David Calvert

Self-Learning Visual Servoing of Robot Manipulator Using Explanation-Based Fuzzy Neural Networks and Q-Learning

A new self-learning visual servoing system for the robot manipulators is proposed. This system includes two main properties: on-line self training and lifelong learning that are implemented by the Q-Learning algorithm and Explanation-based Fuzzy Neural Networks (EBFNN) respectively. We demonstrate that the number of training samples and the training time for a specific robot positioning accuracy can be reduced using explanation-based fuzzy neural networks and the Q-Learning algorithm. The system uses Q-learning to find the optimal policy in conjunction with the reinforcement learning. This policy is used by a robot to reach an object that has been randomly placed in a static workspace. Background knowledge about the robot and its environment is transferred to the robot agent during the learning process using a set of previously trained neural networks. This system learns the optimal policy in order to select the best action that maximizes the cumulative reward received at each time step. This learning approach does not use either a robot or camera model, or require calibration. Simulation results prove the effectiveness of this methodology to improve the learning process and the performance of the self-learning visual servoing system.

Distributed artificial neural network architectures

The computational cost of training artificial neural network (ANN) algorithms limits the use of large systems capable of processing complex problems. Implementing ANNs on a parallel or distributed platform to improve performance is therefore desirable. This work illustrates a method to predict and evaluate the performance of distributed ANN algorithms by analyzing the performance of the comparatively simple mathematical operations, which are used to construct the ANN. The ANN algorithms are divided into simple components: matrix and vector multiplication, matrix processed through a function, competition in a matrix. These basic operational parts are examined individually and it is demonstrated that the computation processes of distributed neural networks can be derived from the composition of these basic operations. Three popular network architectures are examined: multi-layer perceptrons with back-propagation learning, self-organizing map, and radial basis functions network.

Detection of disease outbreaks in pharmaceutical sales: neural networks and threshold algorithms

Syndromic surveillance involves monitoring data that could indicate disease trends a population, such as gastrointestinal illness and respiratory illness. Different types of data can be used to detect potential outbreaks of disease or biological contaminant based on deviations from historical norms. The system discussed in this paper is intended to detect aberration by identifying changes in sequence data that do not match the norms for a given time and location. Artificial neural networks (ANNs) were used to detect changes in the sales trends for over-the-counter (OTC) pharmaceuticals. Early detection of an outbreak allows public health officials to respond faster to potential outbreak situations. Our research examines the application of a multilayer perceptron using back-propagation learning and a moving window of the daily OTC sales values as inputs. The network is trained to identify changes in the sales trends which can be an indicator of a change in the populations health. The sales data exhibits a large amount of variability and the ANN must be trained to process this without prematurely signalling that a change has occurred. The network is trained using multiple years (hundreds) of simulated sales data containing simulated outbreaks. The success of the ANN is determined by its accuracy and by the amount of time (number of days into the outbreak) that the system takes to correctly signal that an anomalous trend is occurring.

A new Canadian interdisciplinary Ph.D. in computational sciences

Abstract In response to growing demands of society for experts trained in computational skills applied to various domains, the School of Computer Science at the University of Guelph is creating a new approach to doctoral studies called an interdisciplinary Ph.D. in computational sciences. The program is designed to appeal to candidates with strong backgrounds in either computer science or an application discipline who are not necessarily seeking a traditional academic career. Thesis based, it features minimal course requirements and short duration, with the student’s research directed by co-advisors from computer science and the application discipline. The degree program’s rationale and special characteristics are described. Related programs in Ontario and reception of this innovative proposal at the institutional level are discussed.

Neural based methods for music composition

Summary form only given, as follows. Several features found in neural networks make them desirable for music composition. These features include the ability to learn complex relationships from examples and to make reasonable generalizations in situations for which they have not been trained. Training consists of exposing the network to examples of music and adjusting its internal representation to match the input examples. It is this internalization of the music which represents many aspects and relationships that humans perceive both consciously and unconsciously. The method used to present the music to the network will therefore largely affect what features the network will extract. Several strategies involved in creating music using neural networks have been examined, and work involving multiple input training which mimics the functions of the right and left hemispheres of the brain has been carried out.<<ETX>>

Analysis of equine gaitprint and other gait characteristics using self-organizing maps (SOM)

Detection and evaluation of lameness by visual assessment requires the examiner to consider several different and rapidly changing body movement patterns. When the patterns become too complex, tools like artificial neural networks (ANN) can be useful. ANNs can be gainfully applied to gait analysis to distinguish stride characteristics and to identify pathological gait. The self-organizing map (SOM) is trained to cluster strain measurement data collected from a single hoof of moving horses. An analysis of the characteristics of the data and the effects of testing different data elements is examined. The method was successful in differentiating certain stride characteristics such as shoeing, gait, speed, and direction of movement and produced a unique model for each horses gait.

A distance based network for sequence processing

This work describes the development of an artificial neural network for sequence modeling and recall. The system described learns an initial data set and treats that as an exemplar for all later comparisons. One of the principles of this work was to design a system that takes advantage of the architectural features common to neural networks. These features are many simple storage locations (weights) and a collection of simple processing elements.

Evaluating symbolic representations in melodic similarity

Symbolic melodic similarity measures have been the subject of considerable investigation for their role in content-based querying, digital musicological analysis, and other data driven applications of Music Information Retrieval (MIR). Despite these efforts, there has been little focus on the representations or encodings employed by symbolic similarity measures, and how each of these representations affects the analysis that follows it. Understanding how these similarity measures behave can improve the way we index and retrieve digital musical content, and offer insights into the underlying musical patterns. This work explores how five melodic encodings, with varying information types and loss, behave using common string matching melodic similarity measures for exact and inexact matching, both globally and locally. The differences in the various symbolic melodic encodings are summarized to provide understanding and context as to when and in what applications these encodings could be applied.

A Robust System for Distributed Data Mining and Preserving-Privacy

Interest in knowledge-based collaborative applications has emerged due to the availability of large volumes of data that can be analyzed through the Internet. Many organizations in several domains are motivated to combine their recodes to improve the reliability and completeness of the extracted knowledge. However, they cannot disclose the records for privacy reasons. This work proposes a new robust multiple classifier system (MCS) called Dgadapt that identifies patterns distributed across multiple sites while avoiding the transfer of any records between those sites to preserve privacy. A classifier is built based on the local patterns of each site. All of these classifiers are transferred to a central site without the local records. On the central site, synthetic patterns are generated randomly. These patterns are labeled using the classifiers. A set of diverse MCSs are created from the classifiers based on the synthetic patterns. The classifiers of each MCS are selected to define subsets of the feature space that closely match the true class regions in different way. This work investigates the effect of using the synthetic patterns on the performance of Dgadapt. Two methods for labeling the synthetic patterns are examined. The first is to select a random classifier from the classifiers to label each synthetic pattern that is generated randomly; the second is to use major voting of the classifiers. The performance of Dgadapt when using real patterns is compared to the performance when using the two types of synthetic patterns. This demonstrates that using the synthetic data labeled based on the local classifiers does not show a significant difference in the performance of Dgadapt when compared to using real data.

Auto-calibration of Support Vector Machines for detecting disease outbreaks

Support Vector Machines (SVM) have several tuning parameters such as the kernel function type. This work proposes to develop an algorithm to calibrate the SVM automatically for detecting disease outbreaks based on Telehealth data. Two sets of simulated data are generated based on real Telehealth calls and an outbreak profile. The Telehealth data is related to respiratory disease syndrome. The outbreak profile is created based on real outbreak data. The first data set is used by the SVM to model the relation between call counts and the occurrence of a respiratory outbreak; however, the other data set is used for testing the resulting model. This model is auto-calibrated by optimizing four parameters using a Genetic Algorithm. These parameters are the tradeoff between the training error and the margin of the classifying hyperplane, kernel function type used, the hyperplane type used and the threshold level at which the occurrence of an outbreak is detected.