New Publications are available for Neural nets
http://dl-live.theiet.org
New Publications are available now online for this publication.
Please follow the links to view the publication.On paradigm shift in computing using soft computing
http://dl-live.theiet.org/content/conferences/10.1049/ic.2009.0116
Conventional technique of computing, which is centrally based on mathematical approaches to problem solving, is considered as hard computing technique. This requires precisely stated analytical model. Many analytical models are available for ideal cases. However, real world problems exist in a nonideal environment. Soft computing techniques, which have drawn their inherent characteristics from biological systems, present effective solutions to these problems. Soft computing refers to computational techniques that include Fuzzy Logic, Artificial Neural Networks and Genetic Algorithms. All of these techniques find their deep roots in artificial intelligence (AI). The paper highlights important features of these computational techniques. Each of these techniques, in their domain has provided very effective solutions to wide range of problems. Attempts have been made to integrate features of these techniques. The paper also focuses on systems of such hybrid soft computing techniques in brief and it is argued that these systems must be autonomous, robust and adaptive in order to be intelligent and stable. Finally, this paper clearly brings out superiority of soft computing techniques over the conventional. (5 pages)Simulating and predicting blood glucose levels for improved diabetes healthcare
http://dl-live.theiet.org/content/conferences/10.1049/cp_20080433
Expert management of blood glucose levels (BGLs) can be difficult to achieve for many people with diabetes, since BGLs are affected in a very complex and non-linear manner by carbohydrate intake (diet), medication (tablets or insulin) and exercise. This paper discusses some of the expert management diabetes software systems that have been developed in the last decade which can accommodate the day- to-day complexities of BGL control, through a BGL prediction / therapy optimization strategy. In particular, the popular AIDA educational diabetes simulator is reviewed, together with a novel BGL simulation system based on artificial neural networks. (4 pages)An application of neural networks to adaptive play out delay in VoIP
http://dl-live.theiet.org/content/conferences/10.1049/cp_20070686
The statistical nature of data traffic and the dynamic routing techniques employed in IP networks results in a varying network delay (jitter) experienced by the individual IP packets which form a VoIP flow. As a result voice packets generated at successive and periodic intervals at a source will typically be buffered at the receiver prior to playback in order to smooth out the jitter. However, the additional delay introduced by the playout buffer degrades the quality of service. Thus, the ability to forecast the jitter is an integral part of selecting an appropriate buffer size. This paper compares several neural network based models for adaptive playout buffer selection and in particular a novel combined wavelet transform/neural network approach is proposed. The effectiveness of these algorithms is evaluated using recorded VoIP traces by comparing the buffering delay and the packet loss ratios for each technique. In addition, an output speech signal is reconstructed based on the packet loss information for each algorithm and the perceptual quality of the speech is then estimated using the PESQ MOS algorithm. Simulation results indicate that proposed Haar-wavelets-packet MLP and statistical-model MLP adaptive scheduling schemes offer superior performance.Perceptually motivated output-based speech quality assessment using neural networks
http://dl-live.theiet.org/content/conferences/10.1049/cp_20040519
A new output-based method for the prediction of subjective speech quality is proposed and its performance evaluated. The method is based on measuring perceptually motivated objective auditory distances between the voiced parts of the speech signal whose quality is to be evaluated to appropriately matching reference vectors extracted from a preformulated codebook. The codebook is formed by optimally clustering a large number of perceptually-based parametric vectors extracted from a database of clean speech signals. The auditory distance measures are then mapped into equivalent subjective scores, represented by the mean opinion scores (MOS), using regression. The required clustering and matching processes are achieved by using an efficient neural network based data mining technique known as the self-organising map. Perceptual speaker-independent parametric representation of the speech is achieved by using a linear prediction (PLP) model and bark spectrum analysis. The reported evaluation results show that the proposed system is robust against speaker, utterance and distortion variations.Determination of machine parameters for internal permanent magnet synchronous motors
http://dl-live.theiet.org/content/conferences/10.1049/cp_20040392
Experimental methods used recently by many researchers for the determination of PMSM parameters have been reviewed. Analysis and experimentations show many shortcomings and inaccuracies involved in those methods. Two novel techniques for better results through the application of linear regression and neural network are presented, and the results from these methods are also included. In addition, computational design values obtained by finite element computations are also compared to the experimental values. (6 pages)Intelligent control schemes for a static compensator connected to a power network
http://dl-live.theiet.org/content/conferences/10.1049/cp_20040355
Two intelligent controllers are designed for a static compensator (STATCOM) connected to a single machine infinite bus power system (SMIB): a novel nonlinear adaptive controller using artificial neural networks based on the indirect adaptive control technique and a Takagi-Sugeno type fuzzy controller. Both schemes provide nonlinear adaptive control with better performance compared to the conventional PI controllers. Simulation results are presented to compare the performances of these controllers with that of the conventional PI controllers. (6 pages)A neural system for automated CCTV surveillance
http://dl-live.theiet.org/content/conferences/10.1049/ic_20030040
We present the Owens Tracker, a complete hybrid neural prefiltering system for tracking pedestrians in a car park, and raising operator attention when unusual activity (defined by pedestrian trajectories) is detected. The system uses a combination of background differencing to detect moving objects, a specialized multiple tracking algorithm to maintain object records, and a two-part neural novelty detection module to detect novel trajectories defined both by short-term and long-term characteristics. The system was developed using data from a commercial industrial park. Experiments demonstrate that it is very robust; it detects and discounts the movement of cars, and can handle problems such as car drop-offs, noise, shadows and reflections. It reliably detected virtually all unusual pedestrian trajectories, but raised a number of false positive alarms. Approximately 50% of these were due to tracking failures, indicating that some improvement in that component of the system would be useful; the others are tolerable, given the system's target deployment as an attention-focussing filter. However, even in its current form, the system has the potential to reduce dramatically the burden of user monitoring, as only about 20 minutes footage from ten hours was identified as needing operator attention - a 30-fold decrease in effort. (5 pages)IP-distributed computer-aided video-surveillance system
http://dl-live.theiet.org/content/conferences/10.1049/ic_20030044
We present a generic, flexible and robust approach for an intelligent real-time video-surveillance system. The proposed system is a multi-camera platform that is able to handle different standards of video inputs (composite, IP, IEEE1394). The system implementation is distributed over a scalable computer cluster based on Linux and IP. Data flows are transmitted between the different modules using multicast technology, video flows are compressed with the MPEG-4 standard and flow control is realized through a TCP-based command network (e.g. for bandwidth occupation control). The design of the architecture is optimized to display, compress, store and playback data and video flows in an efficient way. The platform also integrates advanced video analysis tools, such as motion detection, segmentation, tracking and neural networks modules. The goal of these advanced tools is to provide help to operators by detecting events of interest in visual scenes and storing them with appropriate descriptions. This indexation process allows one to browse rapidly through huge amounts of stored surveillance data and play back only interesting sequences. We report here some preliminary results and we show the potential use of such a flexible system in a third generation video surveillance system. We illustrate the interest of the system in a real case study, which is the surveillance of a reception desk. (5 pages)Expert and machine discrimination of marine flora: a comparison of recognition accuracy of field-collected phytoplankton
http://dl-live.theiet.org/content/conferences/10.1049/cp_20030516
The categorisation of labelling of phytoplankton specimens is carried out manually using microscopes by marine ecologists and taxonomists. Research to automate the task has been on going for many years. Although many systems have been shown to work in small-scale laboratory conditions with cultured populations, few have succeeded when applied to field collected specimens. The reasons are diverse, but are principally due to severely degraded performance of the chosen processing algorithms in the presence of noise and natural morphological variability of the organisms. The application of statistical and neural network pattern learning methods have allowed progress to be made in this difficult area. The machine learning system DiCANN was trained on 128 of the 310 image data set and tested on the 182 samples. This study has highlighted the difficulties facing human ecologists and has shown that automation methods can perform as well as humans on complex categorisations.A comparison of neural network and polynomial models for the approximation of non-linear and anisotropic ferromagnetic materials
http://dl-live.theiet.org/content/conferences/10.1049/ic_20020144
Polynomials fail to give suitable approximations for strongly nonlinear functional mappings. In that case, neural networks can preferably be used. Over the past decade, their popularity steadily increased within various engineering disciplines. Here, it is shown that neural networks must not always be preferred over traditional polynomials. When modeling typical nonlinear and anisotropic magnetic properties for e.g. finite element simulations, both approximations are fairly competitive. (2 pages)GA-RBF neural network based maximum power point tracking for grid-connected photovoltaic systems
http://dl-live.theiet.org/content/conferences/10.1049/cp_20020083
This paper presents a novel GA-RBFNN (genetic algorithm trained radial basis function neural network)-based model to carry out the maximum power point tracking (MPPT) for grid-connected photovoltaic (PV) power generation control systems. The hidden layer of the neural network is self-organised by the GA-based RBF growing algorithm. The trained GA-RBFNN-based MPP model is then employed to predict the maximum power points of a PV array using measured environmental data. The simulation results are compared with the conventional P&O method, and the current/voltage waveforms of the PV panel are presented and discussed.Neural network based techniques for acceleration sensor signal processing: linearisation and diagnosis
http://dl-live.theiet.org/content/conferences/10.1049/ic_20010108
(1 page)Multi-sensor integration and decision level fusion
http://dl-live.theiet.org/content/conferences/10.1049/ic_20010101
Concerns sensor fusion for classification. The many techniques available for fusion at the decision or probability-of-decision level include powerful machine learning and neural network tools. These can be used to design the fusion system that takes the outputs of sensor specific experts and combines them to reach a consensus decision. However, a considerable body of experience suggests that often the best designs are not achieved using the most general tools. These can be cumbersome, require a lot of training data, and may result in over-training because of their capacity to over fit the data. In this paper we argue that by gaining better understanding of the issues involved in the fusion process we should be able to address them individually. The union of these design steps is reflected in an overall architecture which will be the focus of our discussion. We adopt the Bayesian viewpoint and show how this leads to classifier output moderation to compensate for sampling problems. We then discuss how the moderated outputs should be combined to reflect the prior distribution of the models underlying the classifier designs. The final stage of fusion combines the complementary measurement information that may be available to different experts. This process is embodied in an overall architecture which shows why the fusion of raw expert outputs is a nonlinear function and how this function can be realised as a sequence of relatively simple processes. (6 pages)Neural network applications in the water industry
http://dl-live.theiet.org/content/conferences/10.1049/ic_20010111
The operation of water treatment plants is significantly different from most manufacturing industrial operations because raw water sources are often subject to natural perturbations like flood and drought, both of which significantly affect the characteristics of the abstracted water. More recently, improved sensor technology has enabled the successful regulation of variables such as pH and chlorine residual. Without a precise knowledge of the characteristics of the material to be removed, most chemical dosage requirements for primary water treatment are determined from laboratory measurements which are conducted (usually) not less than once a day. This paper gives a brief explanation of water treatment plant operation, and outlines a number of case studies where system knowledge contained within artificial neural networks has been used to provide solutions to operational problems within the water industry. (6 pages)A new demixer scheme for blind source separation using general neural network model
http://dl-live.theiet.org/content/conferences/10.1049/cp_20010077
There has been a surge of interest in blind source separation (BSS) because of its potential applications in several areas of engineering and science such as wireless systems. We propose a new neural network demixing scheme using a general neural network structure for the BSS problem for the instantaneous mixtures. It is shown that the existing feedforward (FF) and feedback (FB) neural network schemes can be reduced from the new general model. The results demonstrate that the new scheme is more robust and offers superior convergence properties.A technique for analyzing artificial neural network based protective relays
http://dl-live.theiet.org/content/conferences/10.1049/cp_20010193
Power system protection researchers and engineers have suggested a number of relay designs based on the artificial neural networks. This paper presents a technique for analyzing the neural outputs of a multilayer feedforward neural network (MFNN) based relay. The proposed technique is applied to a MFNN based fault direction discriminator.IP ware for neural networks
http://dl-live.theiet.org/content/conferences/10.1049/ic_20000416
In order to fill the gap between the increasing silicon capability available to IC designers new design methodologies, based on reusable Intellectual Property (IP) cores, are being developed. This evolution of circuit design has led us to investigate the casting of Artificial Neural Networks (ANNs) into a form suitable for use as IP ware. The Modular Map design is a fully digital implementation of an ANN. It was first implemented using a combination of different design techniques in a 0.65 μm process. Then a new model in the form of a Register Transfer Level (RTL) synthesisable VHDL description suitable for IP ware applications has been developed. This paper presents a comparison of the original implementation of a single neuron in the Modular Map design and a fully synthesised version targeted towards a number of standard cell technologies. Each neuron contains a small local register file which has been implemented using synthesised VHDL, synthesis tool supplied generators, and where available full custom generators. Area performance data for these is given. Casting of the design to a parameterised form suitable for a variety of ANN application areas is also discussed. (6 pages)Genetic learning for direct inverse neural network control
http://dl-live.theiet.org/content/conferences/10.1049/ic_20000347
The application of neural nets (NN) as a direct inverse controller for general nonlinear systems is considered. Since little knowledge of the nonlinear plant is normally available, it is difficult to obtain an analytical expression for its Jacobian. Thus, an emulator is required as a channel to compute the derivative of the output with respect to the input for NN training. Neural net training using genetic algorithms (GA) offers several advantages. No understanding of the plant model is required. Since no derivative computations are involved, it is less likely for these algorithms to get trapped in local minima. The scheme generates individual controllers with the best fitness values. A hybrid coding method and several appropriate modifications of the classical genetic algorithms for NN control purposes are discussed. To overcome the difficulties of saturation and fluctuation in the controller output, the output of the NN controller is obtained as the sum of several small sigmoidal functions. This effectively increases the linear range of operation of controller output without affecting the nonlinear feature of a sigmoidal function. It is noted in this case that, better control is achieved. Fuzzy logic with dynamic features is used to provide an optimal direction for genetic search. It, thus, speeds up the process of convergence by bringing the chromosomes near to the problem space and bringing more exploration amongst the most desirable ones. The method is demonstrated with the control of a single-link flexible manipulator. (4 pages)Neural-learning control of nonlinear dynamical systems
http://dl-live.theiet.org/content/conferences/10.1049/ic_20000346
This paper is concerned with neural-learning control of nonlinear dynamical systems. Two control schemes using neural-learning techniques are introduced. One is adaptive neural control and the other is predictive neural control. An application of neural-learning control to an industrial combustion system is also included. (7 pages)Associative reinforcement learning for discrete-time optimal control
http://dl-live.theiet.org/content/conferences/10.1049/ic_20000342
This paper investigates the application of associative reinforcement learning techniques to the optimal control of linear discrete-time dynamic systems. Associative reinforcement learning involves the trial and error interaction with a dynamic system to determine the control actions that optimally achieve some desired performance index. The methodology can be applied either online or off-line and in a model based or model free manner. Associative reinforcement learning techniques are applied to the optimal regulator (LQR) control of discrete-time linear systems. Adaptive critic designs are implemented and the convergence speed compared for the different approaches. These methods can determine the optimal state and state/action value function and the optimal policy without requiring system models. (4 pages)Application of wavelet transform and neural networks to fault location of a teed-circuit
http://dl-live.theiet.org/content/conferences/10.1049/ic_20000564
A new technique using wavelet transform and neural network for fault location in a teed-circuit is proposed in this paper. Fault simulation is carried out in EMTP96 using a frequency dependent transmission line model. Voltage and current signals are obtained for a single phase (phase-A) to ground fault at every 500 m distance on one of the branches, which is 64.09 km long. Simulation is carried out for 3 cycles (60 ms) with step size at, of 2.J μs to abstract the high frequency component of the signal and every 100 points have been selected as output. Two cycles of waveform, covering pre-fault and post-fault information are abstracted for further analysis. These waveforms are then used in wavelet analysis to generate the training pattern. Two different mother wavelets have been used to decompose the signal, from which the statistical information is abstracted as the training pattern. RBF network was trained and cross-validated with unseen data. (5 pages)A neuro-fuzzy approach for ramp metering
http://dl-live.theiet.org/content/conferences/10.1049/cp_20000113
Ramp metering, in the context of ITS systems, offers several operational features for improving traffic flow, traffic safety and air quality by the regulation of input flow to a freeway or highway. A general fuzzy ramp control algorithm is described and an example is examined. This fuzzy control algorithm is arranged into a neuro-fuzzy architecture and the membership functions are learned by a neural network based method from the US demand-strategy method.Modelling and simulation of a variable speed stand-alone generator system
http://dl-live.theiet.org/content/conferences/10.1049/cp_20000275
Modelling and simulation studies for a diesel driven stand-alone power generating system are presented. A new control strategy for system operation, based on neural networks and fuzzy logic, is reported. Simulation has shown that the resulting controller provides high stability. The system has the advantages of being compact, reliable and cost effective.ANN controlled battery energy storage system for enhancing power system stability
http://dl-live.theiet.org/content/conferences/10.1049/cp_20000416
This paper describes an application of an adaptive artificial neural network (ANN) controller to continuously control the charging and discharging of a battery energy storage system (BESS) to improve the stability of an electric power system. The simulation studies have included a detailed model of the generator including its excitation controller and governor, as well as a comprehensive BESS model, including the DC battery model and the switch operation associated with the power converter. An online training artificial neural network controller is continuously trained to directly control the BESS operation to damp power system oscillations in various power system operating conditions. Simulation results show that this ANN-controller can adaptively learn and update its control strategy to improve the system stability under different system operating conditions.Clustering epileptiform discharges with an adaptive subspace self-organizing feature map: a simulation study
http://dl-live.theiet.org/content/conferences/10.1049/cp_20000344
We present the results of a study where synthetically generated epileptiform discharges (EDs) superimposed on normal background EEG are clustered by means of Kohonen's self-organizing feature map (SOFM) using a set of basis vectors representing adaptive subspaces in place of the more usual weight vector at each node of the network. A training set of synthetic EDs is generated using a spherical head model assuming current dipole ED generators. The synthetic EDs are superimposed onto normal background EEG and a preliminary pre-processing stage is used to extract candidate EDs (CEDs) consisting of ED and non-ED events. The data is clustered using an adaptive subspace algorithm and the resulting map is calibrated using the labeled synthetic data set. Preliminary results show that the SOFM is well suited to clustering the pre-processed CEDs, where strong clusters of real EDs are evident. The next step of this research is to further our investigations into the clustering of EDs using real data extracted from the interictal EEG.The cerebellum as a neuronal machine
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990258
Summary form only given. The exceedingly regular and intricate neuroanatomy of the vertebrate cerebellar cortex has prompted many speculations about its function. The author reviews and assesses models for the functioning of cerebellar cortex, including those suggesting that it is: a timing device, a device enabling spatial navigation, or an associative memory store which is used in motor control. (1 page)Computational delays and habits
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990260
The brain is a slow computer yet humans can skilfully play games where very fast reactions are required. A solution to that problem is to bypass the slow action planning process and map directly perceptions to plans. It is proposed that the cerebellum has a dual function. Whereas in the intermediate and medial areas, it has the well known function of an inverse model of the motor system; in its lateral zone, it learns which plans are the most appropriate responses to a set of future perceptual situations, for a given goal and preselects them via cerebrocortical projections for execution under sensory triggering. This scheme saves planning time but, in absence of cognitive gating, enables sometimes inappropriate habitual behaviour. Hence habits may be a by-product of a computational strategy designed for compensating for computational planning delays. Robot control strategies inspired by the cerebellum are proposed, whereby planning is performed off-line and fed to a sequence learning system or task specific sets of plans are prepared in advance for fast selection using sensory inputs. Such schemes have been implemented using artificial neural networks. (5 pages)Novelty detection in jet engines
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990187
Neural network classifiers can be trained to estimate the posterior probability of a fault occurring given the values of a set of input parameters. With jet engines, however, faults are extremely rare and hence their prior probability is very low. The principle of novelty detection offers an alternative approach to the problem of fault detection. Novelty detection only requires the normal class to be defined. A statistical description of normality is learnt by including normal examples only in the training data; abnormalities are then identified by testing for novelty against this description. A real advantage of novelty detection is that anomalies which have not previously been seen will also be highlighted. (5 pages)The cerebellum and visually controlled movements
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990257
The author reviews some of the data from human and animal models pointing to a critical role for the cerebellum in the control of visually-guided movements. There are now several contrasting theories that suggest what this role may be, evidence from functional imaging, lesion studies, anatomy, and computational modelling supports the theory that the cerebellum forms a forward model of the motor system. This may be used for control (as suggested by the `Smith predictor' hypothesis); it may also underlie a cerebellar role in co-ordination, motor planning and in predicting the sensory consequences of movements. (5 pages)Recurrent disinhibition implements a pointer-map module between cerebellar basket and Purkinje cells
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990259
Summary form only given. Pointers to structures are a powerful software tool for implementing algorithms. Pointing limbs and foveating eyes are important for focal sensory processing in the brain and for motor interaction with the external world. Hahnloser et al. have recently developed a simple neuronal pointer architecture that can express these properties. This framework predicts dynamic features of the recurrent inhibitory interaction between cerebellar basket and Purkinje cells for selective signal enhancement. (1 page)Predictive control using multiple model networks
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990533
The aim of this paper is to describe a nonlinear modelling architecture, called the local model network (LMN), which introduces transparency while offering distinct advantages for nonlinear model-based control. Simulation results for a pH neutralisation process are used to illustrate the performance benefits of LMNs for two novel nonlinear dynamic matrix control schemes. (7 pages)Artificial intelligence approaches to fault diagnosis
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990188
Fault diagnosis of control engineering systems can be based upon the generation of signals which reflect inconsistencies between the fault-free and faulty system operation-so-called residual signals. This paper outlines some recent approaches to the generation of residual signals using methods of integrating quantitative and qualitative system knowledge, based upon AI techniques. (18 pages)Rapid best first retrieval from massive dictionaries with poorly segmented inputs
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990213
The author describes recent extensions to his rapid retrieval system. The initial system was developed to perform best-first retrieval from massive dictionaries given uncertain inputs, and has since been extended to provide a complete and uniform method for incorporating most kinds of contextual knowledge in the pattern recognition process. The modification described is a significant one, whereby the input to the system is now given as a directed graph, rather than a straight sequence of character hypothesis sets. The original formulation only dealt with substitution errors. The extended version copes with insertion and deletion errors, and any number of alternative segmentations of the input, as would naturally arise in cursive script or due to touching printed characters. This research builds on the syntactic neural network (SNN) rapid dictionary search system which is briefly described. (6 pages)Learning to reach via corrective movements
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990261
Summary form only given. When infants begin to perform goal-directed reaching, the kinematics of their reaches show multiple accelerations and decelerations of the hand, which appear to reflect a correcting series of submovements. Although each submovement is an inaccurate correction, the sequence of submovements is often successful in reaching the target. Under some circumstances, such as when learning a novel task or when high accuracy is required, adult behavior also consists of a series of submovements that appear to be corrective in nature. We investigate the hypothesis that this reflects the action of two primary processes: one that regulates the metric properties of the submovements, and one that initiates them at the proper times and in roughly in the correct directions. We hypothesize that the cerebellum is the key structure underlying the first of these processes. The cerebellum is modeled as an adaptive, predictive controller that regulates movement by learning to react in an anticipatory fashion to sensory feedback. (1 page)Cerebellar versus stiffness control
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990262
The mechanical properties of the muscular actuators are important for dynamics (compensation of internal/external loads and disturbances) as well as for spatial cognition. Consider the so called λ-model (Feldman and Levin, (1995)). The key point, from the point of view of space representation, is that that the controlled variable has the same dimensions of the control variable. The proprioceptive space can be defined as a lower-dimensional manifold in the set of all possible configurations and can be represented in a distributed way in terms of a self-organising cortical map. Its dimensionality and geometry is coded into the pattern of lateral connections that are excitatory and recurrent. These connections can be adapted by means of unsupervised Hebbian learning within a (probably innate) behavioural strategy of circular reaction. The same strategy and unsupervised learning paradigm allow the emergence of an internal distributed representation of the exteroceptive space and the nonlinear mapping between the two spaces, exploiting a side-effect of circular reaction. (3 pages)Knowledge extraction and insertion from radial basis function networks
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990372
Neural networks provide excellent solutions for pattern recognition and classification problems. Unfortunately, in the case of distributed neural networks such as the multilayer perceptron it is difficult to comprehend the learned internal mappings. This makes any form of explanation facility such as that possessed by expert systems impractical. However, in the case of localist neural representations the situation is more transparent to examination. This paper examines the quality and comprehensibility of rules extracted from localist neural networks, specifically the radial basis function network. The rules are analysed in order to gain knowledge and insight into the data. We also investigate the benefits of inserting prior domain knowledge into a radial basis function network. (6 pages)Support vector machines: a tutorial overview and critical appraisal
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990359
Summary form only given. There has been much interest in the use of support vector machines (SVM) as an approach to high performance pattern classification. In the linearly separable case, SVMs attempt to position a class boundary so that the margin from the nearest example is maximised. This criterion can be implemented by solving a quadratic programming problem, and the solution turns out to be one in which the class boundary may be expressed as a linear combination of a subset of the training data (the support vectors). The elegance of the QP formulation, and the relationship between control of complexity in this formulation and Vapnik-Chervonenkis dimensions are seen as prime attractions of the SVM method. A related idea in high performance pattern classification is that of boosting multiple classifiers. The author shows that the standard SVM formulation is not robust to noise and explains the performance of boosting algorithms by reference to receiver operating characteristics curves. (1 page)An overview of cerebellar control
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990256
Cerebellar models were first applied for robot dynamics control after Albus' CMAC model was published in 1975. Especially the later implementations by Miller et al. (1989, 1997) for control of a 4-DOF robot arm as well as biped control have demonstrated the power of this approach. Although cerebellar modelling has come a long way since then, applications of such models remain limited to toy problems, not exceeding the complexity of (simulated) two-link robot arms. (2 pages)On-line implementation of a model predictive controller on a multivariable chemical process
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990539
An investigation into neural network model predictive control is described in this paper. The control strategy developed is applied to a laboratory process to control temperature, pH and dissolved oxygen. The main difficulties in control of this process are nonlinearity, coupling effects among variables and long time-delay in the heat exchanger. Parallel neural models are developed from real process data for the use with online model predictive control and off-line simulations. The online control results are demonstrated. (5 pages)Neurofuzzy state estimators and their applications
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990717
Neurofuzzy algorithms have been extensively developed for the real time/online identification of nonlinear a priori unknown dynamical processes. As with all rule base paradigms they suffer from the curse of dimensionality, restricting their practical use to low dimensional control problems. The paper shows how adaptive construction algorithms based on additive decomposition techniques can overcome this problem, to produce parsimonious neurofuzzy models which retain their transparency or interpretability. Not only does this approach extend the applicability of neurofuzzy algorithms, it also enables low complexity controllers and estimators to be derived. In this context neurofuzzy state estimators are derived, which automatically parameterise a Kalman filter for a process state estimate reconstruction from any input/output data source. This approach avoids pitfalls of the extended Kalman filter, and is optimal for local models. The paper discusses real world applications of this new theory of modelling and estimation to helicopter guidance, intelligent driver warning system, communication antennas, autonomous underwater vehicles, ship collision avoidance guidance, and an IFAC benchmark problem. (10 pages)Low-cost optical neural-net torque transducer
http://dl-live.theiet.org/content/conferences/10.1049/ic_19990775
Torque signals are used today across a broad range of automation applications such as dynamometer test stands and web process lines. Some of the drawbacks to using commercial torque transducers include their cost, reliability, and mechanical stiffness. It is possible to utilize an artificial neural network coupled with the photoelastic effect of many polymers to provide a small, low-cost torque sensor. The availability of low-cost torque sensing opens up many new applications that previously did not cost justify the investment in a commercial torque sensor. The sensor material has a high bandwidth which enables sampling high-frequency torque signals. High frequency torque signals are particularly useful for investigating machinery dynamics and for machinery health assessment. (4 pages)The self-organizing map of attribute trees
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991103
The standard version of the self-organizing map (Kohonen (1990)) applies vector data. This paper explains how attribute trees can be used as the learning medium in the self-organizing map. As a data structure, a tree is an optimal presentation of many hierarchical and dynamical objects appearing in natural phenomena and human activities. The proposed approach is based on introducing a distance metric and adjusting schemes for attribute trees. The trees are assumed to be rooted and unordered. The key idea is in heuristic matching which provides approximate results but above all avoids the exponential complexity of exact matching. The feasibility of the suggested methods is demonstrated with an experiment on weather radar imagery.Optimization of surface component mounting on the printed circuit board using SOM-TSP method
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991183
We propose the application of self organising maps-travelling salesman problem (SOM-TSP) method in optimizing the efficiency of surface mounting of electronic parts on the printed circuit board. From the numerical experiment, it was found that the required time for mounting electronic parts can be decreased by our proposed method compared to the built-in method on the mounting-system. As a result, the production output can also be increased using the proposed method.Keyword selection method for characterizing text document maps
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991137
Characterization of subsets of data is a recurring problem in data mining. We propose a keyword selection method that can be used for obtaining characterizations of clusters of data whenever textual descriptions can be associated with the data. Several methods that cluster data sets or form projections of data provide an order or distance measure of the clusters. If such an ordering of the clusters exists or can be deduced, the method utilizes the order to improve the characterizations. The proposed method may be applied, for example, to characterizing graphical displays of collections of data ordered (e.g. with SOM algorithm). The method is validated using a collection of 10000 scientific abstracts from the INSPEC database organized on a WEBSOM document map.Products of experts
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991075
It is possible to combine multiple probabilistic models of the same data by multiplying the probabilities together and then renormalizing. This is a very efficient way to model high-dimensional data which simultaneously satisfies many different low dimensional constraints. Each individual expert model can focus on giving high probability to data vectors that satisfy just one of the constraints. Data vectors that satisfy this one constraint but violate other constraints will be ruled out by their low probability under the other expert models. Training a product of models appears difficult because, in addition to maximizing the probabilities that the individual models assign to the observed data, it is necessary to make the models disagree on unobserved regions of the data space. However, if the individual models are tractable there is a fairly efficient way to train a product of models. This training algorithm suggests a biologically plausible way of learning neural population codes.Multi-layer perceptrons as nonlinear generative models for unsupervised learning: a Bayesian treatment
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991078
In this paper, multilayer perceptrons are used as nonlinear generative models. The problem of indeterminacy of the models is resolved using a recently developed Bayesian method, called ensemble learning. Using a Bayesian approach, models can be compared according to their probabilities. In simulations with artificial data, the network is able to find the underlying causes of the observations despite the strong nonlinearities of the data.A new adaptive architecture: analogue synthesiser of orthogonal functions
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991195
A new adaptive nonlinear (neural-like) architecture, an analogue synthesiser of orthogonal functions which is able to produce a plurality of mutually orthogonal signals as functions of time such as Legendre, Chebyshev and Hermite polynomials, cosine basis of functions, smoothed cosine basis, etc., is proposed. A proof-of-concept breadboard version of the analogue synthesiser is described. The device is characterised by a very fast (approximately 100 iterations) and stable process of signal synthesis. The proposed new device could find applications e.g. in analogue systems of function approximation, in particular as a main unit in an analogue implementation of so-called Chebyshev polynomial-based (CPB) neural networks, as a unit in a fast adaptive alternative to Volterra polynomial neural networks, and also as a preprocessing element (performing some transforms, filtration, etc.) in analogue neural network-based systems of information processing.Self-organizingly emerging activeness architecture realized by coherent neural networks
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991197
Coherent activeness architecture: a self-organizing activeness (or intentionality) architecture using coherent neural networks is proposed for the future brain-type information processing systems. It utilizes coherent neural network modules whose behavior (forward processing, learning, self-organization, etc.) is controlled by their carrier-wave frequencies. The behaviors constructs an orthogonal basis-function set in frequency domain. Therefore the network is able to obtain an arbitrary (continuous/discrete) property as a combinatorial profile. Through this mechanism, an intentionality-controlling network self-organizes adaptively into a symbol-, pattern-, or intermediate processing circuit. The architecture realizes active behaviors such as attentions in recognition tasks and mode selections in motor actions that are considered indispensable for the future brain-type systems. Simulation results are also presented which demonstrate an environment-dependent self-organization of the activeness mechanism.Backtracking deterministic annealing for constraint satisfaction problems
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991221
We present a deterministic annealing approach to the solution of quadratic constraint satisfaction problems with complex interlocking constraints, such as exemplified in polyomino tiling puzzles. We first analyze the dynamical properties of the solution strategies implemented by deterministic annealing (DA) in the analog neural representation of Potts-mean-field (PMF) and penalty-function-based competitive layer model (CLM) neural networks, revealing a similar mechanism. The key idea of our extension of these plain DA approaches is motivated by classical backtracking algorithms. We show that their ability for iterative local pruning of the search space can be implemented within the framework of DA by introducing local temperature parameters which are “reheated” when locally unresolved conflicts occur. To achieve the pruning of the search space, reheating is accompanied by a modification of the constraint-implementing weight matrix to reduce the chance of reentering the same local configuration. The weight changes provide a learning mechanism that facilitates the generation of a solution for subsequent runs. We demonstrate the benefits of the resulting “backtracking deterministic annealing” algorithm (BDA) by applying it to a pentomino tiling problem. We show that the method reliably finds perfect solutions to the task, while the plain DA approach for both PMF and CLM is unable to solve the task in a comparable or even considerably larger number of iterations.A neural network for scene segmentation based on compact astable oscillators
http://dl-live.theiet.org/content/conferences/10.1049/cp_19991191
We show the feasibility of building a neural network for scene segmentation made of astable oscillators. The network is based on Wang and Terman's algorithm, LEGION . However, much simpler astable circuits have substituted the original oscillators so they meet analog microelectronic requirements and can achieve a high integration level. The correct behavior of the modified network and some of its non-idealities inherent to analog VLSI (as time delay effects) are shown by means of simulations.