This article investigates the finite-horizon $H_{\infty } $ tracking control problem for discrete-time (DT) linear systems with partial observation and unknown dynamics from a game-theoretic perspective. Unlike existing reinforcement learning (RL) approaches that primarily address infinite-horizon, time-invariant systems with full state information, our setting requires solving time-varying Riccati equations and developing model-free methods that rely solely on input-output data. To tackle these challenges, we reconstruct the system state from historical input-output trajectories, driving to a data-driven system representation, and we define an input-output-based time-varying $Q$ -function. We then propose two minimax $Q$ -learning algorithms that do not require an initially admissible policy and avoid the use of a discount factor, thereby removing a long-standing obstacle to stability guarantees. Moreover, the framework readily extends to both infinite-horizon and time-varying systems without structural modifications. Convergence is proved theoretically, and the effectiveness of the algorithms is validated through simulations.

This article studies the stability issue of networked switched systems (NSSs) under denial-of-service (DoS) attacks. To address this issue, the derived limitations imposed on both the frequency of DoS attacks on each subsystem and the upper limit of attack duration that each subsystem can tolerate are mode-dependent, which is more efficient and flexible than the current results for NSSs. Moreover, we reveal the relationship between the upper bound of the average maximum tolerable attack duration associated with the corresponding subsystem and the actual mode-dependent average dwell time. Furthermore, we identify that the total tolerable DoS attack duration as a percentage of the system runtime in this article can be higher than existing results. Finally, an example is given to demonstrate the effectiveness of our work.

This article focuses on solving a finite-horizon nonlinear optimal control problem by using the Pontryagin's maximum principle. In practical applications, linearization is a common approach for solving nonlinear dynamical systems. However, it is not universally applicable due to various reasons, such as instability and low accuracy. In contrast to linearization, the inherent challenge in directly solving the above nonlinear optimal control problem lies in addressing the highly coupled nonlinear forward and backward differential equations. In order to address this problem, an equivalent relationship is established between these equations and a new optimization problem. By exploiting the inherent relationship between supervised learning and an optimization problem from the view of a dynamical system, a deep neural network framework is constructed for describing the new optimization problem. Furthermore, a numerical algorithm for optimal control, which is very powerful for a large variety of nonlinear dynamical systems, is implemented by training a deep residual network. Finally, the effectiveness of the algorithm is demonstrated by solving a trajectory tracking control problem for automatic guided vehicle. The obtained results reveal that the proposed control scheme can achieve high-precision tracking.

A new asymptotic Nash equilibrium for a class of linear quadratic mean-field game problem with a finite number of agents is found by using the cost function decomposition method. The cost function value corresponding with this equilibrium is also computed. This value turns out to be smaller than the one obtained by the state average approximation method. The difference is prominent when the number of agents is small, but vanishes as the agents' number tends to infinity.

In this article, we study the feedback Nash strategy of the model-free nonzero-sum difference game. The main contribution is to present the Q -learning algorithm for the linear quadratic game without prior knowledge of the system model. It is noted that the studied game is in finite horizon which is novel to the learning algorithms in the literature which are mostly for the infinite-horizon Nash strategy. The key is to characterize the Q -factors in terms of the arbitrary control input and state information. A numerical example is given to verify the effectiveness of the proposed algorithm.

This paper proposes fast convergent distributed algorithms for weighted average consensus of input data. For acyclic graphs, we give an algorithm that converges to the exact weighted average consensus in a finite number of iterations, equal to the graph diameter. For loopy (cyclic) graphs, we offer two remedies. In the first one, we give another distributed algorithm to enable our average consensus algorithm applicable to a loopy graph by converting it into a spanning tree. In the second one, we consider a slightly modified average consensus problem whose optimal solution approximates the consensus solution with arbitrary precision, and give a modified average consensus algorithm with guaranteed exponential convergence to the optimal solution. The proposed average consensus algorithms enjoy low complexities, robustness to transmission adversaries, and asynchronous implementation. Our algorithms are conceptually different from the popular graph Laplacian approach, and converge much faster than the latter approach.

Linear quadratic regulation (LQR) is very fundamental in modern control theory. The theory has been developing for several decades. However, a more general problem, the LQR problem for stochastic systems with multiple control channels and delays still remains outstanding. To better understand our idea, the paper focuses on dealing with the problem with two control channels and a delay, which is more difficult than the one with a single delayed control channel since the former actually encounters interaction of control channels besides the delay. A new value function is proposed, which is key for finding the predictor-like controller. The idea is also suitable for handling the LQR problem for Ito stochastic systems with multi-control-channel and multi-delay.

In this paper, we are concerned with a new class of mean field games which involve a finite number of agents. With help of conditional mathematical expectation, we obtain necessary and sufficient conditions for the existence of the decentralized open-loop Nash equilibrium for finite-population games. By decoupling a non-standard forward-backward stochastic differential equation, we design a set of decentralized strategies in term of two differential Riccati equations. Instead of the s-Nash equilibrium, the set of decentralized strategies is shown be a Nash equilibrium. Furthermore, we examine the infinite-horizon problem and give a neat condition for solvability of the related algebraic Riccati equation.

This paper addresses the finite-time H∞ filtering for a class of nonlinear singular nonhomogeneous Markov jump systems by T-S fuzzy approximation approach, where the transition probabilities (TPs) are time-varying and unknown. First, by considering a stochastic Lyapunov functional and rendering the time-varying TPs inside a polytope, a sufficient condition on singular stochastic H∞ finite-time boundedness (SSH∞FTB) for the filtering error systems is given. Then, by using the matrix inequality decoupling technique, a novel linear matrix inequality (LMI) condition on the existence of the finite-time H∞fuzzy filter is presented. The fuzzy filter is developed in terms of LMIs ensuring the filtering error system is SSH∞FTB. Compared with the previous ones, the proposed design method in this paper has more freedom, leading to less conservative results. A tunnel diode circuit is provided to illustrate the effectiveness and advantage of the design approach proposed in this paper.

This paper investigates two formation control problems for a leader-follower network in 3-D. One is called the formation marching control problem, the objective of which is to steer the agents to maintain a target formation shape while moving with the synchronized velocity. The other one is called the formation rotating control problem, whose goal is to drive the agents to rotate around a common axis with a target formation. For the above two problems, we consider directed and switching sensing topologies while the communication is assumed to be bidirectional and switching. We develop approaches utilizing barycentric coordinates toward these two problems. Local control laws and graphical conditions are acquired to ensure global convergence in both scenarios.

In this paper, a new distributed algorithm is presented to achieve the consensus of time variations and that of initial time simultaneously. By combining both controller and estimator design methods, it obtains higher synchronization precision with stronger robustness against noisy inputs resulted from crystal oscillators. Furthermore, the control input is ensured bounded to make our algorithm realizable in practical implementation. The implementation of the algorithm allows the receiving end to be event-triggered, while the transmitting end is executed periodically. The performance of the algorithm is illustrated by the given numerical simulations.

Despite of its wide success in many distributed statistical learning applications, the well-known Gaussian belief propagation (BP) algorithm still lacks sufficient understanding at the theoretical level. This paper studies the convergence of Gaussian BP by analyzing the dynamic behaviour of the marginal covariances. We show, under a mild technical assumption, that the information matrices (i.e., the inverses of marginal covariances) are guaranteed to converge exponentially to positive-definite matrices. The convergence rate is explicitly characterized. This result is a key step to the understanding of the dynamic behaviour of the BP iterations.

This paper reviews some main results and progress in distributed multi-agent coordination from a graph Laplacian perspective. Distributed multi-agent coordination has been a very active subject studied extensively by the systems and control community in last decades, including distributed consensus, formation control, sensor localization, distributed optimization, etc. The aim of this paper is to provide both a comprehensive survey of existing literature in distributed multi-agent coordination and a new perspective in terms of graph Laplacian to categorize the fundamental mechanisms for distributed coordination. For different types of graph Laplacians, we summarize their inherent coordination features and specific research issues. This paper also highlights several promising research directions along with some open problems that are deemed important for future study.

The paper studies the formation control problem for distributed robot systems. It is assumed each robot only has access to local sensing information (i.e. the relative positions and IDs of its neighbors). Taking into consideration physical sensing constraints (e.g. limited sensing range) and the motion of the robots over time, it may be noted that the sensing graph for the system is directed and time-varying. This presents a challenging situation for formation control. As an initiative attempt to study this challenging situation, we suppose the sensing graph switches among a family of graphs with certain connectivity properties, under which a switching linear control law is then proposed. We show that for arbitrary dwell times or average dwell times, the proposed control law with properly designed control parameters can ensure global convergence to a desired formation shape. The proposed formation control law can be implemented in a distributed manner while the design of certain control parameters requires some global information.

A distributed steady-state estimator is presented for linear interconnected systems. Stability of the distributed estimator is investigated. Sufficient conditions of stability are deduced based on the state and observation models. Some examples are provided to illustrate the relationship between the stability of the estimator and that of the original dynamics.

A randomly deployed sensor network is typically not completely localizable using distance-based measurements only. Though a necessary and sufficient condition for testing whether a network is localizable has been given in the literature, how to find localizable nodes from a not fully localizable network is still open. In this paper, we try to address a connection between two well-known localization methods, the trilateration method and the WHEEL extension method, by using a graphical tool named Henneberg operations. We also study whether Henneberg operations always guarantee the localizability of a network. The localizability by a Henneberg operation-based algorithm is given. Simulation shows that the performance of this algorithm for finding localizable nodes is very close to a well-known necessary condition called 3-path condition.

In this paper, we consider the quantized quadratic performance control problem for a class of stochastic systems which are subject to multiplicative noises in the measurement, we look for a dynamic output feedback controller to guarantee certain level of performance. By using the sector bound approach to characterize the quantized error, we show that the existence of the solution of quantized quadratic guaranteed cost problem can be found by solving the so-called guaranteed cost control problem of the associated system with sector bound uncertainty. The main result of this paper show that this problem can be effectively solved using linear matrix inequalities (LMIs).

Hysteresis and resonance dynamics are typical adverse effects associated with the piezoelectric (PZT) actuators. To eliminate the loss of positioning accuracy due to these effects, we propose a simple two degree-of-freedom (2DOF) control scheme for the PZT actuators, which consists of a feedforward and a feedback compensator for hysteresis and resonance compensation and for robust step reference tracking. Despite the proposed controller is on par with the existing 2DOF control methods, its specific structure offers significant design simplicity. The experimental results on an actual PZT nanopositioner show that the designed controller can be easily implemented and achieve superior performance for hysteresis and resonance compensation.

A common technical difficulty in target tracking in a wireless sensor network is that individual homogeneous sensors only measure their distances to the target whereas the state of the target composes of its position and velocity in the Cartesian coordinates. That is, the senor measurements are nonlinear in the target state. Extended Kalman filtering is a commonly used method to deal with the nonlinearity, but this often leads to unsatisfactory or even unstable tracking performances. In this paper, we present a new target tracking approach which avoids the instability problem and offers superior tracking performances. We first propose an improved noise model which incorporates both additive noises and multiplicative noises in distance sensing. We then use a maximum likelihood estimator for prelocalization to remove the sensing nonlinearity before applying a standard Kalman filter. The advantages of the proposed approach are demonstrated via experimental and simulation results.

This paper studies system identification of ARMA models whose outputs are subject to finite-level quantization and random packet dropouts. A simple adaptive quantizer and the corresponding recursive identification algorithm are proposed and shown to be optimal in the sense of asymptotically achieving the minimum mean square estimation error. The joint effects of finite-level quantization and random packet dropouts on identification accuracy are exactly quantified. The theoretic results are verified by simulations.

The hybrid filterbank architecture permits implementing accurate, high speed analog-to-digital converters. However, its design is technically involved since perfect reconstruction of the desired samples cannot be achieved in general. In this paper we propose a design method which generalizes existing approaches by dropping the bandlimited assumption on the input signal. The design minimizes the power of the reconstruction error in the samples, for a given input signal power spectrum. We discuss the use of blind techniques to estimate the analysis filterbank parameters as well as the input power spectrum, and we present simulation results to demonstrate the clear advantage of the proposed design, even under input spectrum uncertainties.

Quantized feedback control and quantized estimation have attracted a lot of attention in recent years with many results available on both research topics. In this paper, we investigate connections between quantized feedback control and quantized estimation and try to establish a possible separation principle for quantized output feedback which would allow the control design and state estimation to become independent in a networked control environment. We also consider the use of a variable rate finite-level logarithmic quantizer and show that this may approach the minimum averaged bit rate required for quantized feedback stabilization.

A linear system can be approximated in the time-frequency domain by the composition of an analysis filterbank, a transfer matrix (subband model) and a synthesis filterbank, a method known as subband technique. In this paper we propose an iterative method to jointly optimize the subband model, analysis and synthesis filterbank. To this end we propose a minimization criterion which we solve using the so-called alternating least-squares method. As a possible application we consider the implementation of the so-called head related transfer functions which are used in virtual acoustics. Simulation results suggest that the subband technique, optimized using the proposed method proposed method, is a promising approach.

Both direct and indirect methods exist for continuous-time system identification. A direct method estimates continuous-time input and output signals from their samples and then use them to obtain a continuous-time model, whereas an indirect method estimates a discrete-time model first. Both methods rely on fast sampling to ensure good accuracy. In this paper, we propose a more direct method where a continuous-time model is directly fitted to the available samples. This method produces an exact model asymptotically, modulo some aliasing ambiguity, even when the sampling rate is relatively low.

Narrow-band position error at mid-frequencies around the open-loop crossover frequency can not be effectively reduced using a conventional peak filter, because the attenuation of sensitivity gains has to compromise with the associated decrease of phase margin. This paper presents a general second-order filter that is applicable to reject narrow-band disturbances at any frequency range. The filter zero is designed to minimally degrade the closed-loop system stability and obtain a smooth sensitivity curve around the disturbance frequency. A nonlinear optimization procedure is developed to select the filter parameters such that the statistical position error is minimized. Experimental results of a PZT-actuated head positioning control system on spinstand demonstrate that the add-on filter can further reduce the mid-frequency PES NRRO by 8% and preserve the stability margin of the original control system

We propose a robust interpolation algorithm for model-based spectral estimation. The interpolation data represent information about the half spectrum function associated with a given signal and are computed from an input-to-state filter. Our algorithm allows a large number of noisy interpolation data to be used to optimally fit a half spectrum function of a fixed order. The algorithm involves solving a set of linear matrix inequalities and is thus numerically efficient.

This paper is concerned with density evolution for iterative decoding of low-density parity-check (LDPC) codes. We first study the problem of density evolution computation for regular LDPC codes. For this, we propose a simple computational algorithm based on the ergodicity theory. This method is shown to match very well with explicit calculations of density functions. The second problem we study is about the approach of Gaussian approximation to density evolution. We point out that it is inappropriate to use the mean of the density only to model the iterative decoding process. Instead, both the mean and variance are needed for Gaussian approximation. Finally, we consider the problem of density evolution for irregular LDPC codes. For this, we extend the density evolution algorithm for regular LDPC codes to irregular LDPC codes. We then illustrate that Gaussian approximation is also valid provided that the degree distributions are not wide. A dynamic model is also presented based on Gaussian approximation.

This paper presents a novel physical-modeling approach to wireless systems with multiple antennas. The fundamental problem of modeling the communication channel is studied, where the channel consists of a finite spatial volume for transmitting, a finite spatial volume for reception, and an arbitrary set of reflective-scattering bodies. The number of communication modes (or degrees of freedom) for such a system is calculated, using the procedure developed. We present a simple model for multipath channels, which allows insight into the development of a correlated multiple-input multiple-output (MIMO) channel model. In particular, the model is independent of transmitter and receiver elements and relies on the physical parameters of the channel involved. Our work explains which physical parameters determine the channel model and its channel capacity.

Constant envelope and high spectral efficiency are among the key properties which make continuous phase modulation (CPM) a preferred choice for many applications. Numerous methods are available for designing concatenated CPM systems to achieve additional coding gains. In this paper, we propose a new serial concatenated CPM scheme. Our design uses a multi-alphabet binary CPM as an inner code, a modified turbo code as an outer code, and specially designed interleavers. The extrinsic information transfer (EXIT) chart technique has been employed to assist the outer code design. Using soft decision decoding, we show that our design gives noticeable improvements compared with previous concatenated CPM schemes over additive white Gaussian noise (AWGN) channel.

This paper proposes a stochastic framework for modelling and analysis of turbo decoding. By modelling the input and output signals of a turbo decoder as random processes, we prove that these signals become ergodic when the code block size becomes very large. This basic result allows us to easily model and compute the statistics of the signals in a turbo decoder. Using the ergodicity result and the fact that a sum of lognormal distributions is well approximated using a lognormal distribution, we show that the input-output signals in a turbo decoder, when expressed using the so-called scaled log-likelihood ratios, are well approximated using Gaussian distributions. Combining the two results above, we can model a turbo decoder using two inputs and two outputs (corresponding to the means and variances). Using this model, we have discovered that a typical decoding process is much more intricate than previously known, involving two regions of attractions, several fixed points, and a stable equilibrium manifold at which all decoding trajectories converge.

Cyclic prefix (CP) is commonly used for channel equalization of discrete multitone (DMT) and orthogonal frequency-division multiplexing (OFDM) systems. This is often done in conjunction with a time-domain equalizer (TEQ) for reducing the capacity overhead caused by the CP. However, the use of TEQ greatly increases the computational cost, and is unable to eliminate the need for the CP. In this paper, we propose a subband approach to channel estimation and channel equalization for DMT and OFDM systems. This approach involves splitting the received signals into a number of frequency bands (called subbands), and estimating a constant parameter in each subband. The subband approach is conceptually simple, requires no CP, is much more numerically efficient than the TEQ method, and gives compatible or better estimation errors than the CP-based methods.

This paper proposes a stochastic framework for dynamic modeling and analysis of turbo decoding. By modeling the input and output signals of a turbo decoder as random processes, we prove that these signals become ergodic when the block size of the code becomes very large. This basic result allows us to easily model and compute the statistics of the signals in a turbo decoder. Using the ergodicity result and the fact that a sum of lognormal distributions is well approximated using a lognormal distribution, we show that the input-output signals in a turbo decoder, when expressed using log-likelihood ratios (LLRs), are well approximated using Gaussian distributions. Combining the two results above, we can model a turbo decoder using two input parameters and two output parameters (corresponding to the means and variances of the input and output signals). Using this model, we are able to reveal the whole dynamics of a decoding process. We have discovered that a typical decoding process is much more intricate than previously known, involving two regions of attraction, several fixed points, and a stable equilibrium manifold at which all decoding trajectories converge. Some applications of the stochastic framework are also discussed, including a fast decoding scheme

We consider the standard (point-wise) linear channel model for MIMO wireless systems in terms of a continuous operator channel. We show analytically that the point-wise representation overestimates the (true) modal connection strengths and produces artificially distorted channel singular values. Analytic results are compared with simulations for simple channel models and the convergence of the point-wise model toward the continuous model is shown.

For stochastic analysis of single-rate linear systems, a desirable property for stochastic signals is ergodicity in the mean and correlation. Unfortunately, as we show, the ergodicity property may not be preserved under downsampling and uniformly stable linear filtering. This poses a serious problem for stochastic analysis of multirate linear systems. We introduce the notion of strong ergodicity which is preserved under a number of important multirate operations including downsampling, upsampling and time-variant uniformly stable linear filtering. We provide conditions for stochastic processes to be strongly ergodic. Using this result, we show that both independent random processes and bounded deterministic signals are strongly ergodic in the mean and correlation.

This paper presents a new physical modeling approach for wireless systems with multiple antennas. The fundamental problem of modelling the communications channel if we are given an arbitrary spatial volume for transmitting, an arbitrary spatial volume for receiving, and a set of scattering bodies is studied. We show how to calculate the number of communication modes, both for direct (point-to-point) and for scattering environments. Our work explains the physical parameters, which determine the channel model and its channel capacity.

This paper deals with the problem of computing channel capacity for wireless communications systems with multi-antennas. The system is modeled using a transmit volume, a receive volume and a set of reflective scatterers. Based on the author's previous work on channel modeling of such a system, we give a multi-input-multi-output model for communications between arbitrary antenna arrays in two volumes. Together with a novel noise model for describing interferences, we derive formulas for channel capacity and capacity bounds. In particular, our results give conclusions which agree with intuition: (i) in the case where the transmitter knows the channel, the channel capacity saturates to a constant when the number of transmit/receive antennas increases beyond a certain threshold; (ii) in the case where the transmitter des not know the channel, the channel capacity still saturates when the receive antennas increases but decays when the number of transmit antennas increases beyond a certain threshold.

In this paper we propose two equalization methods for orthogonal frequency-division multiplexing (OFDM) modulation. These methods are intended to be alternatives to the known cyclic prefix equalization (CPE) method, and are derived using ideas from subband identification. The first proposed method is called subband equalization (SBE) and the second one is called enhanced cyclic prefix equalization (ECPE). The performance of different methods is compared in terms of bit error rate (BER) after equalization, convergence rate and computational cost in the presence of noise and fast fading. Both SBE and ECPE outperform the CPE in terms of BER and convergence rate. The SBE does not require the use of CP, and therefore saves channel capacity, but needs more computation than CPE, ECPE requires the use of CP, but its cost is compatible to that of CPE.

Integral quadratic constraints (IQC) arise in many optimal and/or robust control problems. The IQC approach can be viewed as a generalization of the classical multiplier approach in the absolute stability theory. In this paper, we study the relationship between the two approaches for robust stability analysis. The key result shows that for many applications, the existence of an IQC is equivalent to the existence of a multiplier. Because the multiplier approach is typically simpler and more intuitive, this result suggests that the multiplier approach may be more useful than the IQC approach in many applications.

We present a general multiple path scattering model for multiple transmit, multiple receive wireless systems. The model is generated using physical modelling of the scatterers surrounding transmit and the receive arrays. The condition of the resulting multiple-input multiple-output (MIMO) transfer matrix is then examined. A key parameter /spl eta/ which determines the channel condition is identified. This parameter depends on the local scatter geometries, the separation of arrays, and the wavelength of the transmitted signal. We show that there exists a critical value for this parameter at which the channel condition changes sharply. The implication is that the promised linear growth in channel capacity may not eventuate if the separation of the transmit and receive arrays is large (/spl ap/10/spl times/ or 20/spl times/ at 2 GHz) compared with the distance from array elements to local scatterers. Monte-Carlo simulations are used to demonstrate these claims.

We study the problem of filterbank design for the subband identification method in the oversampled case. We aim to design the filterbanks such that the coupling between adjacent subbands is minimized. This allows the subband model of the system to be diagonal, thus simplifying the complexity of identification. Solutions are given to two minimization criteria: the worst-input case and white input case.

The adaptive-subspace self-organizing map (ASSOM) proposed by Kohonen is a recent development in self-organizing map (SOM) computation. In this paper, we propose a method to realize ASSOM using a neural learning algorithm in nonlinear autoencoder networks. Our method has the advantage of numerical stability. We have applied our ASSOM model to build a modular classification system for handwritten digit recognition. Ten ASSOM modules are used to capture different features in the ten classes of digits. When a test digit is presented to all the modules, each module provides a reconstructed pattern and the system outputs a class label by comparing the ten reconstruction errors. Our experiments show promising results. For relatively small size modules, the classification accuracy reaches 99.3% on the training set and over 97% on the testing set.

The so-called analytic center approach to parameter estimation has been proposed as an alternative to the well-known least squares approach. This new approach offers a parameter estimate that is consistent with the past data observations, has a simple geometric interpretation, and is computable sequentially. We study the asymptotic performance of the analytic center approach and show that the resulting estimate converges to the true parameter asymptotically, provided some mild conditions are satisfied. These conditions involve some weak persistent excitation and independence between noise and regressor, similar to the least squares case. This result is used to derive a new parameter estimation approach which offers both good transient and asymptotic performance.

The problem of passivity analysis finds important applications in many signal processing systems such as digital quantizers, decision feedback equalizers and digital and analog filters. This paper considers the passivity analysis problem for a large class of systems which involve uncertain parameters, time delays, quantization errors, and unmodeled high order dynamics. By characterizing these and many other types of uncertainty using a general tool called integral quadratic constraints (IQCs), we present a solution to the problem of robust passivity analysis. More specifically, we determine if a given uncertain system is robustly passive. The solution is given in terms of the feasibility of a linear matrix inequality (LMI) which can be solved efficiently.

Designing optimal filter banks for subband coding applications has recently attracted considerable attention. In particular, the authors have developed an adaptive algorithm based on stochastic gradient descent (SGD) that enables one to optimize two channel paraunitary filter banks in an on-line fashion. The idea has also been extended to the case of tree-structured filter banks. The computational complexity of the algorithm proposed is proportional to N/sup 2/ where N is the number of stages in the paraunitary lattice. In this paper we derive a fast algorithm which reduces the amount of computation to O(N). We also show that the new algorithm can be implemented using an IIR lattice. Some issues regarding numerical stability of the IIR implementation are also discussed.

This paper describes a modular classification system for handwritten digit recognition based on the elastic net model. We use ten separate elastic nets to capture different features in the ten classes of handwritten digits and represent an input sample from the activations in each net by population decoding. Compared with traditional neural networks based discriminant classifiers, our scheme features fast training and high recognition accuracy.

This paper propose a vector quantization (VQ) technique to solve the problem of handwritten signature verification. A neural 'gas' model is trained to establish a reference set for each registered person with handwritten signature samples. Then a test sample is compared with all the prototypes in the reference set and the system outputs the label of the writer of the word. Several different feature extraction methods are compared and good results have been obtained by the VQ technique.

We consider a multi-input, multi-output lattice realization for linear time-varying analysis banks which are all pass. Such a realization was given by Vaidyanathan and Mitra (1985) for linear time invariance (LTI) systems; and under certain conditions generalizes to the linear time varying (LTV) case. Moreover, our implementation is simpler than the one presented by Vaidyanathan et al. Finally, we describe the anticausal inverse of a lattice realization which is used in the synthesis bank.

A procedure for the adaptation of multistage multirate filters is developed in the context of a system identification perspective. The system to be identified is the cascade of a multistage decimator followed by a multistage interpolator. For this structure, an identifiability condition is established which is sufficient to guarantee that the stagewise systems may be uniquely determined from input-output data in the ideal, exact model order case. An algorithm is then described that achieves exponential convergence of the system parameter estimates to their desired values given satisfaction of these identifiability conditions.

This study investigates the usefulness of wavelet transforms in phoneme recognition.Both discrete wavelet transforms (DWT) and sampled continuous wavelet transforms (SCWT) are tested.The wavelet transform is used as a part of the front-end processor which extracts feature vectors for a speakerindependent HMM-based phoneme recognizer.The results are evaluated on a portion of TIMIT corpus consisting of 30293 phoneme tokens for training and 14489 phoneme tokens for testing.The test results suggest that SCWT gives considerably better recognition rate than DWT.On the other hand, the improvement of SCWT over Mel-scale cepstral coecients appears to be marginal.

In this paper we studied a self-organization principle that input should be best reconstructed from a factorial distributed hidden representation, which has been addressed in the literature recently. An auto-encoder network is trained by the Least Mean Square Error Reconstruction (LMSER) while the redundance in the representation is reduced by a proposed anti-Hebbian scheme, in which a penalty term called Receptive Field Overlapping Index (RFOI) is combined into the objective function for enhancing competition among nodes in the network. Our learning scheme provides a way for balancing the cooperation and competition necessary for the self-organization process thus realizes the multiple causes model, which accounts for an observed data by combining assertions from the discovered causes or features in the data. Our experiment results demonstrate again the powerful information processing capability inherent to the popular weighted sum followed by sigmoid squashing. Comparing with previous probability theory based multiple causes models, our scheme is much easier to implement and quite reliable.