexity receiver. The structure of the proposed receiver consists of a bank of correlators and a Hopfield network where the number of units is proportional to both the number of users and the length of a data sequence demodulated at a time. Computer simulation results are presented to compare the performance of the proposed receiver with those of the conventional multiuser detectors. It is shown that the proposed receiver significantly outperforms the correlation receiver, decorrelating detector and multistage detector, and provides suboptimum performnace.