Thesis (Ph.D.)--Boston University / Advances in remote sensing and associated capabilities are expected to proceed in a number of ways in the era of the Earth Observing System (EOS). More complex multitemporal, multi-source data sets will become available, requiring more sophisticated analysis methods. This research explores the applications of artificial neural networks in land-cover mapping, forward and inverse canopy modeling and change detection. For land-cover mapping a multi-layer feed-forward neural network produced 89% classification accuracy using a single band of multi-angle data from the Advanced Solidstate Array Spectroradiometer (ASAS). The principal results include the following: directional radiance measurements contain much useful information for discrimination among land-cover classes; the combination of multi-angle and multi-spectral data improves the overall classification accuracy compared with a single multi-angle band; and neural networks can successfully learn class discrimination from directional data or multi-domain data. Forward canopy modeling shows that a multi-layer feed-forward neural network is able to predict the bidirectional reflectance distribution function (BRDF) of different canopy sites with 90% accuracy. Analysis of the signal captured by the network indicates that the canopy structural parameters, and illumination and viewing geometry, are essential for predicting the BRDF of vegetated surfaces. The inverse neural network model shows that the R2 between the network-predicted canopy parameters and the actual canopy parameters is 0.85 for canopy density and 0.75 for both the crown shape and the height parameters. [TRUNCATED]

A common problem in the area of non-linear function optimisation is that of not being able to guarantee finding the global optimum of the function in a feasible time especially when local optima exist. This problem applies to various areas of heuristic search. One of these areas concerns standard training techniques for feedforward neural networks. The element of heuristic search consists of attempting to find a neural weight state corresponding to the lowest training error. This problem may be termed the local minimum problem. The local minimum problem is addressed for feedforward neural networks. This is done by first establishing the conditions under which local minimum interference for the training process is to be expected. A target based approach to subgoal chaining in supervised learning is then investigated. This is a method to improve travel for neural networks by directing it more precisely through local subgoals than may be achieved through a more distant goal. It is shown however that linear subgoal chains are not sufficient to overcome the local minimum problem. Two novel training techniques are presented which use non-linear subgoal chains and are examined for their capability to address the local minimum problem. It is found that attempting to target a neural network to do something it cannot may lead to suboptimal training. It is also found that targeting a network to do something it is capable of generally leads to successful training. A novel system is presented which is designed to create optimal realisable targets for unrealisable goals. This allows neural networks to subsequently achieve the optimal weight state through a sufficiently powerful training method such as subgoal chaining. The results are shown to be consistent with the theoretical expectations.

No description available.

This thesis documents the research that has led to the development of a number of methodologies for combining existing artificial intelligence and statistical techniques into a form appropriate for the development of an intelligent appraisal system for use in the residential property appraisal profession. The methodologies illustrate how regression based appraisal models, previously restricted to homogeneous data, can be applied to heterogeneous data without significant loss in accuracy. The majority of research, previous to this, has addressed this problem by manually selecting homogeneous sub-regions from a heterogeneous parent region. However, the main drawback with this approach is that the segregation of parent regions into sub-regions relies upon a significant amount of a priori knowledge pertaining to the location of the property being valued. The requirement for a commercial residential property appraisal system is one that given sufficient training evidence can automatically learn how to value a property in any region and be able to modify this knowledge over time. Two methodologies are proposed within the thesis to address this requirement. The first, using a technique known as the Kohonen Self Organising Map, makes an assumption that residential properties that share sufficient characteristics can be appraised using the same function The Kohonen Self Organising Map is used to cluster properties with respect to their property characteristics and locational characteristics represented using a mortgage transaction database and UK Census statistics. Aptness of each cluster to define a homogeneous subset suitable to train a regression model, such as multiple regression analysis or a neural network, is estimated using a form of 'nearest neighbour' analysis. The second methodology, improves on the previous by transforming the static 'cluster then observe' solution to a more dynamic one using a Genetic Algorithm to evolve good clusters from those that at first inspection were mediocre. Another issue that has hindered the development of intelligent residential property appraisal systems has been the inability of such models to express their underlying functional form. This is addressed from two perspectives in this thesis: Rules are derived that describe the characteristic make-up of the formed clusters and, alternative modelling techniques are used to generate the final training models that are able to express their functional form as a set of induced rules. The work contained within this thesis demonstrates the feasibility of such an automatic stratification approach. Also, the research illustrates that by observing the characteristics of the generated clusters formed, a useful insight into both the underlying reasoning of the generated models and also of the locational and financial makeup of the subject location can be gained.

Global navigation satellite systems (GNSS), such as the Global Positioning System (GPS) have been increasingly used in navigation and tracking of vehicles. Using GPS, certain positioning errors and limitations, such as multipath effects and the geometric position of the satellites (DOP) or signal obstructions by high buildings, trees and terrain, have to be considered. Generally travel on road or footpath, map-matching algorithms can be used to correlate the computed system location with a digital map network. Map Matched GPS (MMGPS) is a test-bed simulator for researching algorithms and techniques to reduce the error in position provided by a low cost stand-alone GPS receiver. In order to correctly map-match the GPS positions, a decision about the correct road can be difficult, especially at road junctions, slip roads or almost parallel roads. Investigations into the use of artificial neural networks (ANNs) for reliability and accuracy improvement of map-matched GPS positioning was initiated in previous research [Winter, 2002]. However, there are generally strong interference effects that lead to slow learning and poor generalization when a single ANN is trained to perform different subtasks on different occasions [Jacobs et al., 1991], e.g. correct transport network (TN) segment selection considering different TN geometry. Interference can be reduced by training a system composed of several different "expert" ANNs using a TN geometry indicator to decide which of the experts should be used for each training case. An aim of this research was the design, development and implementation of such a modular neural network (MNN). This work uses a new measure for indicating TN geometry, directly derived from GPS positions in MMGPS. An improvement of more than 50% to traditional map-matching techniques was achieved using the proposed MNN approach, when the correct road could not be uniquely identified by map-matching.

The creation of anatomically correct three-dimensional joints for the simulation of humans is a complex process, a key difficulty being the correction of invalid joint configurations to the nearest valid alternative. Personalised models based on individual joint mobility are in demand in both animation and medicine [1]. Medical models need to be highly accurate animated models less so, however if either are to be used in a real time environment they must have a low temporal cost (high performance). This work briefly explores Support Vector Machine neural networks as joint configuration classifiers that group joint configurations into invalid and valid. A far more detailed investigation is carried out into the use of topologically evolved feed forward neural networks for the generation of appropriately proportioned corrective components which when applied to an invalid joint configuration result in a valid configuration and the same configuration if the original configuration was valid. Discontinuous vector fields were used to represent constraints of varying size, dimensionality and complexity. This culminated in the creation corrective quaternion constraints represented by discontinuous vector fields, learned by topologically evolved neural networks and trained via the resilient back propagation algorithm. Quaternion constraints are difficult to implement and although alternative methods exist [2-6] the method presented here is superior in many respects. This method of joint constraint forms the basis of the contribution to knowledge along with the discovery of relationships between the continuity and distribution of samples in quaternion space and neural network performance. The results of the experiments for constraints on the rotation of limb with regular boundaries show that 3.7 x lO'Vo of patterns resulted in errors greater than 2% of the maximum possible error while for irregular boundaries 0.032% of patterns resulted in errors greater than 7.5%.

The thesis is written in chapter form. Chapter 1 describes some of the history of neural networks and its place in the field of artificial intelligence. It indicates the biological basis from which neural network approximation are made. Chapter 2 describes the properties of neural networks and their uses. It introduces the concepts of training and learning. Chapters 3, 4, 5 and 6 show the perceptron and adaline in feedforward and recurrent networks particular reference is made to regression substitution by "group method data handling. Networks are chosen that explain the application of neural networks in classification, association, optimization and self organization. Chapter 7 addresses the subject of practical inputs to neural networks. Chapter 8 reviews some interesting recent developments. Chapter 9 reviews some ideas on the future technology for neural networks. Chapter 10 gives a listing of some neural network types and their uses. Appendix A gives some of the ideas used in portfolio selection for the Johannesburg Stock Exchange. / Computing / M. Sc. (Operations Research)

A new methodology was developed for flow regime identification in pipes. The method utilizes the pattern recognition abilities of Artificial Neural Networks and the unprocessed time series of a system-monitoring-signal. The methodology was tested with synthetic data from a conceptual system, liquid level indicating capacitance signals from a Horizontal flow system and with a pressure difference signal from a S-shape riser. The results showed that the signals that were generated for the conceptual system had all their patterns identified correctly with no errors whatsoever. The patterns for the Horizontal flow system were also classified very well with a few errors recorded due to original misclassifications of the data. The misclassifications were mainly due to subjectivity and due to signals that belonged to transition regions, hence a single label for them was not adequate. Finally the results for the S-shape riser showed also good agreement with the visual observations and the few errors that were identified were again due to original misclassifications but also to the lack of long enough time series for some flow cases and the availability of less flow cases for some flow regimes than others. In general the methodology proved to be successful and there were a number of advantages identified for this neural network methodology in comparison to other ones and especially the feature extraction methods. These advantages were: Faster identfication of changes to the condition of the system, inexpensive suitable for a variety of pipeline geometries and more powerful on the flow regime identification, even for transitional cases.

Neural networks need to be able to guarantee their intrinsic generalisation abilities if they are to be used reliably. Mitchell's concept and version spaces technique is able to guarantee generalisation in the symbolic concept-learning environment in which it is implemented. Generalisation, according to Mitchell, is guaranteed when there is no alternative concept that is consistent with all the examples presented so far, except the current concept, given the bias of the user. A form of bidirectional convergence is used by Mitchell to recognise when the no-alternative situation has been reached. Mitchell's technique has problems of search and storage feasibility in its symbolic environment. This thesis aims to show that by evolving the technique further in a neural environment, these problems can be overcome. Firstly, the biasing factors which affect the kind of concept that can be learned are explored in a neural network context. Secondly, approaches for abstracting the underlying features of the symbolic technique that enable recognition of the no-alternative situation are discussed. The discussion generates neural techniques for guaranteeing generalisation and culminates in a neural technique which is able to recognise when the best fit neural weight state has been found for a given set of data and topology.

Neurons and networks of the brain may use various strategies of computation to provide the neural substrate for sensation, perception, or cognition. To simplify the scenario, two of the most commonly cited neural codes are firing rate and temporal coding, whereby firing rates are typically measured over a longer duration of time (i.e., seconds or minutes), and temporal codes use shorter time windows (i.e., 1 to 100 ms). However, it is possible that neurons may use other strategies. Here, we highlight three methods of computation that neurons, or networks, of the brain may use to encode and/or decode incoming activity. First, we explain how single neurons of the brain can utilize a neuronal oscillation, specifically by employing a ‘spike-phase’ code wherein responses to stimuli have greater reliability, in turn increasing the ability to discriminate between stimuli. Our focus was to explore the limitations of spike-phase coding, including the assumptions of low firing rates and precise timing of action potentials. Second, we examined the ability of single neurons to track the onset of network bursting activity, namely ‘burst predictors’. In addition, we show that burst predictors were less susceptible to an in vitro model of neuronal stroke (i.e., excitotoxicity). Third, we discuss the possibility of distributed processing with neuronal networks of the brain. Specifically, we show experimental and computational evidence supporting the possibility that the population activity of cortical networks may be useful to downstream classification. Furthermore, we show that when network activity is highly variable across time, there is an increase in the ability to linearly separate the spiking activity of various networks. Overall, we use the results of both experimental and computational methods to highlight three strategies of computation that neurons and networks of the brain may employ.