Abstract Problems such as data compression, pattern recognition and artificial intelligence often deal with a large data sample as observations of an unknown object. An effective method is proposed to fit hyperplanes to data points in each hypercubic subregion of the original data sample. Corresponding to a set of affine linear manifolds, the locally fitted hyperplanes optimally approximate the object in the sense of least squares of their perpendicular distances to the sample points. Its effectiveness and versatility are illustrated through approximation of nonlinear manifolds Möbius strip and Swiss roll, handwritten digit recognition, dimensionality reduction in a cosmological application, inter/extrapolation for a social and economic data set, and prediction of recidivism of criminal defendants. Based on two essential concepts of hyperplane fitting and spatial data segmentation, this general method for unsupervised learning is rigorously derived. The proposed method requires no assumptions on the underlying object and its data sample. Also, it has only two parameters, namely the size of segmenting hypercubes and the number of fitted hyperplanes for user to choose. These make the proposed method considerably accessible when applied to solving various problems in real applications.

As a consequence of aging population and an increasing prevalence of obesity and diabetes there are more and more patients with heart failure. This leads to a lack of professionals who can treat them and to escalating costs. An interesting solution appears to be home telemonitoring with an intelligent clinical decision support system. In this paper, the use of cumulative information estimations for risk assessment of heart failure patients with such a system is analyzed. These cumulative information estimations are utilized for creation of an algorithmic model using fuzzy decision trees that combine decision trees and notions of fuzzy logic. The algorithmic model employs mutual cumulative information and relative mutual cumulative information for association of an important piece of data about the patients with a decision node. The risk assessment with the presented solution is analyzed from the point of view of minimization of life-threatening situations and minimization of costs. Comparisons with a Bayesian network method, a nearest neighbor method, and a logistic regression method show it is a promising solution.

Real live clinical data often present itself with a number of usual challenges, such as class imbalance, high dimensionality and missing data. There is the added complexity of the data being distributed non-uniformly and skewed. Thus the performance of classical classification methods with this type of data is lower than with other types of data. Classification based on Bayes is often suggested as a better method, however, the typical assumption made for Bayes, such as variable and data distributions are not satisfied by real clinical data. This paper focuses on improving the performance of Bayesian classifiers but also on how the underlying structures of the data affects the performance. Thus this paper will focus on Bayesian methodologies, namely use of non-parametric Kernel Density Estimation (KDE) and Tree Augmented Naïve Bayes (TAN). The aim is to measure the performance on the heart failure dataset and by focusing on how the data structure improves the classification. The missing data present in the clinical heart failure datasets are replaced using two imputation methods and results compared. We also apply the imputed datasets on three classifiers including J48 (decision tree), naïve Bayesian multinomial and Bayesian network. The experiments show an improvement on the naïve Bayes using KDE, however TAN achieves significant improvement with the different missing value imputation methods. It is seen that TAN not only improves performance of the classifier, but also enhances prediction accuracy while maintaining efficiency and model simplicity.

The prevalence of heart failure is 2-3% of the adult population and it is expected to grow. Half of all patients diagnosed with it die within four years. To minimize life-threatening situations and to minimize costs, it is interesting to predict mortality rates for a patient with heart failure. In this paper, a fuzzy decision tree based on classification ambiguity and a fuzzy decision tree based on cumulative information estimations are presented. They are employed on a heart failure data fuzzified on the basis of medical expert knowledge. After a transformation of fuzzy decision trees, the use of medical expert knowledge allows us to create a group of fuzzy rules that is easily interpretable by medical experts. Our study shows that different types of fuzzy decision trees can have significantly different accuracy results and interpretability.

The characteristics of live clinical datasets comprise of: high dimensionality, missing data, class imbalance, non-normal distribution, noisy and inconsistency. This paper focuses on missing data and high dimensionality, both of which are known to affect classification and the design of decision support systems. Naive Bayes classifier was employed to explore how both complexities are handled. Experimental result showed that imputing missing values does not improve the performance of naive Bayes. However, a wrapper subset evaluator that employed forward and backward search strategies for reducing the data dimension had an effect on performance. The methods determined an optimal set of features and reduced numerical and time complexity while maintaining a high degree of accuracy. These methods were compared in two different datasets, a life lab and iris dataset.

This paper deals with fault-tolerant control of a network controlled systems (NCS) problem, where the sensors, actuator and controller are inter-connected via a communication network. A procedure is proposed for controlling a system over a network using the concept of an NCS-information-packet which is an augmented vector comprising control moves and fault flags. The size of this packet is used to define a completely fault tolerant NCS. The behavior and control of this scheme is illustrated by way of an example, where the plant is being controlled over a network. Implicit in this paper is that appropriate FDI schemes exist within the set up. The software environment used is MATLABcopy and LABVIEWcopy. The results illustrate that the scheme is tolerant to faults.

Anti-Personal landmines are a significant barrier to economic and social development in a number of countries. Several sensors have been developed but each one will probably have to find, if it exists, a specific area of applicability, determined by technological as well as economical or even social factors, and possibly other sensors to work with using some form of sensor fusion. A significant issue concerns the safety of the deminer. For every 5000 mines removed, one deminer is killed. The design of an accurate sensor may reduce the amount of time needed to determine whether a landmine exists, but does not increase the safety of the deminer. Since the safety issues during the eradication process are of great concern, the use and integration of cheap and simple robots and sensors in humanitarian demining has been raised. This removes the deminer from close contact with mines. This paper is concerned mainly with the work that has been done in the area of robotics and landmine detection, the problems involved, the issues that have been overlooked and the future of robots in the field.

Cluster analysis is traditionally an unsupervised data. reduction technique. However, by unifying ideas from both cluster analysis and rough set theory, the inherent structure of a data set can be expressed as a rule set which, in turn, can be modified in a supervised manner to obtain a decision rule set with minimal ambiguity. This process of encasing knowledge extraction in an algorithmic framework results in an optimal process for decision support.

The human motor-control system has a hierarchical and decentralized structure, and building such a control system for a multiple robot system would require a decomposed model. The difficulty in decomposing complex robotic systems is due to interactions between robots, and the paper proposes a control architecture that controls the manipulator joints and interactions separately. A decomposable model for an N - n degrees-of-freedom multimanipulator/object system handling an object is derived. This model is then used to design a Lyapunov-based fuzzy logic controller for the system by solving linear matrix equalities. It is shown that this controller is closed-loop stable and a stable suboptimal controller for the system may be designed using bounds.

When designing a Takagi-Sugeno (TS) fuzzy controller, reducing the number of rules makes it easier to derive a stable Lyapunov based controller using linear matrix inequalities (LMIs). The paper describes an intuitive way of reducing a fuzzy ruleset by eliminating rules that are very similar in terms of parameters and membership functions. It is then suggested that a more formal approach to this intuitive idea may be found in rough set theory.

This paper describes the aspects involved in modelling a multi-robotic system from a human motor control perspective. Building a control system for a multi-robot system that works in a similar fashion to the human motor control system would require a decomposable model. Decomposition of a complex robotic system is difficult due to the interactions between the subsystems. The proposed method of separating the interconnections is applied with the aid of fuzzy modelling to derive a fully decomposable model of two manipulator robots handling a common object.

This work provides a framework for the approximation of a dynamic system of the form x/spl dot/=f(x)+g(x)u by dynamic recurrent neural network. This extends previous work in which approximate realisation of autonomous dynamic systems was proven. Given certain conditions, the first p output neural units of a dynamic n-dimensional neural model approximate at a desired proximity a p-dimensional dynamic system with n>p. The neural architecture studied is then successfully implemented in a nonlinear multivariable system identification case study.