Deborah A. Stacey

Clustering unlabeled data with SOMs improves classification of labeled real-world data

We show the use of a self organizing map to cluster unlabeled data and to infer possible labelings from the clusters. Our inferred labels are presented to a multilayer perceptron along with labeled data, performance is improved over using only the labeled data. Results are presented for a number of popular real-world benchmark problems from domains other than text. This shows one way in which unlabeled data can be used to enhance supervised learning in a general-purpose neural network.

A solution to vicinity problem of obstacles in complete coverage path planning

In real world applications there exist arbitrarily shaped obstacles in the workspace during complete coverage path planning of cleaning robots. A cleaning robot should be able to sweep in a variety of corners and in the vicinity of arbitrarily shaped obstacles in an indoor environment. Consequently, the robot is required not only to effectively avoid the obstacles, but also to delicately cover every area in the vicinity of obstacles. In the paper, a solution to vicinity problem of obstacles in complete coverage path planning is proposed using neural-neighborhood analysis. The path planner is a biologically inspired neural network. The proposed model is capable of planning a real-time path to reasonably cover every area in the vicinity of obstacles. The robot path is autonomously generated through the dynamic neural activity landscape of the neural network and the previous robot location. The effectiveness of the proposed approach is verified through computer simulations.

Real-time path planning with deadlock avoidance of multiple cleaning robots

In this paper, a cooperative sweeping strategy with deadlock avoidance of complete coverage path planning for multiple cleaning robots in a changing and unstructured environment is proposed, using biologically inspired neural networks. Cleaning tasks require a special kind of trajectory being able to cover every unoccupied area in specified cleaning environments, which is an essential issue for cleaning robots and many other robotic applications. Multiple robots can improve the work capacity, share the cleaning tasks, and reduce the time to complete sweeping tasks. In the proposed model, the dynamics of each neuron in the topologically organized neural network is characterized by a shunting neural equation. Each cleaning robot treats the other robots as moving obstacles. The robot path is autonomously generated from the dynamic activity landscape of the neural network, the previous robot location and the other robot locations. The proposed model algorithm is computationally efficient. The feasibility is validated by simulation studies on three cases of two cooperating cleaning robots. The multiple cleaning robots sweeping will not be trapped in deadlock situations.

A hierarchical artificial neural network system for the classification of cervical cells

The task of cervical cell classification can be divided into four sub-tasks: (1) the isolation of single cells, cell clusters and clumps as well as artifacts, (2) the segmentation of the cell image into nucleus and cytoplasm, (3) the extraction of cell features such as size and density of the nucleus and cytoplasm, grey level extrema, fractal dimension, texture parameters and shape measures, and (4) the use of these features to classify the cell as normal or abnormal. The final problem of formulating a diagnostic decision based on these data is a multivariate statistical one, to which there are many theoretical and practical solutions. Palcic et al. (1992) have performed a discriminant function analysis of a large set of such measurements, and have achieved a high predictive accuracy. This paper describes a solution for the cell classification task which utilizes a hierarchical system of artificial neural networks (ANNs) using backpropagation (BP) and achieves extremely high accuracy.<<ETX>>

Feature subset selection via multi-objective genetic algorithm

Real-world datasets tend to be complex, large in size, and may contain many irrelevant features. Eliminating such irrelevant features can significantly improve the performance of a data mining algorithm. In this paper, we propose a multi-objective genetic algorithm that finds a set of Pareto-optimal feature subsets that works as a wrapper around a standard back-propagation algorithm. We also introduce a novel mechanism called the least-crowded selection algorithm that maximizes the diversity of the solutions returned by the algorithm. We justify the proposed method by theoretically and empirically comparing it to the backpropagation neural network and the simple genetic algorithm for feature selection.

Understanding the Parallel Programmer

As low-cost multiprocessing reaches a wider market, a greater number of programmers will need to be trained for parallel programming. Current studies exploring usability engineering for parallel programming focus only on experienced parallel programmers. This paper applies the card sorting method used in psychology research to understanding the software needs of the novice parallel programmer. This paper demonstrates that novices organize parallel problems by domain type whereas experts use parallel communication type.

Adaptive exploration in reinforcement learning

The exploration/exploitation trade-off is a difficult problem for a reinforcement learning agent. A non-stationary environment coupled with current connectionist implementations of reinforcement learning algorithms is a recipe for disaster. Towards a solution for such situations we introduce a novel technique, called past-success directed exploration, and an implementation of reinforcement learning algorithms based on the fuzzy ARTMAP architecture. We compare through experimentation features of a traditional approach with our own.

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.

Solving the Traveling Salesman Problem Using the Enhanced Genetic Algorithm

An extension to the Enhanced Genetic Algorithm (EGA) analysis of Gary W. Grewal, Thomas C. Wilson, and Deborah A. Stacey [1] is introduced and applied to the TSP. Previously the EGA had successfully handled constraint-Satisfaction problems, such as graph coloring. This paper broadens the application of the EGA to the specific NP-hard problem, the Traveling Salesman Problem (TSP). The first part of this paper deals with the unique features of the EGA such as running in an unsupervised mode, as applied to the TSP. In the second part, we present and analyze results obtained by testing the EGA approach on three TSP benchmarks while comparing the performance with other approaches using genetic algorithms. Our results show that the EGA approach is novel and successful, and its general features make it easy to integrate with other optimization techniques.

ART2 based classification of sparse high dimensional parameter sets for a simulation parameter selection assistant

This paper presents the design and creation of a simulation parameter selection assistant (SPSA) that helps modeling researchers choose meaningful values for their complex simulations, and encourages collaboration between teams searching through high dimensional parameter spaces. Proposed simulation parameters are compared to past runs using adaptive resonance theory to measure similarity with the goals of preventing repetitive exploitations of parameters and of encouraging the exploration of new regions of the parameter space. The assistant was designed to be used as part of a high performance animal disease spread simulator but is general and modular enough to be easily adapted to other simulation and search domains.