Autonomous and semiautonomous robots are certain to play a major role in several areas in the future, from the battlefield to the household. Countless different methodologies have been applied to solve the problem of mobile robot navigation, with varying degrees of success. SAMUEL, a genetic algorithm based system for evolving semi-autonomous agent behaviors, has proven successful in generating the necessary rule sets for navigating a simple environment. Fuzzy AR TMAP (FAM) neural networks have also been applied in a similar fashion, again with success. In this thesis, a hybrid system is developed. The system fuses both SAMUEL and FAM neural networks, using SAMUEL to develop rule sets for which the FAM provides motion prediction information. The FAM motion predictor serves as an input to the genetic algorithm, so the genetic algorithm can utilize this capability without modification. A simulation using the hybrid system is developed and run, demonstrating how agents controlled by the system would respond to an example mission. The effectiveness of this approach is compared to SAMUEL's ability to complete this task unaided. Finally, open-source source code is made available.

Research linking neural networks and statistics has been at two ends of a spectrum: either highly theoretical or application specific. This research attempts to bridge the gap on the spectrum by exploring the robustness of regression analysis and backpropagation networks in conducting data analysis. Robustness is viewed as the degree to which a technique is insensitive to abnormalities in data sets, such as violations of assumptions. The central focus of regression analysis is the establishment of an equation that describes the relationship between the variables in a data set. This relationship 1s used primarily for the prediction of one variable based on the known values of the other variables. Certain assumptions have to be made regarding the data in order to obtain a tractable solution and the failure of one or more of these assumptions results in poor prediction. The assumptions underlying linear regression that are used to characterize data sets in this research are characterized by: (a) sample size and error variance, (b) outliers, skewness, and kurtosis, (c) multicollinearity, and (d) nonlinearity and underspecification. By using this characterization, the robustness of each technique is studied under what is, in effect, the relaxation of assumptions one at a time. The comparison between regression and backpropagation is made using the root mean square difference between the predicted output from each technique and the actual output. / Ph. D.

This thesis investigates the applicability of backpropagation neural networks to production order dispatching in a dynamic, single machine shop where the achievement of multiple performance measures is desired. There has been relatively little research done in this area so the objectives center around the determination of information and parameters which lead to improved network performance with respect to learning as well as decision making. Results of the research showed that many of the qualities inherent to backpropagation neural networks were compatible with the requirements of the dispatching activity. The networks that were trained and tested had the ability to implicitly map the complex functional relationships between inputs reflecting system status and desired performance and outputs which represented appropriate coefficients used to determine job priority. Once trained they displayed good generalization capabilities when exposed to information they had never been exposed to before. Most importantly, they provided the basis for a complex dispatching procedure which utilized considerable shop floor information to make completely dynamic, real time dispatching decisions. Guidelines and generalizations for similar applications were developed including: input selection and presentation formats, effective training parameters, the effect of using purely dynamic vs. historical data as shop status inputs, the effect of compromising desired performance measure inputs, and the effect of changes in the underlying shop parameters. / M.S.

Nowadays, deep neural networks based software have been widely applied in many areas including safety-critical areas such as traffic control, medical diagnosis and malware detection, etc. However, the software engineering techniques, which are supposed to guarantee the functionality, safety as well as fairness, are not well studied. For example, some serious crashes of DNN based autonomous cars have been reported. These crashes could have been avoided if these DNN based software were well tested. Traditional software testing, debugging or repairing techniques do not work well on DNN based software because there is no control flow, data flow or AST(Abstract Syntax Tree) in deep neural networks. Proposing software engineering techniques targeted on DNN based software are imperative. In this thesis, we first introduced the development of SE(Software Engineering) for AI(Artificial Intelligence) area and how our works have influenced the advancement of this new area. Then we summarized related works and some important concepts in SE for AI area. Finally, we discussed four important works of ours. Our first project DeepTest is one of the first few papers proposing systematic software testing techniques for DNN based software. We proposed neuron coverage guided image synthesis techniques for DNN based autonomous cars and leveraged domain specific metamorphic relation to generate oracle for new generated test cases to automatically test DNN based software. We applied DeepTest to testing three top performing self-driving car models in Udacity self-driving car challenge and our tool has identified thousands of erroneous behaviors that may lead to potential fatal crash. In DeepTest project, we found that the natural variation such as spatial transformations or rain/fog effects have led to problematic corner cases for DNN based self-driving cars. In the follow-up project DeepRobust, we studied per-point robustness of deep neural network under natural variation. We found that for a DNN model, some specific weak points are more likely to cause erroneous outputs than others under natural variation. We proposed a white-box approach and a black-box approach to identify these weak data points. We implemented and evaluated our approaches on 9 DNN based image classifiers and 3 DNN based self-driving car models. Our approaches can successfully detect weak points with good precision and recall for both DNN based image classifiers and self-driving cars. Most of existing works in SE for AI area including our DeepTest and DeepRobust focus on instance-wise errors, which are single inputs that result in a DNN model's erroneous outputs. Different from instance-wise errors, group-level errors reflect a DNN model's weak performance on differentiating among certain classes or inconsistent performance across classes. This type of errors is very concerning since it has been found to be related to many real-world notorious errors without malicious attackers. In our third project DeepInspect, we first introduced the group-level errors for DNN based software and categorized them into confusion errors and bias errors based on real-world reports. Then we proposed neuron coverage based distance metric to detect group-level errors for DNN based software without requiring labels. We applied DeepInspect to testing 8 pretrained DNN models trained in 6 popular image classification datasets, including three adversarial trained models. We showed that DeepInspect can successfully detect group-level violations for both single-label and multi-label classification models with high precision. As a follow-up and more challenging research project, we proposed five WR(weighted regularization) techniques to repair group-level errors for DNN based software. These five different weighted regularization techniques function at different stages of retraining or inference of DNNs including input phase, layer phase, loss phase and output phase. We compared and evaluated these five different WR techniques in both single-label and multi-label classifications including five combinations of four DNN architectures on four datasets. We showed that WR can effectively fix confusion and bias errors and these methods all have their pros, cons and applicable scenario. All our four projects discussed in this thesis have solved important problems in ensuring the functionality, safety as well as fairness for DNN based software and had significant influence in the advancement of SE for AI area.

Thesis submitted in compliance with the requirements for the Master's Degree in Technology: Industrial Engineering, Department of Industrial Engineering, Durban University of Technology, 2009. / This research focuses on the design and implementation of an intelligent machine vision and sorting system that can be used to sort objects in an industrial environment. Machine vision systems used for sorting are either geometry driven or are based on the textural components of an object’s image. The vision system proposed in this research is based on the textural analysis of pixel content and uses an artificial neural network to perform the recognition task. The neural network has been chosen over other methods such as fuzzy logic and support vector machines because of its relative simplicity. A Bluetooth communication link facilitates the communication between the main computer housing the intelligent recognition system and the remote robot control computer located in a plant environment. Digital images of the workpiece are first compressed before the feature vectors are extracted using principal component analysis. The compressed data containing the feature vectors is transmitted via the Bluetooth channel to the remote control computer for recognition by the neural network. The network performs the recognition function and transmits a control signal to the robot control computer which guides the robot arm to place the object in an allocated position. The performance of the proposed intelligent vision and sorting system is tested under different conditions and the most attractive aspect of the design is its simplicity. The ability of the system to remain relatively immune to noise, its capacity to generalize and its fault tolerance when faced with missing data made the neural network an attractive option over fuzzy logic and support vector machines.

Thesis submitted in compliance with the requirements for the Master's Degree of Technology: Electrical Engineering - Light Current, Durban University of Technology, 2009. / This research is devoted to the development of an intelligent image motion detection system based on artificial neural networks (ANN’s). Object motion detection, non-stationary image isolation and extraction, and image filtering is investigated, with the intention of developing a system that will overcome some of the shortcomings associated with the performance of conventional motion detection systems. Motion detection and image extraction finds popular application in medical imagery and engineering based diagnostics systems. Conventional image processing systems utilise Digital Signal Processing (DSP) to perform the non-stationary image motion detection function. Aliasing and filtering are problematic processes in DSP based image processing systems. The proposed ANN motion detection system overcomes some of these shortcomings. The study compares the performance of conventional DSP systems to that of the proposed ANN based system. The excellent noise immunity, ability to generalise and robustness of the ANN system is exploited in the design of the motion detection system. The ANN’s are arranged as ensembles in order to improve the computation time of the proposed motion detection system. A hybrid system comprising DSP and ANN ensembles is also proposed in the study. The hybrid system exploits the positive characteristics of DSP and ANN’s within a single system. The performance of the pure ANN system and the hybrid system is compared to that of DSP systems, using the image’s signal-to-noise ratio and computation times as a basis for comparison.

Infrared (IR) imagery collected by geostationary satellites provides useful information about the dirunal evolution of cloud systems. These IR images can be analyzed to indicate the location of clouds as well as the pattern of cloud top temperatures (Tbs). During the past several decades, a number of different approaches for estimation of rainfall rate (RR) from Tb have been explored and concluded that the Tb-RR relationship is (1) highly nonlinear, and (2) seasonally and regionally dependent. Therefore, to properly model the relationship, the model must be able to: (1) detect and identify a non-linear mapping of the Tb-RR relationship; (2) Incorporate information about various cloud properties extracted from IR image; (3) Use feedback obtained from RR observations to adaptively adjust to seasonal and regional variations; and (4) Effectively and efficiently process large amounts of satellite image data in real -time. In this study, a kind of artificial neural network (ANN), called Modified Counter Propagation Network (MCPN), that incorporates these features, has been developed. The model was calibrated using the data around the Japanese Islands provided by the Global Precipitation Climatology Project (GPCP) First Algorithm Intercomparison Project (AIP-I). Validation results over the Japanese Islands and Florida peninsula show that by providing limited ground-truth observation, the MCPN model is effective in monthly and hourly rainfall estimation. Comparison of results from MCPN model and GOES Precipitation Index (GPI) approach is also provided in the study.

Thesis (MEng)--University of Stellenbosch, 2000. / ENGLISH ABSTRACT: The intent of this thesis is to examine the solving of problems with neural networks. Three cases are investigated: the calculation of a Visual Condition Index (VCI), the determination ofthe reseal need, and the back-calculation of E-moduli from measured deflection basins. The calculation of a Visual Condition Index (VCI) is a very good example of how a neural network can be applied to reach a conclusion through the association of a number of facts with one single outcome. VISual assessments of the road condition are done on a yearly basis and the Assessor gives his impression of the condition of a road. A neural network simulates the association between the inputs of elements of distress on the road and the eventual assessment of the overall condition expressed as the VCI, very well. Reseal need is determined by the Provincial Administration: Western Cape (PAWC) with a Reseal Expert System. Data produced by the expert system was used to train a neural network to determine the reseal need. The strength of using these two methods in combination is shown. Meaningful results could not be obtained due to insufficient data in certain categories. Deflection measurements with a Falling Weight Deflectometer are meaningful indicators of pavement strength. Back-calculation is used to calculate E-moduli of pavement layers which can be used in a mechanistic approach to estimate remaining pavement life from pavement response. Conventional backcalculation programs, when implemented in a pavement management system, result in very long computing times due to the large volumes of data available. Neural networks offer the alternative of very fast processing, making the implementation of back-calculation in real-time possible. It is shown that neural networks can back-calculate E-moduli, but with varying degrees of success. The main problem identified is the basis on which the dataset used to train neural networks, is generated using linear elastic theory. The biggest limitation in the linear elastic theory is that non-linear and stress dependent behaviour of materials cannot be simulated, two aspects that have a major influence on the back-calculated E-moduli. Improvements in the data generation process using a theory that accommodates non-linear and stress dependent behaviour of materials may result in improved performance of the neural networks. It is also shown that it is very difficult to design a single neural network that can be successfully used on all the possible pavement types. It is better to identify representative pavement types and train neural networks for each of these. Neural networks can be applied with success in the pavement management field and the combination of Expert Systems, Neural Networks and Fuzzy Logic can be a very powerful method to solve complicated problems. Care should be taken in the design of the neural networks and a good understanding ofthe data is a prerequisite for success. / AFRIKAANSE OPSOMMING: Die bedoeling met die tesis is om die vermoë van neurale netwerke om probleme op te los, te ondersoek. Drie gevalle word beskou: die berekening van 'n Visuele Toestand Indeks (VTI), die bepaling van die herseël behoefte en die terugberekening van die E-moduli vanaf defleksie metings. Die berekening van die VTI demonstreer die vermoë van neurale netwerke om,deur middel van die assosiasie tussen 'n hele aantal veranderlikes tot 'n enkele uitkoms, tot 'n gevolgtrekking te kom. Visuele opnames van paaie word op 'n jaarlikse basis gedoen waar die opnemer sy indrukke gee van die toestand van die pad. In Neurale netwerk simuleer die assosiasie tussen die insette (waargenome gebreke) en die uiteindelike toestands beskrywing van die pad, uitgedruk as die VTI, baie goed. Die Provinsiale Administrase: Wes-Kaap bepaal die jaarlikse herseëlbehoefte met behulp van 'n Herseël Ekspertstelsel. Die uitsette van hierdie stelsel is gebruik om 'n neurale netwerk op te lei om die herseëlbehoefte te bepaal. Die voordele om die twee stelsels saam aan te wend, word getoon. Betekenisvolle resultate kom nie bekom word nie vanweë onvoldoende inligting in sekere kategorieë. Defleksiemetings deur 'n vallende-gewig meetapparaat is betekenisvolle indikators van die plaveiselsterkte. Die E-moduli van die plaveisellae word bepaal deur terugberekenings vanaf defleksiemetings. Hierdie Emoduli kan gebruik word om met behulp van meganistiese metodes die oorblywende leeftyd van 'n plaveisel te bepaal. Konvensionele terugberekenings programme, geïmplementeer in In plaveiselbestuurstelsel, neem lank om die groot hoeveelheid defleksiemetings te verwerk. Neurale netwerke bied die alternatief van die intydse berekening van E-moduli vanweë die besonder hoë berekeningspoed wat behaal word. In hierdie tesis word aangetoon dat neurale netwerke aangewend kan word om die terugberekenigs te doen, maar met 'n wisselende mate van sukses. Die gebruik van die lineêre elastiese teorie om die data vir die neurale netwerke te genereer, word as 'n probleem geïdentifiseer. Die grootste tekortkoming wat met die lineêre elastiese teorie ondervind word is dat dit nie die nie-lineêre en spanningsafhanklike gedrag van materiale voldoende simuleer nie. Beide hierdie twee aspekte het 'n groot invloed op die akkuraatheid van terugberekende E-moduli. Verbeteringe in die generering van data deur 'n teorie te gebruik wat nie-lineêre en spanningsafhanklike gedrag van materiale behoorlik simuleer, mag lei tot 'n beter prestasie van die neurale netwerke. Dit word ook getoon dat dit moeilik is om 'n enkele neurale netwerk te ontwerp wat suksesvol gebruik kan word op alle plaveiseltipes. Dit is beter om verteenwoordigende plaveiseltipes te identifiseer en dan neurale netwerke vir elkeen te ontwerp. Neurale netwerke kan met sukses in die plaveiselbestuur veld toegepas word en die kombinasie van ekspertsteiseis, neurale netwerke en vaagheidstelsels (fuzzy) kan tot kragtige metodes lei om komplekse probleme op te los. Sorg moet aan die dag gelê word met die ontwerp van neurale netwerke en 'n goeie begrip van die data is 'n voorvereiste vir sukses.

published_or_final_version / Industrial and Manufacturing Systems Engineering / Doctoral / Doctor of Philosophy

published_or_final_version / Mechanical Engineering / Doctoral / Doctor of Philosophy