My Refereed Publications and Abstracts

Most of the papers are available on request, others can be found on line via the usual providers


2007

Kouchakpour, P., Zaknich, A., and Braunl, T.

Population variation in genetic programming, Information Sciences, Vol. 177, No. 17, pp. 2438-3452, September 2007.

 

A population variation scheme is proposed, where by the size of the population is varied during the execution of the genetic programming process with the aim of reducing the computational effort with respect to that of Standard Genetic Programming (SGP). Within this new scheme the initial population size is varied with respect to the initial population size of the SGP such that the worst case computational effort is never greater than that of the SGP. Various schemes for altering population size under this proposal are investigated using a comprehensive range of standard problems to determine whether the nature of the "population variation", i.e. the way the population is varied during the search, has any significant impact on GP performance. It is subsequently shown that the proposed population variation schemes do have the capacity to provide solutions at a lower computational cost compared with the SGP.

 

Farrokhi, D., Togneri, R., and Zaknich, A.

Speech enhancement of non-stationary noise based on controlled forward march averaging

International Symposium on Communications and Information Technologies 2007 (ISCIT), Sydney, Australia, pp. 1551-1555, 17-19th October 2007.

 

A pre and post processing technique is proposed to enhance the speech signal of highly non-stationary noisy speech. The purpose of this research has been to buils on current speech enhancement algorithsm to produce an improved algorithm for enhancement of speech contaminated with non-stationary babble type noise. The pre processing involves two stages. In stage one, the variance of the noisy speech spectrum is reduced by untilizing the Discrete or Prolate Spheroid Sequence (DPSS) multi-taper algorithm plus a Controlled Forward Moving Average (CFMA) technique. We introduced the CFMA algorithm to smooth and reduce variance of the estimated non-stationary noise spectrum. In the second stage the noisy speech power spectrum is de-noised by applying Stein’s Unbiased Risk Estimator (SURE) wavelet thresholding technique. In the third layer, use is made of a noise estimation algorithm with rapid adaptation for a highly non-stationary noise environment. The noise estimate is updated in three frequency sub-bands, by averaging the noisy speech power spectrum using a frequency dependent smoothing factor, which is adjusted, based on a signal presence probability factor. In the fourth layer a spectral subtraction algorithm is used to enhance the speech signal, by subtracting each estimate dnoise from the original noisy speech. The new proposed processing is then applied to the complete signal when the speech enhancement is processed using segmental speech enhancement. The enhanced signal is further improved by applying a soft wavelet thresholding technique to the un-segmented enhanced speech at the final processing stage. The results shows improvements both quantatively and qualitatively compaed to the speech enhancement that does not apply the CFMA algorithm.

 

Lee, G. E., Bahri, P. A., Shastri, S. S., and Zaknich, A.

A multi-category decision support framework for the Tennessee Eastman problem,” European Control Conference, Kos, Greece, Vol. ?, No. ?, pp. ?-?, 2-5th July 2007. Accepted for Publication on 24-1-07

 

The paper investigates the feasibility of developing a classification framework, based on support vector machines (SVM), with the correct properties to act as a decision support system for an industrial process plant, such as the Tennessee Eastman process. The system would provide support to the technicians who monitor plants by signalling the occurrence of abnormal plant measurements marking the onset of a fault condition. To be practical such a system must meet strict standards, in terms of low detection latency, a very low rate of false positive detection and high classification accuracy. Experiments were conducted on examples generated by a simulation of the Tennessee Eastman process and these were preprocessed and classified using a support vector machine. Experiments also considered the efficacy of preprocessing observations using Fischer Discriminant Analysis and a strategy for combining the decisions from a bank of classifiers to improve accuracy when dealing with multiple fault categories.

 

Lee, C. S., Braunl, T., and Zaknick, A.

An Adaptive T-S type Rough-Fuzzy Inference System (ARFIS) for Pattern Classification, IEEE International Conference of the North American Fuzzy Information Processing Society  (NAFIPS), San Diego, USA, pp. 117-122, 24-27th June 2007.

 

The Rough-Fuzzy hybridization scheme has become of research interest in pattern classification over the past decade. The present paper proposes a new Adaptive T-S type Rough-Fuzzy Inference System (ARFIS) for pattern classification. Rough set theory is utilized to reduce the number of attributes and also to obtain a minimal set of decision rules based on input-output data sets. A T-S type fuzzy inference system is constructed by the automatic generation of membership functions and rules by the fuzzy c-means clustering algorithm and rough set theory, respectively. The generated T-S type rough-fuzzy inference system is adjusted by the least-squares fit and a conjugate gradient descent algorithm towards better performance with a validity checking for the minimal set of rules. The proposed ARFIS is able to reduce the number of rules which increases exponentially when more input variables are involved and also to assess the validity of the minimized decision rules. The performance of the proposed ARFIS is compared with other existing pattern classification schemes using Fisher's Iris and Wisconsin breast cancer data sets and shown to be very competitive.

 

Legg, M.W., Duncan, A.J., Zaknich, A., and Greening, M.V.

Analysis of impulsive biological noise due to snapping shrimp as a point process in time

 MTS/IEEE Conference Oceans07-Europe, Aberdeen, Scotland, pp. 1-6, 18-21st June 2007.

 

Impulsive biological noise produced by snapping shrimp provides an important contribution to the ambient acoustic noise in warm, coastal waters. The challenge is to understand and model the properties of shrimp noise to reduce the impact on sonar and underwater acoustic telemetry systems. Shrimp snaps are impulsive events occurring apparently at random. The short duration of each snap allows these events to be modeled as a point process in time. Point processes are used to model many naturally occurring phenomena including neuron firings, seismic events, radioactive decay, lightning discharges and shot noise in semiconductors. In this paper, point process analysis techniques are applied to real shrimp noise. Inter-snap interval histogram and Fano-factor analysis provide strong evidence that the snaps are not homogeneous Poisson distribution distributed in time. Further analysis based on the rate function suggests that the data may be more appropriately modeled by a doubly stochastic Poisson process.

 

Haque, S., Togneri, R., and Zaknich, A.

A temporal auditory model with adaptation for automatic speech recognition,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Honolulu, USA, Vol. IV, pp. 1141-1144, 15-20th April 2007.


Rapid and short-term adaptation and dynamic are mechanisms of human auditory system. An auditory model based on zero-crossings with peak amplitudes (ZCPA) was used as a front-end for automatic speech recognition (ASR) with the perceptual property of adaptation as determined by psychoacoustic observations. The model performance was evaluated on the isolated digits (TIDIGITS) database using continuous density HMM recogniser in additive noise environment. Experimental results indicate that the ASR performance of the ZCPA may be improved with adaptation over the static baseline performance in white Gaussian and factory noise. The perceptual front-end was also evaluated with dynamic (delta and delta-delta) features added to the adaptation. It was observed that adaptation with dynamic features performed better in factory, babble and car noise over a wide range of SNR values.

 

Hanselmann, T., Noakes, L., and Zaknich, A.

Continuous adaptive critics, IEEE Transactions on Neural Networks, Vol. 18, No. 3, pp. 631-647, March 2007.

 

A continuous-time formulation of an adaptive critic design (ACD) is investigated. Connections to the discrete case are made, where backpropagation through time (BPTT) and real-time recurrent learning (RTRL) are prevalent. Practical benefits are that this framework fits well with plant descriptions and that any standard integration routine with adaptive step-size does and adaptive sampling for free. A second-order actor-critic adaptation using Newton’s method is established for fast actor convergence for a general plant and critic. Also, a fast critic update for concurrent actor-critic training is introduced to immediately apply necessary adjustments of critic parameters induced by actor updates to keep the Bellman optimality correct to first-order approximation after actor changes. Thus, critic and adaptor updates may be performed at the same time until some substantial error build up in the Bellman optimality or temporal difference equation, when a traditional critic training needs to be performed and then another interval of concurrent actor-critic training may resume. 

2006

Haque, S., Togneri, R., and Zaknich, A.

Zero-crossings with adaptation for automatic speech recognition.

Eleventh Australasian International Conference on Speech Science and Technology, Auckland, New Zealand, pp. 199-204, 6-8th December 2006.

 

An auditory model based on zero-crossings with peak amplitudes (ZCPA) was used as a front-end for automatic speech recognition (ASR) with the perceptual property of adaptation as determined by psychoacoustic observations, The model performance was evaluated on the isolated digits (TIDIGITS) database using continuous density HMM recogniser in additive noise. Experimental results indicate that the ASR performance of the ZCPA may be improved with adaptation over the static baseline performance in white Gaussian and factory noise. The perceptual front-end was also evaluated with dynamic (delta and delta-delta) features added to the adaptation. It was observed that adaptation with dynamic features performed better in factory, babble and car noise over a wide range of SNR values. The recognition performances were compared with the baseline MFCC. The performance of the dynamic ZCPA with adaptation was better than the dynamic MFCC in white noise.

 

Zaknich, A., and Lee, G.E.

An audio equalisation linear phase FIR filter design method using RBF based smoothing and interpolation, Fourth International Conference on Intelligent Sensing and Information Processing (ICISIP), Bangalore, India, pp. 109-114, 15-18th December 2006.

 

A method for the design of audio equalisation linear phase FIR filters is developed by using a variant of the Tuneable Approximate Piecewise Linear Regression (TAPLR) method to model the required FIR magnitude frequency response. This TAPLR model incorporates a set of contiguous piecewise linear (affine) sections which are coupled and smoothed by a single tuning parameter associated with a set of bandwidth weighted Radial Basis Functions (RBFs) assigned to each linear section. The main difference between this variant and the original TAPLR method is that it incorporates RBFs with variable bandwidths that can be centred and set according to standard nonlinear frequency spacings as used for audio system response measurements, thereby producing a more accurate response model. The TAPLR smoothing mechanism is used to achieve the required degree of FIR filter band limiting to avoid aliasing effects and the Gibbs phenomenon. Some typical audio FIR filter design examples are provided to show the value and versatility of the method as compared with the standard windowing design approach.

 

Zaknich, A., and Lee, G.E.

Arbitrary audio FIR filter design by Bode plot smoothing using tuneable approximate piecewise linear regression,” Joint New Zealand Acoustical Society and Australian Acoustical Society Conference, Christchurch, New Zealand, pp. 219-223, 20-22nd November 2006.

 

A method for the design of arbitrary minimum or linear phase FIR filters is developed for audio applications by using the Tuneable Approximate Piecewise Linear Regression (TAPLR) method to smooth the required FIR magnitude frequency response from a prototype Bode plot model. The TAPLR method incorporates a set of contiguous piecewise linear (affine) sections, which are coupled and smoothed by a single tuning parameter associated with a set of amplitude weighted Radial Basis Functions (RBFs) assigned to each linear section. The Bode plot also consists of a set of contiguous linear asymptotes plotted on a log-log scale, which makes it a perfect candidate for modelling and smoothing by the TAPLR method. The TAPLR smoothing turns the artificial asymptotic magnitude plot into a realisable magnitude response curve, which can be made to be band-limited with a finite impulse response by the appropriate degree of smoothing. A typical FIR filter design example for audio system equalisation is provided to show the value and versatility of the method. Also,  two Bode plot filter prototypes are presented to show how well the new modelling approach can capture them and adapt them to suitable band-limited FIR filter designs.

 

Lee, C. S., Braunl, T., and Zaknich, A.

A Rough-Fuzzy controller for autonomous mobile robot navigation, IEEE International Conference on Intelligent Systems (ICIS), London, United Kingdom, pp. 679-682, 4-6th September 2006.

 

This paper presents a new development of a rough-fuzzy controller for an autonomous mobile robot based on rough set and fuzzy set theory. It has been tested in different environments with the Saphira simulation software. The proposed approach provides an improvement in uncertainty reasoning by using a rough-fuzzy controller, resulting in better wall-following behavior performance as compared against other controllers. The rough-fuzziness of the input data leads to the enhanced uncertainty reasoning process by calculating the roughly approximated fuzzified value of the input, which makes the system more robust and reliable

2005

Legg, M.W., Duncan, A.J., Zaknich, A., and Greening, M.V.

An exploratory analysis of non-Poisson temporal behaviour in snapping shrimp noise.

Annual Conference of the Australian Acoustical Society, Acoustics 2005, Busselton, Western Australia, pp. 399-403, 9th-11th November 2005.

 

Snapping shrimp are a well known interference source for underwater sonar and communication systems, particularly in shallow and harbour waters.  The noise produced by snapping shrimp is highly impulsive and the amplitude statistics are non-Gaussian.  Impulsive noise is most often modelled in a way that implicitly assumes that the temporal statistics are Poisson.  The Poisson assumption implies that a snap from any shrimp is completely independent of snaps from other shrimp.  This paper reports on an exploratory analysis of non-Poisson temporal behaviour in snapping shrimp noise using real acoustic data from different geographic locations in Australian coastal waters.  The analysis makes use of various statistical techniques applied to snaps detected in high-pass filtered data using a threshold technique.  Attempts are made to eliminate multi-path effects, which can introduce correlations between snap arrivals, from other possible effects such as interactions between shrimp.  The results are compared and contrasted between different geographic locations.

 

Zaknich, A.

A loudspeaker response interpolation model based on one-twelfth octave interval frequency acoustic measurements.

Annual Conference of the Australian Acoustical Society, Acoustics 2005, Busselton, Western Australia, pp. 133-137, 9th-11th November 2005.

 

A practical loudspeaker frequency response interpolation model is developed using a modification of the Tuneable Approximate Piecewise Linear Regression (TAPLR) model that can provide a complete magnitude and phase response over the full frequency range of the loudspeaker. This is achieved by first taking standard one-twelfth octave frequency interval acoustic intensity measurements at a one meter distance in front of the loudspeaker. These measurements are inserted directly into the formulation, which then requires only minimal tuning to achieve an magnitude response model to better than +/- 1 dB error as compared with the magnitude of the Fourier transform of the impulse response for typical hi-fi loudspeakers. The Hilbert transform can then be used to compute the corresponding phase response directly from the resulting magnitude response. Even though it is initially based on consecutive piecewise linear sections this new model provides a continuous smooth interpolation between the measured values that is much more satisfactory than normal piecewise linear segment interpolation and much simpler to do than polynomial interpolation. It only requires the tuning of a single parameter to control the degree of smoothness from a stair step response at one extreme to a straight mean horizontal line at the other. It is easy to find the best tuning parameter value in between these two extremes by either trial and error or by the minimisation of a mean squared interpolation error.

 

Haque, S., Togneri, R., and Zaknich, A.

A zero-crossing perceptual model for robust speech recognition.

Inter-University Postgraduate Electrical Engineering Symposium, Edith Cowan University, Perth, Western Australia, pp. 60-65, 27th September 2005.

 

The traditional speech recognition systems based on linear prediction and spectral/cepstral analysis can only partially fulfil the speech recognition as it severely degrades under noise and environmental mismatched conditions.  Alternatively, auditory models,  based on the properties of human sound perception in the peripheral auditory system, depend on the time-frequency response of the basilar membrane, the neural activity pattern and other observed psychophysical properties of hearing. The disadvantage with these models are that they are computationally intensive and are dependent on several free parameters such as zero crossing level values, derivative window lengths and the number of frequency bins which are frequently selected by trial and error. In this paper, an auditory model based on the zero crossing peak amplitude, which is reasonably low order, computationally efficient and implemented with minimum choice of parameters is presented. It is shown that for an isolated digit recognition task, improved performances can be obtained compared to the Mel Frequency Cepstral Coefficients (MFCC) and the Perceptual Linear Prediction (PLP) methods in presence of  additive Gaussian noise.

 

Hanselmann, T., Noakes, L., and Zaknich, A.

Continuous adaptive critic designs.

IEEE-INNS International Joint Conference on Neural Networks (IJCNN), Montreal, Canada, Vol. 5, pp. 3001-3006, 31st July-4th August 2005.

 

A continuous formulation of an adaptive critic design (ACD) is investigated. Connections to the discrete case are made, where backpropagation through time (BPTT) and real-time recurrent learning (RTRL) are prevalent. A second order actor adaptation, based on Newton’s method, is established for fast actor convergence. Also a fast critic update for concurrent actor-critic training is outlined that keeps the Bellman optimality correct to first order approximation after actor changes.

 

Zaknich, A.

Principles of adaptive filters and self-learning systems.

Springer-Verlag, Series on Advanced Textbooks in Control and Signal Processing, 22nd July 2005, ISBN 1-85233-984-5.

 

This book can be used as a textbook for a one semester undergraduate or postgraduate introductory course on adaptive and self-learning systems for signal processing applications. Important topics are introduced and discussed sufficiently to give the reader adequate background to be able to confidently pursue them at depth in more advanced literature. Tutorial problems and exercises identify the significant points and are designed to demonstrate the practical relevance of the theory. Answers to the tutorials are given to aid topic understanding. To further facilitate course preparation a typical course outline and sample examination material is provided. The book’s topics are presented in a progressive sequence from a short introduction to adaptive filters, linear systems, and stochastic process theory, to system and signal modelling, the classical Wiener filter, the Kalman filter, spectral analysis theory, classical adaptive filters, adaptive control systems and the application of adaptive filtering, through to  nonclassical adaptive systems. It offers a comprehensive review of linear and stochastic theory as well as a design guide for the application of the least squares estimation method and Kalman filters. Although the book concentrates on the more established adaptive systems theory an introduction to neural networks, fuzzy logic and genetic algorithms as adaptive systems is also included to provide a more generic perspective to the topic of adaptive learning. A significant further offering of the book is a method to seemlessly combine a set of both classical and/or nonclassical adaptive systems to form a powerful self-learning engineering solution method. This method is referred to as sub-space adaptive filtering and is suitable for solving very complex nonlinear problems typical of the underwater acoustic signal processing and other equally difficult application domains.

2004

Hanselmann, H., Noakes, L., Zaknich, A., and Savkin, A.

A hybrid dynamical system with robust switching control by action dependent heuristic dynamic programming.

IEEE International Joint Conference on Neural Networks (IJCNN), Budapest, Hungary, Vol. 3, pp. 1799-1804, 25-29th July 2004.

 

In this paper a hybrid dynamical system with linear plant characteristics but unknown state, disturbance and observation inputs is considered and controlled by switching between fixed linear output feedback controllers. Using state estimation based on Kalman filtering and solving a Riccati equation, a dynamic programming solution based on estimated state can be obtained and a switching sequence for output feedback controllers can be deduced. However, solving the dynamic programming equation is difficult in practice due to the ‘curse of dimensionality’. Action Dependent Heuristic Dynamic Programming (ADHPD), also known as Q-learning, is applied to achieve an approximate dynamic programming solution based on piecewise quadratic interpolation and explicit determination of extremal values.

 

Zaknich, A.

A loudspeaker response model using tuneable approximate piecewise linear regression.

IEEE International Joint Conference on Neural Networks (IJCNN), Budapest, Hungary, Vol. 4, pp. 2711-2716, 25-29th July 2004.

 

A practical loudspeaker frequency response interpolation model is developed using a Tuneable Approximate Piecewise Linear Regression (TAPLR) model that can provide a complete amplitude and phase response over the full frequency range of the loudspeaker. This is achieved by taking a finite number of standard one-twelfth octave frequency amplitude measurements at a one meter distance in front of the loudspeaker. The Hilbert transform can be used to compute corresponding phase response from the amplitude response.

2003

Zaknich, A.

An integrated sensory-intelligent system for underwater acoustic signal-processing applications.

IEEE Journal of Oceanic Engineering, Vol. 28, No. 4, pp. 750-759, October 2003.

 

A generic Integrated Sensory-Intelligent System (ISIS) is developed for underwater acoustic signal processing applications. ISIS constantly monitors the current acoustic channel conditions and smoothly integrates the outputs of the most appropriate signal processing procedures or algorithms available to it for those conditions. The system is based on a generalisation of a Tuneable Approximate Piecewise Linear (TAPL) model derived from the Modified Probabilistic Neural Network (MPNN). This model was designed to seamlessly integrate a set of local linear signal processing algorithms within a given multidimensional data space. Depending on the input signal distortions, determined by environmental effects, ISIS automatically weighs and adds the outputs from a set of processing algorithms working in parallel. The weighting is related to the “closeness” of each algorithm to the sensed input signal characteristics or some other measured environmental state. A single tuning parameter is used to smoothly and seamlessly select appropriately amongst the parallel processing algorithm outputs. A very small tuning parameter value selects the closest most appropriate algorithm output. At the other extreme a fixed, weighted average of all the algorithm outputs is produced with a very large value. Otherwise, a dynamic weighed average of all algorithm outputs is achieved with values in between. Some features and benefits of ISIS are demonstrated with an illustrative linear sweep chirp signal detector estimation problem characterised by extreme variable Doppler conditions.

 

Zaknich, A.

A practical sub-space adaptive filter.

Neural Networks, Vol. 16, Nos 5/6, pp. 833-839, June/July 2003.

 

A Sub-Space Adaptive Filter (SSAF) model is developed using, as a basis,  the Modified Probabilistic Neural Network (MPNN) and its extension the Tuneable Approximate Piecewise Linear Regression (TAPLR) model. The TAPLR model can be adjusted by a single smoothing parameter continuously from the best piecewise linear model in each sub-space to the best approximately piecewise linear model over the whole data space. A suitable value in between ensures that all neighbouring piecewise linear models merge together smoothly at their boundaries. This model was developed by altering the form of the MPNN, a network used for general nonlinear regression. The MPNN’s special  structure allows it to be easily used to model a process by appropriately weighting piecewise linear models associated with each of the network’s radial basis functions. The model has now been further extended to allow each piecewise linear model section to be adapted separately as new data flows through it. By doing this, the proposed SSAF model represents a learning/filtering method for nonlinear processes that provides one solution to the stability/plasticity dilemma associated with standard adaptive filters.

 

Zaknich, A.

An adaptive sub-space filter model.

IEEE International Joint Conference on Neural Networks (IJCNN), Portland, Oregon, USA, Vol. 2, pp. 1464-1468, 20-24th July 2003.

 

A tuneable approximate piecewise linear regression model [1] has been previously developed, which can be adjusted by a single smoothing parameter continuously from the best piecewise linear model in each sub-space to the best approximate piecewise linear model over the whole data space. A suitable value in between ensures that all neighbouring piecewise linear models merge together smoothly at their boundaries. This model was developed by making relatively minor changes to the form of the Modified Probabilistic Neural Network [2], a network used for general nonlinear regression. The special Modified Probabilistic Neural Network (MPNN) structure allows it to be easily used to model a process by appropriately weighting piecewise linear models associated with each of the network’s radial basis functions, which together cover the data space. The model has now been further extended to allow each piecewise linear model section to be adapted separately as new data flows through it. By doing this, the proposed Adaptive Sub-Space Filter (ASSF) model represents a learning/filtering method for nonlinear processes that provides a solution to the stability/plasticity dilemma associated with standard adaptive filters.

 

Zaknich, A.

Neural networks for intelligent signal processing.

World Scientific Publishing, Series on Advanced Biology and Logic-Based Intelligence, Vol. 4, 2003, ISBN 981-238-305-0.

 

This book provides a thorough theoretical and practical introduction to the application of neural networks to pattern recognition and intelligent signal processing. It has been tested on students, unfamiliar with neural networks, who were able to pick up enough details to successfully complete their masters or final year undergraduate projects. The text also presents a comprehensive treatment of a class of neural networks called common bandwidth spherical basis function NNs, including the probabilistic NN, the modified probabilistic NN and the general regression NN.

2002

Young, J., Hanselmann, T., Zaknich, A., and Attikiouzel, Y.

Fine tuning the algebraic perceptron equaliser to increase the separation margin.

Inter-University Postgraduate Electrical Engineering Symposium, Murdoch University Rockingham Campus, Perth, Western Australia, 2nd October 2002.

 

The Algebraic Perceptron (AP) maps inputs vectors to a high dimensional feature space where linear separation is possible. Using a geometrical separation technique, an arbitrary hyperplane can be constructed using a very small number of data points. However, the margin of the separating hyperplane is not always close to the maximal margin. Therefore, the generalisation error is not always minimised. In this paper we propose a method to tune a kernel parameter to increase the non-optimal margin found by the AP algorithm. Simulations have confirmed improvement in terms of robustness and bit-rate error-rate. Comparison with the Support Vector Machine equaliser shows that for similar bit-rate-error, the tuned AP equaliser is superior with respect to the network size and training time.

2001

Young, J., Hanselmann, T. Zaknich A., and Attikiouzel, Y.
Adaptive complex modified probabilistic neural network in digital channel equalisation.
IEEE 8th Australian and New Zealand International Conference on Intelligent Information Processing Systems, Perth, Western Australia, Vol 1, pp. 247-251, 18-21st November 2001.

A novel adaptive technique is proposed for the complex-valued Modified Probabilistic Neural Network (MPNN). The adaptive feature is desirable when using the MPNN in channel equalization to track time-varying channels. The MPNN is initially trained using the clustering technique. When training is completed, the network is switched to decision directed mode and the network parameters are adapted using stochastic gradient-based algorithms in an unsupervised manner. Simulations show that the equalizer was able to efficiently equalize 4-QAM symbol sequences transmitted through non-linear, slowly time-varying channels.

Young, J., Hanselmann, T., Zaknich A., and Attikiouzel, Y.
Algebraic perceptron in digital channel equalization.
IEEE International Joint Conference on Neural Networks (IJCNN), Washington, DC, USA, pp. 2889-2892, 14-19th July 2001.

This paper investigates the application of the Algebraic Perceptron to solve the problem of channel equalization. The focus is on the particular case where the degree of intersymbol interference is severe. In recent years, some researchers have applied the Support Vector Machine to the same application and found valuable results. However, the Support Vector Machine requires solving a constrained optimisation problem with quadratic programming, which is not a trivial task. Like the Support Vector Machine, the Algebraic Perceptron also achieves linear separation in the high dimensional feature space, but with much reduced calculation requirement. The tradeoff is that the separation surface is not a maximal margin one. In the simulation, it was found that for some channels the Algebraic Perceptron performed better than the Support Vector Machine. Further, given a more complete training set, the performance of the Algebraic Perceptron can match the performance of the Support Vector Machine.

Young, Y., Zaknich, A., and Attikiouzel, Y.
Center reduction algorithm for the modified probabilistic neural network equalizer.
IEEE International Joint Conference on Neural Networks (IJCNN), Washington, DC, USA, pp. 1966-1970, 14-19th July 2001.

The applicability of the Modified Probabilistic Neural Network (MPNN) to channel equalization can be severely limited by the size of the network. The size of the network grows exponentially with the order of the channel and the dimension of the input vectors. As a result, the standard network is practical only for low order channels with small input alphabet size. An algorithm is proposed to alleviate such an undesirable constraint by finding a much smaller network representation with a similar decision surface.

2000

Zaknich, A., and Attikiouzel, Y.
A tuneable approximate piecewise linear model derived from the modified probabilistic neural network.
IEEE Signal Processing Workshop on Neural Networks for Signal Processing (NNSP), Sydney, Australia, pp. 45-53, 11-13th December 2000.

A simple model which can be adjusted by a single smoothing parameter continuously from the best piecewise linear model in each linear subregion to the best approximately piecewise linear model overall is developed for multivariate general nonlinear regression. The model provides an accurate, smooth approximately piecewise linear model to cover the entire data space. It provides a logical basis for extrapolation to regions not represented by training data, based on the closest piecewise linear model. This model has been developed by making relatively minor changes to the form of the Modified Probabilistic Neural Network, which is a network used for general nonlinear regression. The Modified Probabilistic Neural Network structure allows it to be used to model data by weighting piecewise linear models associated with each of the network's radial basis functions in the data space.

Jan, T., Zaknich, A., and Attikiouzel, Y.
Separation of signals with overlapping spectra using signal characterisation and hyperspace filtering.
IEEE Symposium on Adaptive systems for Signal Processing, Communications, and Control, Lake Louise, Alberta, Canada, pp. 327-332, 1-4th October 2000.

For separation of signals with overlapping spectra, classical linear filters fail to perform effectively. Nonlinear filters such as Volterra filters or Artificial Neural Networks (ANNs) can perform better but their implementations are often impractical due to their computational. In this paper an ANN based hyperspace signal model is used to separate signals with overlapping spectra. The computational complexity of the ANN is reduced significantly by a simple feature extraction utilising the unique temporal characteristics of the signals. The results show that difficult signal separation and filtering can be achieved efficiently by employing an ANN and effective feature extraction.

1999

Pathirana, P. N., and Zaknich, A.
Muti-path suppression for multiple swept carrier signals via time delay spectrometry concepts.
2nd International Conference on Information Communications and Signal Processing (ICICS'99), Mandarin, Singapore, CDROM 6 pages, 7-10th December 1999.

Multi-path suppression of signal acquisition systems is an important issue in a variety of applications including, sonar, underwater and mobile radio communications, medical ultrasound imaging and also in testing and design of acoustical systems. Time delay spectrometry immerged from classical loudspeaker testing systems where it was employed as a way of isolating a desired reflected or direct signal from other reflections in a reverberant environment. The traditional use of time delay spectrometry was mainly for characterising direct field audio system responses. Here, it is presented in a communication context with the introduction of the idea of non-overlapping bands of multiple linear frequency swept carriers each containing different amplitude modulated message signals. The data rate is increased over that achieved by a single linearly swept carrier over a given time interval by effectively folding this longer sweep signal into a shorter time interval. This folding creates the non-overlapping bands. Since the carriers for each band are all linearly swept signals they can each be recovered using the time delay spectrometry and gain the benefit of multi-path suppression.

Finch, B., Zaknich, A., and Cook, G.
Further development of time delay spectrometry.

138th Meeting of the Acoustical Society of America, pp. 2279, November 1999.

The accuracy of conventional time delay spectrometry (TDS) measurements are limited by constraints on signal parameters. These constraints are ultimately limited by the resolution with which the frequency response can be measured. Poletti, Cook, and others have shown that more exact measurements, without these constraints, are possible. These theories require that the system response to the full complex chirp be known. This is usually done by exciting the system with two separate orthogonal sweep signals. In this paper, this theory is developed in a manner similar to that done by Vanderkooy for the conventional TDS measurement, to show that the more exact system response can be deduced from a single linear sweep. The new development is supported by experimental results comparing the measured system response of a series resonant circuit against the conventional TDS results originally reported by Vanderkooy. This method provides practitioners more convenience in making exact TDS measurements, and allows for further exploration of the application of this technique in time varying environments.

Chao, M. T, Braunl, T., and Zaknich, A.
Visually-guided obstacle avoidance in office environments.
6th International Conference on Neural Information Processing-jointly with the 7th Australian and New Zealand International Conference on Intelligent Information Processing Systems, and the 5th New Zealand International Conference on Artificial Neural Networks and Expert Systems, and the 11th Australian Conference on Neural Networks, Perth, Australia, pp. 650-655, 16-20th November 1999.

This paper describes an indoor autonomous vision-based obstacle avoidance robot system. The vision part of the system converts forward looking greyscale camera images into edge images using Canny edge detection. Both edge image and sonar ranging information is used as stimuli by the behaviours that make up the reactive part of the system. These behaviours all run concurrently and they couple perception to actions to generate motor responses. A priority based subsumption coordinator selects the most appropriate response to direct the robot away from obstacles.

Minchin, G., and Zaknich, A.
A design for FPGA implementation of the probabilistic neural network.
6th International Conference on Neural Information Processing-jointly with the 7th Australian and New Zealand International Conference on Intelligent Information Processing Systems, and the 5th New Zealand International Conference on Artificial Neural Networks and Expert Systems, and the 11th Australian Conference on Neural Networks, Perth, Australia, pp. 556-559, 16-20th November 1999.

A design concept is introduced for the implementation of the Probabilistic Neural Network classifier using standard binary Field Programmable Gate Array logic. It is an efficient hardware design concept which substitutes fixed point binary valued vector components for real valued ones and uses a top hat spherical basis function in conjunction with a city block distance measure without significantly affecting classifier performance for some practical problems.

McGibney, S., and Zaknich, A.
Unsteady airflow classification by artificial neural networks.
6th International Conference on Neural Information Processing-jointly with the 7th Australian and New Zealand International Conference on Intelligent Information Processing Systems, and the 5th New Zealand International Conference on Artificial Neural Networks and Expert Systems, and the 11th Australian Conference on Neural Networks, Perth, Australia, pp. 1094-1099, 16-20th November 1999.

A Multi-Layer Perceptron classifier is applied to the classification of gas flow states. A number of suitable discriminant features are determined heuristically for the categorisation of gas flow states including; background, laminar flow, transition 1, transition 2 and turbulent flow. This technique can be used to develop an automatic real-time classifier for gas flow.

Zaknich, A., and Attikiouzel, Y.
The classification of sheep and goat feeding phases from acoustic signals of jaw sounds.
3rd International Conference on Knowledge-Based Intelligent Information Engineering Systems (KES), Adelaide, Australia, pp. 158-161, 31st August - 1st September 1999.

This paper describes and documents investigatory work for the detection and measurement of sheep rumination and mastication time periods from jaw sounds transmitted through the skull. The rumination and mastication time periods were determined by a neural network classifier using a combination of time and frequency domain features extracted from successive 10 second acoustic signal lengths. It is shown that spectral features contain most of the information required for good classification.

Rennick, G., Attikiouzel, Y., and Zaknich, A.
Machine grading and blemish detection in apples.
IEEE 5th International Symposium on Signal Processing and its Applications, Brisbane, Australia, pp. 567-570, 22-25th August 1999.

Five classifiers including the K-means, Fuzzy c-means, K-nearest neighbour, Multi-Layer Perceptron Neural Network and Probabilistic Neural Network classifiers are compared for application to colour grade classification and detection of bruising of Granny Smith apples. A number of suitable discriminate features are determined heuristically for the categorisation of four classes including: high grade fruit, high grade fruit with bruising or blemishes, off-grade fruit, and off-grade fruit with bruising or blemishes. Robust features based on intensity statistics are extracted from enhanced monochrome images produced by special transformation from original RGB images. The best of the five classifiers using the optimal feature set, is shown to outperform human graders viewing the same images.

Hanselmann, T., Zaknich, A., and Attikiouzel, Y.
Connection between BPTT and RTRL.
IEEE 3rd International Multiconference on Circuits, Systems, Communications and Computers (IMACS), Athens, Greece, pp. 190-193, 4-8th July 1999.

This paper shows the connection between the Backpropagation Through Time (BPTT) algorithm, its truncation form with truncation depth h, and the Recurrent Real Time Learning (RTRL) algorithm. The comparison is done by looking at a fully connected recurrent network, which is based on the same error function and calculations, using exact ordered derivatives. Two kinds of formulas, based on total ordered derivatives, for BPTT(h) are given and proven to be equivalent. Of the two formulae, the second one, can be interpreted by a target modification in the case of h ® Ą . Further, a combination of BPTT and RTRL and their implications as well as their interpretations and uses are outlined.

Zaknich, A.
Efficient kernel functions for the general regression and modified probabilistic neural networks.
IEEE International Joint Conference on Neural Networks (IJCNN), Washington, DC, USA, paper number 636, 10-16th July 1999.

Four spherical kernel functions and two associated distance measures for the General Regression and Modified Probabilistic Neural Networks are compared using four classification and four nonlinear filtering data sets. The standard Gaussian kernel is compared with three efficient functions - the tophat, triangle and a quadratic form kernel function. The standard Euclidean distance measure and more computationally efficient Hamming distance measure are also compared.. The work shows that the computationally efficient combination of quadratic kernel and Hamming distance measure can produce comparable results with the traditional Gaussian kernel with Euclidean distance measure.

Jan, T., and Zaknich, A.
An adjustable model for linear to nonlinear regression.
IEEE International Joint Conference on Neural Networks (IJCNN), Washington, DC, USA, paper number 635, 10-16th July 1999.

A basic limitation of all data-driven approximation methods is their inability to extrapolate accurately once the input is outside of the training data range. This paper examines the effectiveness and utility of combining a linear regression model with the General Regression Neural Network or Modified Probabilistic Neural Network for better linear extrapolation and function approximation. For a given set of training data, this combination provides a way of fine tuning the model by the adjustment of a single smoothing parameter.

Hanselmann, T., Zaknich, A., and Attikiouzel, Y.
Learning functions and their derivatives using Taylor series and neural networks.
IEEE International Joint Conference on Neural Networks (IJCNN), Washington, DC, USA, paper number 637, 10-16th July 1999.

This paper describes a design based on the Taylor series to approximate a function and its derivatives. After being trained, derivatives are obtained in a fast feed forward evaluation without the need for backpropagation or forward perturbation. The Taylor network is basically an implementation of the Taylor series of a function. However, instead of only having one expansion point, it uses a function of expansion points and takes account of the order of the Taylor series by biasing individual terms of the Taylor series. A simple experiment to learn a sinusoid and its first derivative.

Attikiouzel, Y., Tan, C. H., and Zaknich, A.
A neural network classifier for radio station scanning and analysis.
International Workshop on Mobile Communications, Crete, Greece, pp. 112-116, 24-26th June 1999.

Three neural network classifiers including the Probabilistic Neural Network, Kohonen's Self-Organising Map and a Multi-Layer Perceptron are compared for application to the classification of audio frequency signals from radio receivers. A number of suitable discriminate features are determined heuristically for the categorisation of six signal classes including: music, speech, silence, random noise, telemetry data signals and Morse code. This class selection can be used to develop an automatic radio station scanning system to select the class of program material desired by the listener and the others.

1998

Zaknich, A., and Baker, S.K.
A real-time system for the characterisation of sheep feeding phases from acoustic signals of jaw sounds.
Australian Journal of Intelligent Information Processing Systems (AJIIPS), Vol. 5, No. 2, pp. 103-110, Winter 1998.

This paper describes a four-channel real-time system for the detection and measurement of sheep rumination and mastication time periods by the analysis of jaw sounds transmitted through the skull. The system is implemented using an 80486 personal computer, a proprietary data acquisition card (PC-126) and a custom made variable gain preamplifier and bandpass filter module. Chewing sounds are transduced and transmitted to the system using radio microphones attached to the top of the sheep heads. The system's main functions are to detect and estimate rumination and mastication time periods, to estimate the number of chews during the rumination and mastication periods, and to provide estimates of the number of boli in the rumination sequences and the number of chews per bolus. The individual chews are identified using a special energy threshold detector. The rumination and mastication time periods are determined by neural network classifier using a combination of time and frequency domain features extracted from successive 10 second acoustic signal blocks.

Zaknich, A., and Attikiouzel, Y.
A comparison of template matching with neural network approaches in the recognition of numeric characters hand-stamped in aluminium.
IEEE International Workshop on Intelligent Signal Processing and Communications Systems, Melbourne, Australia, pp. 98-102, 4-6th November 1998.

This paper investigates a variety methods of feature extraction for the recognition of numeric digits stamped in aluminium from very poor grey-scale images. Comparisons are made between the traditional template matching method and other approaches incorporating artificial neural networks. The best methods based on artificial neural networks outperformed the template matching method as expected. It was evident that the better the preprocessing and feature extraction the better the neural network performance.

Zaknich, A., and Attikiouzel, Y.
An unsupervised clustering algorithm for the modified probabilistic neural networks.
IEEE International Workshop on Intelligent Signal Processing and Communications Systems, Melbourne, Australia, pp. 319-322, 4-6th November 1998.

The Modified Probabilistic Neural Network is a general regression method which has fundamental similarities to Specht's General Regression Neural Network. It is able to achieve performances better than or equal to the General Regression Neural Network with a much smaller network size. However, the Modified Probabilistic Neural Network size is still large compared to other types of neural networks. Consequently, methods are required to reduce the size further without sacrificing the network's simplicity. A simple one-pass unsupervised clustering algorithm is applied to the Modified Probabilistic which significantly reduces the network size with minimal performance loss.

Zaknich, A.
Introduction to the modified probabilistic neural network for general signal processing applications.
IEEE Transactions on Signal Processing, Vol. 46, No. 7, pp. 1980-1990, July 1998.

This paper introduces a practical and easy to understand network for signal processing called the Modified Probabilistic Neural Network. It begins with a short introduction to the application of artificial neural networks to signal processing followed by a background and review of the Modified Probabilistic Neural Network theory. The Modified Probabilistic Neural Network is a regression technique similar to Specht's General Regression Neural Network which is based on a single radial basis function kernel whose bandwidth is related to the noise statistics. It has advantages in application to time and spatial series signal processing problems because it is constructed directly and simply from the training signal waveform characteristics or features. An illustrative example involving noisy Doppler shifted swept frequency sonar signal detection compares the effectiveness of the first and second order Volterra, Multi- Layer Perceptron Neural Network, Radial Basis Function Neural Network, General Regression Neural Network and Modified Probabilistic Neural network filters, demonstrating some features of the Modified Probabilistic Neural Network for practical design.

Cook, G., and Zaknich, A.
Chirp sounding underwater acoustic channels.
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seattle, Washington, USA, Vol. IV, pp. 2521-2524, 12-15th May 1998.

Characterisation of the shallow water acoustic communications channel involve the analysis of sounding data. Chirp signals have many properties which make them an attractive choice for channel sounding. They are easily generated and channel responses can be processed in the time or frequency domain for channel estimation. In the rapidly varying shallow water environment time domain techniques are most appropriate. In this case weighting windows can be used to reduce clutter in the estimate. A channel sounding experiment is described which employs very simple hardware to generate and record chirp responses for off line processing.

Zaknich, A.
An efficient modified probabilistic neural network hardware implementation for zero crossing thresholded binary signals.
IEEE International Joint Conference on Neural Networks (IJCNN), Alaska, USA, pp. 256-261. 256-261, 4-9th May 1998.

An efficient form of the modified probabilistic neural network is developed for the detection of Doppler shifted zero crossing thresholded binary chirp signals and other similar signals. The normal modified probabilistic neural network algorithm is based on a Gaussian radial basis function and the Euclidean distance measure requiring complex arithmetic operations. By using a simple tophat radial basis function and hamming distance measure in conjunction with binary signals it is possible to simplify the repetitive arithmetic operations to produce a more efficient form of the modified probabilistic neural network. This new form is can produce more accurate correlator detector outputs than a multiple correlator detector system for moderate to high signal to noise ratios.

Zaknich, A.
A nonlinear correlator detector for Doppler shifted chirp signals.
Australian Journal of Intelligent Information Processing Systems (AJIIPS), Vol. 4, No. 2, pp. 142-150, Winter 1997.

The modified probabilistic neural network is applied to the problem of detecting Doppler shifted chirp signals. It is shown to be superior in performance to a single linear correlator, multiple parallel linear correlators, a quadratic filter and multi-layer perceptron and radial basis function neural network filters. Not only is the modified probabilistic neural network more accurate in respect to amplitude and time but it performs much better in the presence of noise. It has the desirable feature that it may be trained with noiseless data and still perform very well as noise is added.

1997

Zaknich, A., and deSilva, C.J.S.
Adaptive learning schemes for the modified probabilistic neural network.
IEEE Third International Conference on Algorithms and Architectures for Parallel Processing, Melbourne, Australia, pp. 597-610, 8-12th December 1997.

The Modified Probabilistic Neural Network was initially derived from Specht's Probabilistic Neural Network classifier and developed for nonlinear time series analysis. Essentially, the Modified Probabilistic Neural Network could be described as a vector quantised reduced form of Specht's General Regression Neural Network. It is typically trained with a known set of representative data pairs. This is quite satisfactory for stationary data statistics but for the nonstationary case it is necessary to be able to adapt the network during operation. This paper describes adaptive learning schemes for the Modified Probabilistic Neural Network for both stationary and nonstationary data statistics. A nonlinear control problem is used to illustrate and compare the Modified Probabilistic Neural Network's learning ability with respect to the General Regression and Radial Basis Function Neural Networks. The issues of data searching and sorting related to network building are also discussed.

Zaknich, A., and Attikiouzel, Y.
General regression techniques based on spherical kernel functions for intelligent processing.
Chapter in Brain-like computing and intelligent information systems, Editors - S. I. Amari and N. Kasabov, Springer-Verlag, Singapore, pp. 189-211, 1997.

This chapter reviews and compares a set of nonparametric spherical kernel based general regression techniques having similarities to the statistical work of Nadaraya and Watson. The set includes Specht's General Regression Neural Network, the Modified Probabilistic Neural Network introduced by Zaknich et al, the method of Moody and Darken and the method of Radial Basis Functions developed by Powell. All these methods are based on Bayesian estimation theory, can be realised by a parallel artificial neural network architecture and are trained by supervision. It is shown that these methods have some advantages over other neural networks in relation to brain-like intelligent processing applications. The Modified Probabilistic Neural Network has special advantages when used to solve time and spatial series processing problems because it is constructed directly and simply from the training series waveform characteristics. An illustrative intelligent control example compares the effectiveness and advantages of the Modified Probabilistic Neural Network over the General Regression and the Radial Basis Function Neural Networks which in turn all have advantages over the Multi-Layer Perceptron Neural Network and other similar types.

Zaknich, A., and Attikiouzel, Y.
A fast adaptive neural network system for intelligent control.
IEEE International Conference on Systems, Man and Cybernetics, Vol. II, Orlando, Florida, pp. 1023-1027, 12-15th October 1997.

An intelligent control system needs to adapt to new dynamics very quickly but also retain knowledge of past dynamics to be able to act effectively and quickly for repeat occurrences. One solution is to model the system with two neural networks in parallel whereby one network is trained a priori with a wide range of historical dynamics while the second one, is allowed to adapt itself to make up the differences between the first model and the real-time dynamics. Within this scheme, as the second network is called to adapt itself, the first one can be progressively trained to learn the new dynamics without adversely affecting the old training. A strategy of this type can be achieved very effectively using the Modified Probabilistic Neural Network because it is constructed with local radial kernel functions and its adaptation mechanism is computationally simple and very fast. This is demonstrated using a complex nonlinear system whose characteristics suddenly change after initial training and then switch back to the original characteristics. Comparisons are made with other networks to show important advantages of the Modified Probabilistic Neural Network.

Zaknich, A.
Characterisation of aluminium hydroxide particles from the Bayer process using neural network and Bayesian classifiers.
IEEE Transactions on Neural Networks, Vol. 8, No. 4, pp. 919-931, July 1997.

An automatic process of isolating, identifying and characterising individual aluminium hydroxide particles from the Bayer process in scanning electron microscope grey-scale images of samples is described. It has been possible to effectively automate this process, previously performed by human experts, by various image processing algorithms and judicious use of neural network and Bayesian classifiers. As aluminium hydroxide particles can be quite amorphous and very different from particle to particle this necessitated the use of neural network and Bayesian classifiers in various stages to resolve complex nonlinear decisions and anomalies. The process has been achieved in two stages; the isolation and identification of individual particles in the images and then the classification of each particle. The isolation process is shown to isolate and then correctly identify 96.9% of the objects as complete and single particles after a 15.5% rejection of questionable objects. The sample set had a possible 2455 particles taken from 384 256x256 pixel images. Of the 15.5% initially rejected, 14.2% were correctly rejected and only 1.3% represented acceptable isolated particles incorrectly rejected. The rejection of unsatisfactory objects is necessary to compensate for the errors in the isolation process. With a 0% rejection the accuracy drops to 91.8% which represents the accuracy of the isolation process alone. After the particles are isolated they are classified according to five general property types related to morphology and texture. These five were determined by expert opinion and they are named shape, single crystal protrusions, texture, crystal size, agglomeration. The 2455 particle samples were preclassified in each of these five property types by a human expert and the data was used to train the five classifiers to embody the expert knowledge. The system was designed to be used as a research tool to determine and study relationships between particle properties and plant parameters in the production of smelting grade alumina by the Bayer process.

Sia, S., and Zaknich, A.
Neural network edge detectors for separation of particles in 2-D grey-scale images.
13th International Conference on Digital Signal Processing (DSP-97), Santorini, Greece, pp. 1141-1144, 2nd-4th July 1997.

A method is developed for edge detection in grey-scale images of particles using artificial neural network (ANN) classifiers. The edge detection is used to separate the individual particles in the image especially where the particles are touching each other. Once the particles are separated individual measurements can be made to compile accurate size-distribution information. Images of eighty calibrated gravel stones are used to test the method. Gravel stones are adopted as abstracts for the general class of irregularly shaped particles. Methods are developed to reduce the number of feature vectors and features in order to speed up the edge detection process without a loss in performance. Comparisons are made between Probabilistic Neural Network, Gaussian Mixture Model Bayesian and Multi-layer Perceptron classifiers for edge detection.

Zaknich, A., and Attikiouzel, Y.
A modified probabilistic neural network signal processor for nonlinear signals.
13th International Conference on Digital Signal Processing (DSP-97), Santorini, Greece, pp. 291-294, 2nd-4th July 1997.

This paper introduces a practical and very effective network for nonlinear signal processing called the Modified Probabilistic Neural network. It is a regression technique which uses a single radial basis function kernel whose bandwidth is related to the noise statistics. It has special advantages in application to time and spatial series signal processing problems because it is constructed directly and simply from the training signal waveform characteristics or features.

Zaknich, A.
A vector quantisation reduction method for the probabilistic neural network.
IEEE International Conference on Neural Networks (ICNN), Vol II, Houston, Texas, USA, pp. 1117-1120, 9-12th June 1997.

This paper introduces a vector quantisation method to reduce the Probabilistic Neural Network classifier size. It has been derived from the Modified Probabilistic Neural Network which was developed as a general regression technique but can also be used for classification. It is a very practical and easy to implement method requiring a very low level of computation. The method is described and demonstrated using 4 different sets of classification data.

Pathirana, P., and Zaknich, A.
Identification of surfaces by acoustics using time delay spectrometry and neural network classifiers.
IEEE International Conference on Neural Networks (ICNN), Vol I, Houston, Texas, USA, pp. 31-36, 9-12th June 1997.

The identification of surfaces using incident sound waves is associated with a variety of different applications including, sonar, seabed scanning and medical ultrasound imaging. The biologically innocuous nature, applicability, and simplicity involved in generation and measurement, makes sound inherently a more attractive agent for most applications. Time delay spectrometry can be employed as away of isolating a desired reflected signal from other reflections dramatically increasing the signal to noise ratio of the receiver of a neural network based classification system with the analysis of its performance will be introduced in this paper as a successful implementation of the proposed methodology.

1996

Cook, G., and Zaknich, A.
Theory and implementation of extensions to time delay spectrometry.
Audio Engineering 6th Regional Conference, Melbourne, VIC, Australia, preprint 4299, 10-12 September 1996.

Two new results are added to the theory of Time Delay Spectrometry (TDS). The first is the discovery of a fundamental property of the complex linear frequency sweep. This property allows for perfect reconstruction of the desired spectrum. The second result is a refinement to the definition of space equivalent bandwidth. Implementation issues are addressed and a computer simulation is described which exploits these new results. Precise expressions for error are developed.

McKinnon, M., and Zaknich, A.
Scratch noise filtering using neural networks.
Audio Engineering 6th Regional Conference, Melbourne, VIC, Australia, preprint 4304, 10-12 September 1996.

A neural network based method of scratch noise removal from phonograph recordings is described. A design for scratch detection and filtering is presented along with an examination of the pre-processing and feature extraction incorporated. It is able to remove scratch noise with a minimal loss of high frequency information and is superior to classical approaches.

Cook, G., and Zaknich, A.
Extensions to the theory of time delay spectrometry.
IEEE Tencon (Digital Signal Processing Applications), Perth, WA, Australia, pp. 763-768, 27-29 November 1996.

Time Delay Spectrometry (TDS) is a technique used to measure both frequency and time responses of acoustic and electrical systems. TDS theory is presented in both intuitive and mathematical terms. The theoretical limitations of current TDS technology are described. A new extension to TDS theory is introduced which has the potential to significantly improve TDS measurements. Implementation issues are addressed and expressions for the system error are derived. The results of computer simulation are presented.

Zaknich, A., and Attikiouzel, Y.
Modified probabilistic neural network hardware implementation schemes.
IEEE Tencon (Digital Signal Processing Applications), Perth, WA, Australia, pp. 167-172, 27-29 November 1996.

The modified probabilistic neural network for nonlinear time series analysis was developed and introduced in 1991. It effectively represents a simple family of clustering met hods for reducing the size of Specht's general regression neural network and retaining all its benefits. Three hardware implementation schemes for the most basic form of the modified probabilistic neural network are described. The first is an optoelectronic implementation and the other two are Very Large Scale Integration designs: a virtual implementation and a fully parallel implementation.

1995

Zaknich, A., and Attikiouzel, Y.
Application of artificial neural networks to nonlinear signal processing,
Computational Intelligence: A Dynamic System Perspective.
IEEE Press, pp. 292-311, November 1995.

A brief review of the application of artificial neural networks to nonlinear signal processing is presented followed by Poggio's identification of parametric approximation theory as a classical framework for this type of problem. To help develop the idea of nonlinear discrete time series signal processing a general signal processing model is defined for reference. The static, dynamic and adaptive ANN supervised learning models are offered as solutions to the nonlinear signal processing problem. A number of practical issues, including feature extraction, related to the application of ANNs to nonlinear signal processing are discussed and developed. Some nonlinear signal filtering applications are cited and an illustrative example involving the removal of impulse plus wide band random noise from a speech signal is developed. In this example the effectiveness of linear finite impulse response and nonlinear Quadratic, Third Order Volterra, Backpropagation (Multilayer Perceptron), General Regression Neural Network and Modified Probabilistic Neural Network filters are compared. The paper concludes with a discussion of the comparative features of the main ANN types and some special advantages of the Modified Probabilistic Neural Network filter over the others, especially for nonlinear time series signals.

Zaknich, A., and Attikiouzel, Y.
The modified probabilistic neural network as a nonlinear correlator detector.
IEEE International Conference on Neural Networks Proceedings, Vol. 1, Perth, WA, Australia, pp. 309-313, November 1995.

A Nonlinear correlator detector for the detection of a signal class with some intra class variance is developed using the Modified Probabilistic Neural Network and the General Regression Neural Network. An application, involving the detection of regular tone bursts transmitted over a poor and noisy radio channel subjected to fading, random noise and impulse noise effects, is used to show the effectiveness of the method as compared to a linear correlator.

Zaknich, A., and Attikiouzel, Y.
Detection of sodium oxalate needles in optical images using neural network classifiers.
IEEE International Conference on Neural Networks Proceedings, Vol. 4, Perth, WA, Australia, pp. 1699-1702, November 1995.

A description is given of a PC based system for the automatic detection, counting and sizing of sodium oxalate needles in optical microscope images predominated by a background of hydrate particles. The system is primarily based on a neural network classifier which is fed by a feature vector derived from grayscale dynamically thresholded binary images. A Backpropagation neural network (BPN) was adopted for technical reasons, but any of the other neural network classifiers could have been used. Comparative results are given for the Backpropagation, Probabilistic (PNN), General Regression (GRNN) neural networks and a Gaussian model, which show the utility and validity of the neural network approach.

Zaknich, A., and Attikiouzel, Y.
Time series characterisation schemes for the modified probabilistic neural network.
Australian Journal of Intelligent Information Processing Systems, Vol. 2, No. 2, pp. 1-11, June 1995.

A brief review of the basic Modified Probabilistic Neural Network algorithm is presented, followed by two new extensions for applications to time series analysis, especially to nonlinear signal filtering. Both methods rely on a systematic selection of radial basis centres, based on quantised time series waveform characterisation schemes. The methods are similar to Specht's General Regression Neural Network, and the method proposed by Moody and Darken. The main differences are that the new methods offer very simple and systematic network reduction mechanisms, and very fast training times without the need for complex computations. These makes them very suitable for hardware implementation using present and foreseeable technology.

Zaknich, A., and Attikiouzel, Y.
Application of the modified probabilistic neural network to the enhancement of noisy short wave radio time and Morse code signals.
Australian Journal of Intelligent Information Processing Systems, Vol. 2, No. 3, pp. 9- 14, September 1995.

This paper investigates nonlinear and linear techniques for the enhancement of short wave radio time and Morse code signals, corrupted by typical channel effects such as fading, random noise and impulse noise. Comparative results are given for solutions based on the Modified Probabilistic Neural Network, General Regression, Backpropagation neural networks, and first, second and third order Volterra Filters, which demonstrate the advantages of the neural network approach to this type of problem.

1994

Zaknich, A., and Cornell, J.
Application of neural networks to image analysis in the alumina industry.

Image Analysis Seminar, Fremantle, WA, Australia, pp. 27-37, September 1994.

Alcoa of Australia Limited uses the Bayer process to refine aluminium. This involves the digestion of gibbsite from bauxite in caustic soda and subsequent re-precipitation of the gibbsite in a continuous precipitation circuit. The gibbsite is calcinated and shipped to smelters. Two image analysis systems based on artificial neural networks are described. They are a hydrate particle characterisation system and a oxalate needle counting system The first system automatically identifies hydrate particles in scanning electron microscope images and characterises them with respect to shape, texture, crystallite size and degree of agglomeration. The second system automatically discriminates oxalate crystals from hydrate particles in specially prepared optical microscope images and counts and sizes the oxalate needles.

Klein, L., Baker, S. K., Purser, D. B., Zaknich, A., and Bray, A. C.
Telemetry to monitor sounds of chews during eating and rumination by grazing sheep.
Australian Society of Animal Production Conference, Perth, Western Australia, Vol. 20, p. 423, 5th July 1994.

The objective was to monitor eating and ruminating behaviour of grazing sheep, using computer software that would analyse in real-time mode the audio input from microphones on the sheep's heads.

1993

Zaknich, A., and Attikiouzel, Y.
Automatic optimisation of the modified probabilistic neural network for pattern recognition and time series analysis.
First Australian and New Zealand Conference on Intelligent Information Systems, Perth, Western Australia, pp. 152-156, 1-3rd December 1993.

The Modified Probabilistic Neural Network (MPNN) is based on Specht's Probabilistic Neural Network (PNN) and was developed primarily for time series analysis. It can also be used for pattern classification as can the PNN but it has the added benefit that it can be automatically trained using a convergent optimisation algorithm. The algorithm automatically finds the optimum smoothing factor, given a set of training and testing data, by minimizing the mean squares error between network outputs and desired responses using recurrent parabolic curve fitting. The algorithm is described and its effectiveness shown on time series analysis and pattern recognition problems. A short review of the PNN and MPNN is also given.

1992

Attikiouzel, Y., and Zaknich, A.
Applications of the probabilistic neural network.

Canadian Conference on Electrical and Computer Engineering (CCECE), Ontario, Canada, pp. TM6.11.1-TM6.11.4, 13-16th September 1992.

Specht introduced a one pass learning algorithm called the Probabilistic Neural Network (PNN) for classification, mapping and associative memory. This paper describes how the standard PNN and a modified version was used to develop a number of successful applications for industry and agriculture. Brief introductions to the PNN and modified PNN algorithms are given as well as summaries of other applications reported in the literature.

Zaknich, A., and Attikiouzel, Y.
A probabilistic neural network edge detector for 2 dimensional gray scale images.
Australian Neural Network Conference, pp. 274-277, February 1992.

A Probabilistic Neural Network classifier is used to design an edge detection algorithm for two dimensional gray scale images of rocks. It is shown how a very sparse training set, made up of a small number of example vectors, is sufficient to generalise the edge detection of each rock. The data is also applied to the Backpropagation neural network classifier for comparison. Many industrial image processing problems, for which it is difficult to develop a priori processing rules can be solved conveniently in a similar way.