Anthony Zaknich

Continuous-Time Adaptive Critics

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 in well with plant descriptions given by differential equations and that any standard integration routine with adaptive step-size does an adaptive sampling for free. A second-order actor adaptation using Newtons 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 actor 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

Perceptual features for automatic speech recognition in noisy environments

The performances of two perceptual properties of the peripheral auditory system, synaptic adaptation and two-tone suppression, are compared for automatic speech recognition (ASR) in an additive noise environment. A simple method of synaptic adaptation as determined by psychoacoustic observations was implemented with temporal processing of speech utilizing a zero-crossing auditory model as a pre-processing front end. The concept is similar to RASTA processing, but instead of bandpass filters, a high-pass infinite impulse response (IIR) filter is used. It is shown that rapid synaptic adaptation may be implemented by temporal processing using the zero-crossing algorithm, not otherwise implementable in the spectral domain implementation. The two-tone suppression was implemented in the zero-crossing auditory model using a companding strategy. Recognition performances with the two perceptual features were evaluated on isolated digits (TIDIGITS) corpus using continuous density HMM recognizer in white, factory, babble and Volvo noise. It is observed that synaptic adaptation performs better in stationary white Gaussian noise. In presence of non-stationary non-Gaussian noise, however, no improvements or a degradation is observed. Moreover, a reciprocal effect is observed with two-tone suppression, with better performance in non-Gaussian real-world noise and degradation in stationary white Gaussian noise.

Dynamic population variation in genetic programming

Three innovations are proposed for dynamically varying the population size during the run of the genetic programming (GP) system. These are related to what is called Dynamic Population Variation (DPV), where the size of the population is dynamically varied using a heuristic feedback mechanism during the execution of the GP with the aim of reducing the computational effort compared with Standard Genetic Programming (SGP). Firstly, previously developed population variation pivot functions are controlled by four newly proposed characteristic measures. Secondly, a new gradient based pivot function is added to this dynamic population variation method in conjunction with the four proposed measures. Thirdly, a formula for population variations that is independent of special constants is introduced and evaluated. The efficacy of these innovations is examined using a comprehensive range of standard representative problems. It is shown that the new ideas do have the capacity to provide solutions at a lower computational cost compared with standard genetic programming and previously reported algorithms such as the plague operator and the static population variation schemes previously introduced by the authors.

A modified probabilistic neural network (PNN) for nonlinear time series analysis

A modified PNN (probabilistic neural network) is proposed that can be used for nonlinear time series analysis without loss of the advantages offered by D.F. Spechts PNN architecture (1988, 1990). It is shown how the Gaussian radial basis function, expressed as a Parzen probability density function estimator, can be used to estimate and implement nonlinear mappings, applied to time series data. The performance of this modified PNN is demonstrated by showing its effectiveness in smoothing a sinusoidal signal which has been compressed in amplitude and then corrupted with wideband non-Gaussian noise. The network is also compared with the multipass learning backpropagation network and the relative merits of the proposed modified PNN are discussed.<<ETX>>

A vector quantisation reduction method for the probabilistic neural network

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.

Machine grading and blemish detection in apples

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.

A practical sub-space adaptive filter

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 MPNNs special structure allows it to be easily used to model a process by appropriately weighting piecewise linear models associated with each of the networks 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.

A framework of adaptive T-S type Rough-Fuzzy Inference Systems (ARFIS)

The rough-fuzzy hybridization scheme has become of research interest in a variety of areas over the past decade. The present paper proposes a general framework for adaptive T-S type rough-fuzzy inference systems (ARFIS) for many practical applications. Rough set theory is utilized to reduce the number of attributes and 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 the rough set approach, respectively. The generated T-S type rough-fuzzy inference system is then adjusted by the least squares fit and the conjugate gradient descent algorithm towards better performance with a validity checking for the generated minimal set of rules. The proposed framework of 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 framework of ARFIS is compared with other existing approaches in a variety of application areas and shown to be very competitive.

A Temporal Auditory Model with Adaptation for Automatic Speech Recognition

Rapid and short-term adaptation are dynamic 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 recognizer 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.

An adjustable model for linear to nonlinear regression

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 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 as well as providing linear extrapolation.