OpenUCT :: Browsing by Author "Durbach, Ian"

Browsing by Author "Durbach, Ian"

Now showing 1 - 20 of 23

Open Access
Adapting Large-Scale Speaker-Independent Automatic Speech Recognition to Dysarthric Speech
(2022) Houston, Charles; Britz, Stefan S; Durbach, Ian
Despite recent improvements in speaker-independent automatic speech recognition (ASR), the performance of large-scale speech recognition systems is still significantly worse on dysarthric speech than on standard speech. Both the inherent noise of dysarthric speech and the lack of large datasets add to the difficulty of solving this problem. This thesis explores different approaches to improving the performance of Deep Learning ASR systems on dysarthric speech. The primary goal was to find out whether a model trained on thousands of hours of standard speech could successfully be fine-tuned to dysarthric speech. Deep Speech – an open-source Deep Learning based speech recognition system developed by Mozilla – was used as the baseline model. The UASpeech dataset, composed of utterances from 15 speakers with cerebral palsy, was used as the source of dysarthric speech. In addition to investigating fine-tuning, layer freezing, data augmentation and re-initialization were also investigated. Data augmentation took the form of time and frequency masking, while layer freezing consisted of fixing the first three feature extraction layers of Deep Speech during fine-tuning. Re-initialization was achieved by randomly initializing the weights of Deep Speech and training from scratch. A separate encoder-decoder recurrent neural network consisting of far fewer parameters was also trained from scratch. The Deep Speech acoustic model obtained a word error rate (WER) of 141.53% on the UASpeech test set of commands, digits, the radio alphabet, common words, and uncommon words. Once fine-tuned to dysarthric speech, a WER of 70.30% was achieved, thus demonstrating the ability of fine-tuning to improve upon the performance of a model initially trained on standard speech. While fine-tuning lead to a substantial improvement in performance, the benefit of data augmentation was far more subtle, improving on the fine-tuned model by a mere 1.31%. Freezing the first three layers of Deep Speech and fine-tuning the remaining layers was slightly detrimental, increasing the WER by 0.89%. Finally, both re-initialization of Deep Speech's weights and the encoder-decoder model generated highly inaccurate predictions. The best performing model was Deep Speech fine-tuned to augmented dysarthric speech, which achieved a WER of 60.72% with the inclusion of a language model.
Open Access
Agent-based model of the market penetration of a new product
(2014) Magadla, Thandulwazi; Durbach, Ian; Scott, Leanne
This dissertation presents an agent-based model that is used to investigate the market penetration of a new product within a competitive market. The market consists of consumers that belong to social network that serves as a substrate over which consumers exchange positive and negative word-of-mouth communication about the products that they use. Market dynamics are influenced by factors such as product quality; the level of satisfaction that consumers derive from using the products in the market; switching constraints that make it difficult for consumers to switch between products; the word-of-mouth that consumers exchange and the structure of the social network that consumers belong to. Various scenarios are simulated in order to investigate the effect of these factors on the market penetration of a new product. The simulation results suggest that: ■ A new product reaches fewer new consumers and acquires a lower market share when consumers switch less frequently between products. ■ A new product reaches more new consumers and acquires a higher market share when it is of a better quality to that of the existing products because more positive word-of-mouth is disseminated about it. ■ When there are products that have switching constraints in the market, launching a new product with switching constraints results in a higher market share compared to when it is launched without switching constraints. However, it reaches fewer new consumers because switching constraints result in negative word-of-mouth being disseminated about it which deters other consumers from using it. Some factors such as the fussiness of consumers; the shape and size of consumers' social networks; the type of messages that consumers transmit and with whom and how often they communicate about a product, may be beyond the control of marketing managers. However, these factors can potentially be influenced through a marketing strategy that encourages consumers to exchange positive word-of-mouth both with consumers that are familiar with a product and those who are not.
Open Access
An exploration of media repertoires in South Africa: 2002-2014
(2019) Bakker, Hans-Peter; Durbach, Ian
This dissertation explores trends in media engagement in South Africa over a period from 2002 until 2014. It utilises data from the South African Audience Research Foundation’s All Media and Products Surveys. Using factor analysis, six media repertoires are identified and, utilising structural equation modelling, marginal means for various demographic categories by year are estimated. Measurement error is determined with the aid of bootstrapping. These estimates are plotted to provide visual aids in interpreting model parameters. The findings show general declines in engagement with traditional media and growth in internet engagement, but these trends can vary markedly for different demographic groups. The findings also show that for many South Africans traditional media such as television remain dominant.
Open Access
Blow me down: A new perspective on Aloe dichotoma mortality from windthrow
(BioMed Central, 2014-03-18) Jack, Samuel L; Hoffman, Michael T; Rohde, Rick F; Durbach, Ian; Archibald, Margaret
Background: Windthrow, the uprooting of trees during storms associated with strong winds, is a well-established cause of mortality in temperate regions of the world, often with large ecological consequences. However, this phenomenon has received little attention within arid regions and is not well documented in southern Africa. Slow rates of post-disturbance recovery and projected increases in extreme weather events in arid areas mean that windthrow could be more common and have bigger impacts on these ecosystems in the future. This is of concern due to slow rates of post-disturbance recovery in arid systems and projected increases in extreme weather events in these areas. This study investigated the spatial pattern, magnitude and likely causes of windthrown mortality in relation to other forms of mortality in Aloe dichotoma, an iconic arid-adapted arborescent succulent and southern Africa climate change indicator species. Results: We found that windthrown mortality was greatest within the equatorward summer rainfall zone (SRZ) of its distribution (mean = 31%, n = 11), and was derived almost exclusively from the larger adult age class. A logistic modelling exercise indicated that windthrown mortality was strongly associated with greater amounts of warm season (summer) rainfall in the SRZ, higher wind speeds, and leptosols. A statistically significant interaction term between higher summer rainfall and wind speeds further increased the odds of being windthrown. While these results would benefit from improvements in the resolution of wind and substrate data, they do support the hypothesised mechanism for windthrow in A. dichotoma. This involves powerful storm gusts associated with either the current or subsequent rainfall event, heavy convective rainfall, and an associated increase in soil malleability. Shallow rooting depths in gravel-rich soils and an inflexible, top-heavy canopy structure make individuals especially prone to windthrown mortality during storms. Conclusions: Results highlight the importance of this previously unrecognised form of mortality in A. dichotoma, especially since it seems to disproportionately affect reproductively mature adult individuals in an infrequently recruiting species. Smaller, more geographically isolated and adult dominated populations in the summer rainfall zone are likely to be more vulnerable to localised extinction due to windthrow events.
Open Access
Classification and visualisation of text documents using networks
(2018) Phaweni, Thembani; Durbach, Ian; Varughese, Melvin; Bassett, Bruce
In both the areas of text classification and text visualisation graph/network theoretic methods can be applied effectively. For text classification we assessed the effectiveness of graph/network summary statistics to develop weighting schemes and features to improve test accuracy. For text visualisation we developed a framework using established visual cues from the graph visualisation literature to communicate information intuitively. The final output of the visualisation component of the dissertation was a tool that would allow members of the public to produce a visualisation from a text document. We represented a text document as a graph/network. The words were nodes and the edges were created when a pair of words appeared within a pre-specified distance (window) of words from each other. The text document model is a matrix representation of a document collection such that it can be integrated into a machine or statistical learning algorithm. The entries of this matrix can be weighting according to various schemes. We used the graph/network representation of a text document to create features and weighting schemes that could be applied to the text document model. This approach was not well developed for text classification therefore we applied different edge weighting methods, window sizes, weighting schemes and features. We also applied three machine learning algorithms, naïve Bayes, neural networks and support vector machines. We compared our various graph/network approaches to the traditional document model with term frequency inverse-document-frequency. We were interested in establishing whether or not the use of graph weighting schemes and graph features could increase test accuracy for text classification tasks. As far as we can tell from the literature, this is the first attempt to use graph features to weight bag-of-words features for text classification. These methods had been applied to information retrieval (Blanco & Lioma, 2012). It seemed they could also be applied to text classification. The text visualisation field seemed divorced from the text summarisation and information retrieval fields, in that text co-occurrence relationships were not treated with equal importance. Developments in the graph/network visualisation literature could be taken advantage of for the purposes of text visualisation. We created a framework for text visualisation using the graph/network representation of a text document. We used force directed algorithms to visualise the document. We used established visual cues like, colour, size and proximity in space to convey information through the visualisation. We also applied clustering and part-of-speech tagging to allow for filtering and isolating of specific information within the visualised document. We demonstrated this framework with four example texts. We found that total degree, a graph weighting scheme, outperformed term frequency on average. The effect of graph features depended heavily on the machine learning method used: for the problems we considered graph features increased accuracy for SVM classifiers, had little effect for neural networks and decreased accuracy for naïve Bayes classifiers Therefore the impact on test accuracy of adding graph features to the document model is dependent on the machine learning algorithm used. The visualisation of text graphs is able to convey meaningful information regarding the text at a glance through established visual cues. Related words are close together in visual space and often connected by thick edges. Large nodes often represent important words. Modularity clustering is able to extract thematically consistent clusters from text graphs. This allows for the clusters to be isolated and investigated individually to understand specific themes within a document. The use of part-of-speech tagging is effective in both reducing the number of words being displayed but also increasing the relevance of words being displayed. This was made clear through the use of part-of-speech tags applied to the Internal Resistance of Apartheid Wikipedia webpage. The webpage was reduced to its proper nouns which contained much of the important information in the text. Training accuracy is important in text classification which is a task that can often be performed on vast amounts of documents. Much of the research in text classification is aimed at increasing classification accuracy either through feature engineering, or optimising machine learning methods. The finding that total degree outperformed term frequency on average provides an alternative avenue for achieving higher test accuracy. The finding that the addition of graph features can increase test accuracy when matched with the right machine learning algorithm suggests some new research should be conducted regarding the role that graph features can have in text classification. Text visualisation is used as an exploratory tool and as a means of quickly and easily conveying text information. The framework we developed is able to create automated text visualisations that intuitively convey information for short and long text documents. This can greatly reduce the amount of time it takes to assess the content of a document which can increase general access to information.
Open Access
Counting animals in ecological images
(2022) Pillay, Nakkita; Durbach, Ian; Dufourq, Emmanuel
In the field of ecology, counting of animals to estimate population size and prey abundance is important for the conservation of wildlife. This involves analyzing large volumes of image, video or audio data and manual counting. Automating the process of counting animals would be invaluable to researchers as it will eliminate the tedious time-consuming task of counting. The purpose of this dissertation is to address manual counting in images by implementing an automated solution using computer vision. This research applies a blob detection algorithm primarily based on the determinant of the Hessian matrix to estimate counts of animals in aerial images of colonies in a user-friendly web application and trains an object detection model using deep convolutional neural networks to automatically identify and count penguin prey in 2053 images extracted from animal-borne videos. The blob detection algorithm reports an average relative bias of less than 6% and the YOLOv3 object detection model automatically detects jellyfish, school of fish and fish with a mean average precision of 82,53% and counts with an average relative bias of -17,66% over all classes. The results show that applying traditional computer vision methods and deep learning on data-scarce and data-rich situations respectively, can save ecologists an immense amount of time used on manual tedious methods of analysis and counting. Additionally, these automated counting methods can contribute towards improving wildlife conservation and future studies.
Open Access
Decision support for the production and distribution of electricity under load shedding
(2016) Rakotonirainy, Rosephine Georgina; Durbach, Ian; Nyirenda, Juwa
Every day national power system networks provide thousands of MW of electric power from generating units to consumers. This process requires different operations and planning to ensure the security of the entire system. Part of the daily or weekly operation system is the so called Unit Commitment problem which consists of scheduling the available resources in order to meet the system demand. But the continuous growth in electricity demand might put pressure on the ability of the generation system to sufficiently provide supply. In such case load shedding (a controlled, enforced reduction in electricity supply) is necessary to prevent the risk to system collapse. In South Africa at the present time, a systematic lack of supply has meant that regular load shedding has taken place, with substantial economic and social costs. In this research project we study two optimization problems related to load shedding. The first is how load shedding can be integrated into the unit commitment problem. The second is how load shedding can be fairly and efficiently allocated across areas. We develop deterministic and stochastic linear and goal programming models for these purposes. Several case studies are conducted to explore the possible solutions that the proposed models can offer.
Open Access
Deep neural networks for video classification in ecology
(University of Cape Town, 2020) Conway, Alexander; Durbach, Ian
Analyzing large volumes of video data is a challenging and time-consuming task. Automating this process would very valuable, especially in ecological research where massive amounts of video can be used to unlock new avenues of ecological research into the behaviour of animals in their environments. Deep Neural Networks, particularly Deep Convolutional Neural Networks, are a powerful class of models for computer vision. When combined with Recurrent Neural Networks, Deep Convolutional models can be applied to video for frame level video classification. This research studies two datasets: penguins and seals. The purpose of the research is to compare the performance of image-only CNNs, which treat each frame of a video independently, against a combined CNN-RNN approach; and to assess whether incorporating the motion information in the temporal aspect of video improves the accuracy of classifications in these two datasets. Video and image-only models offer similar out-of-sample performance on the simpler seals dataset but the video model led to moderate performance improvements on the more complex penguin action recognition dataset.
Open Access
Detection and Isolation of Prey Capture Events in Animal-Borne Images
(2021) Chirwa, Temweka S; Durbach, Ian; Dufourq, Emmanuel
Understanding the foraging habits and prey availability for a species is crucial. Prey availability is crucial to a species' survival and sustainability of the food pyramid. Identifying the type of prey consumed also allows ecologists to determine the energy received, while the duration and extent of foraging bouts provide information about the energy expended. With recent advancements in technology, data collection has become more accessible, and animal-borne video cameras are an increasingly popular mechanism for collecting information about foraging and other behaviour. Video recorders collect large volumes of data but create a bottleneck as data processing is still predominantly done manually. This process is time-consuming and costly, even with the assistance of crowdsourced tasks. Advancements in deep learning, and its applications to computer vision, provide opportunities to apply these tools to ecological problems, such as the processing of data from animal-borne video recorders. Speeding up the annotation process allows more time to be spent focused on the ecological research questions. This dissertation aims to develop detection and isolation models that will assist in the processing of visual data, namely images from animal-borne videos. The first model used for detection will perform an image classification determining whether prey is present or not. Images found to have prey present will then be presented to the second model for isolation that identifies exactly where within the image the prey is and labels the type of prey. The models were trained on video data of little penguins (Eudyptula minor ), whose main prey in this investigation are small fish, predominantly anchovies, and jellyfish. The image classification model based on the ResNet architecture achieved 85% accuracy with precision and recall values of 0.85 and 0.85 respectively on its test set. The object detection model based on the You Only Look Once (YOLO) framework achieved a mean average precision of 60% on its test set. However, the models did not perform well enough on unseen full length videos to be used without human supervision or to serve as alternatives to manual labelling. Rather, the models can be used to guide researchers to areas that may contain prey events.
Open Access
Machine learning methods for individual acoustic recognition in a species of field cricket
(2018) Dlamini, Gciniwe; Durbach, Ian
Crickets, like other insects, play a vital role in maintaining a balance in the ecosystem. Therefore, the ability to identify individual crickets is crucial as it enables ecologists to estimate important population metrics such as population densities, which in turn are used to investigate ecological questions pertaining to these insects. In this research, classification models were developed to recognise individual field crickets of the species Plebeiogryllus guttiventris based solely on the audio recordings of their calls. Recent advances in technology have made data collection easier, and consequently, large volumes of data, including acoustic data, have become available to ecologists. The task of acoustic animal identifications thus requires the utilisation of models that are well suited for training large datasets. It is for this very reason that convolutional neural networks (CNN) and recurrent neural networks (RNN) were utilised in this research. The results of these models were compared to results of a baseline random forest (RF) model as RFs can also be used to make acoustic classifications. Mel-frequency cepstral coefficients (MFCC), raw acoustic samples as well as two temporal features were extracted from each chirp in the cricket recordings and used as inputs to train the machine learning models. The raw acoustic samples were only used in the deep neural network (DNN) models (CNNs and RNNs) as these models have been successful in training other raw forms of data such as images (for example, Krizhevsky et al. (2012)). Training on the MFCC features was conducted in two ways: the DNN models were trained on MFCC matrices that each spanned a chirp, whereas the RF models were trained on the MFCC frame vectors. This is because RF are only able to train on vector representations of observations, not matrices. The frame-level MFCC predictions obtained from the RF model were then aggregated into chirp-level predictions to facilitate the comparison with the other classification models. The best classification performance was achieved by the RF model trained on the MFCC features with a score of 99.67%. The worst performance was observed from the RF model trained upon the temporal features, which scored 67%. The DNN models attained on average 98.6% classification accuracies when trained on both MFCC features and the raw acoustic samples. These results show that individual recognition of the crickets using acoustics can be achieved with great success through the use of machine learning. Moreover, the performance of the deep learning models when trained upon the raw acoustic samples indicate that the feature (MFCC) extraction step can be bypassed; the deep learning machine algorithms can be trained directly on the raw acoustic data and still achieve great results.
Open Access
Measuring the rebound effect of energy efficiency initiatives for the future: A South African case study
(2011) Davis, Stephen; Cohen, Brett; Hughes, Alison; Durbach, Ian; Nyatsanza, Kudakwashe
The rebound effect is a phrase which was originally defined to refer to the extent to which energy efficiency improvements are lost due to subsequent behavioural changes. This report documents almost three years of research work that set out to quantify the rebound effect of energy efficiency initiatives in South Africa’s residential sector, and to explore ways of mitigating that effect using awareness and education. Society is in an era where energy commodities are characterised by constrained supply, increasing demand, and higher prices, and where the harmful social and environmental externalities resulting from the conversion of primary into useful energy can no longer be ignored. Part of the solution to the sustainable energy provision and consumption challenge has focused on the technology devices used to convert primary and secondary energy to useful energy that can be used for lighting, water heating, space heating and cooling (and a host of other end-uses). Given that all energy demand can ultimately be traced to the energy required for survival, a study of the residential sector is the natural place to begin.
Open Access
Mining a large shopping database to predict where, when, and what consumers will buy next
(2020) Halam, Bantu; Durbach, Ian
Retailers with electronic point-of-sale systems continuously amass detailed data about the items each consumer buys (i.e. what item, how often, its package size, how many were bought, whether the item was on special, etc.). Where the retailer can also associate purchases with a particular individual for example, when an account or loyalty card is issued, the buying behaviour of the consumer can be tracked over time, providing the retailer with valuable information about a customer's changing preferences. This project is based on mining a large database, containing the purchase histories of some 300 000 customers of a retailer, for insights into the behaviour of those customers. Specifically, the aim is to build three predictive models, each forming a chapter of the dissertation; forecasting the number of daily customers to visit a store, detecting changes in consumers' inter-purchase times, and predicting repeat customers after being given a special offer. Having too many goods and not enough customers implies loss for a business; having too few goods implies a lost opportunity to turn a profit. The ideal situation is to stock the appropriate number of goods for the number of customers arriving, so you can minimize loss, and maximize profit. To attend to this problem, in the first chapter we forecast the number of customers that will visit a store each day to buy any product (i.e. store daily visits). In the process we also carry out a time-series forecasting methods comparison, with the main aim of comparing machine learning methods to classical statistical methods. The models are fitted into a univariate time-series data and the best model for this particular dataset is selected using three accuracy measures. The results showed that there was not much difference between the methods, but some classical methods slightly performed better than the machine learning algorithms, and this was consistent with outcomes obtained by Makridakis et al. (2018) on similar comparisons. It is also vital for retailers to know when there has been a change in their consumers purchase behaviour. This change can either be the time between purchases, change in brand selection or change in market share. It is critical for such changes to be detected as early as possible, as speedy detection can help managers act before incurring loses. In the second chapter, we use change-point models to detect changes in consumers' inter-purchase times. Change-point models are approaches that offer a flexible, general-purpose solution to the problem of detecting changes in customer historic behaviour. This multiple change-point model assumes that there is a sequence of underlying parameters, and that this sequence is partitioned into contiguous blocks. These partitions are such that the parameter values are equal within, and different between blocks, whereby a beginning of a block is considered to be a change point. This changepoint model is fitted to consumers inter-purchase times (i.e. we model time between purchases) to see whether there were any significant changes on the consumers buying behaviour over a one year purchase period. The results showed that, depending on the length of the sequences, minority to a handful of customers do experience changes in their purchasing behaviours, with the longer sequences having more changes than the shorter ones. The results seemed to be different to those obtained by Clark and Durbach (2014), but analysing a portion of sequences of same lengths as those analysed in Clark and Durbach (2014), lead to similar results. Increasing sales growth is also vital for retailers, and there are various possible ways in which this can be achieved. One of the strategies is what is referred to as up-selling (whereby a customer is persuaded to make an additional purchase of the same product or purchase a more expensive version of the product.) and cross-selling (whereby a retailer sells a different product or service to an existing customer). These involve campaigning to customers and sell certain products, and sometimes include incentives in the campaign with the aim of exposing customers to these products hoping they will become repeat customers afterwards. In Chapter 3 we build a model to predict which customers are likely to become repeat customers after being given a special offer. This model is fitted to customers' time between two purchases, which makes the input time-series data, and is sequential in nature. Therefore, we build models that provide a good way for dealing with sequential inputs (i.e. convolutional neural networks and recurrent neural networks), and compare them to models that do not take into account the sequence of the data (i.e. feedforward neural networks and decision trees). The results showed that, inter-purchase times are only useful when they are about the same product, as models did no better than random if inter-purchase times were from a different product in the same department. Secondly, it is useful to take the order of the sequence into account, as models that do this do better than those who do not, with the latter not doing any better than a null model. Lastly, while none of the models performed well, deep learning models perform better than standard classification models and produce some substantial lift.
Open Access
Modelling household responses to energy efficiency interventions via system dynamics and survey data
(Stellenbosch University, 2010) Davis, Stephen; Durbach, Ian
An application of building a system dynamics model of the way households might respond to interventions aimed at reducing energy consumption (specifically the use of electricity) is described in this paper. A literature review of past research is used to build an initial integrated model of household consumption, and this model is used to generate a small number of research hypotheses about how households possessing different characteristics might react to various types of interventions. These hypotheses are tested using data gathered from an efficiency intervention conducted in a town in the South African Western Cape in which households were able to exchange regular light bulbs for more efficient compact fluorescent lamp light bulbs. Our experiences are (a) that a system dynamics approach proved useful in advancing a non-traditional point of view for which, for historical and economic reasons, data were not abundantly available; (b) that, in areas where traditional models are heavily quantitative, some scepticism to a system dynamics model may be expected; and (c) that a statistical comparison of model results by means of empirical data may be an effective tool in reducing such scepticism.
Open Access
Modelling the range-wide density patterns of the Arthroleptella lightfooti using acoustic monitoring data
(2019) Poongavanan, Jenicca; Altwegg, Res; Durbach, Ian; Measev, John
Species distributions are often limited by environmental factors and according to the abundant—centre hypothesis, abundance should be highest Where the environment is most favourable for the species. So, do the same environmental factors determine occurrence and abundance patterns inside the range? I examined this question using Arthroleptella lightfooti, a species of frog from the family of Pyxicephalidae, endemic to the mountains of the Cape peninsula. South Africa. I used density estimates obtained from acoustic Spatially Explicit Capture Recapture (aSCR) methods and data from an acoustic survey using an array of 6 microphones to construct the ﬁrst Peninsula wide population-density surface for this visually cryptic but acoustically active species. The analysis consisted of three stages. The first involved creating two sets of data from the original: one shows whether the species is present or not and the other indicates the density when the species is present. The second stage consisted of fitting a Hurdle Model to the data where the presence data is modelled using logistic regression and the density data is separately modelled using ordinary linear regression. The third stage involved combining the two models to estimate the expected density of the species. Conﬁdence intervals were built using non-parametric bootstrapping. It was found that covariates explaining variation in occurrence were not the same as those explaining variation in density, suggesting that processes determining occurrence were not always those determining density. Of the environmental conditions examined, although predictive of occurrence, were generally poor predictors of A. lightfooti density. Presence of the Lightfoot’s moss frog was largely explained by topographic features and availability of water. In contrast. predictions of density were only weakly related to these same environmental factors and in some cases contradicting one another. The second part of this study produces the first Peninsula wide population density surface of A. 11'ghtfo0t1'. At the same time, it assesses the ability of using opportunistically collected presence-only records in combination with the higher quality density data to improve the estimation of expected population-density surface of A. Iightfooti. The presence-only records were constructed into a habitat suitability map using an ensemble of species distribution models. The habitat suitability map was then integrated in the modelling framework as a covariate in order to improve the estimation of expected population—(lensity surface of A. liglitfooti. However, the habitat suitability covariate resulted as being uninformative.
Open Access
Natural Language Processing on Data Warehouses
(2019) Maree, Stiaan; Durbach, Ian
The main problem addressed in this research was to use natural language to query data in a data warehouse. To this effect, two natural language processing models were developed and compared on a classic star-schema sales data warehouse with sales facts and date, location and item dimensions. Utterances are queries that people make with natural language, for example, What is the sales value for mountain bikes in Georgia for 1 July 2005?" The first model, the heuristics model, implemented an algorithm that steps through the sequence of utterance words and matches the longest number of consecutive words at the highest grain of the hierarchy. In contrast, the embedding model implemented the word2vec algorithm to create different kinds of vectors from the data warehouse. These vectors are aggregated and then the cosine similarity between vectors was used to identify concepts in the utterances that can be converted to a programming language. To understand question style, a survey was set up which then helped shape random utterances created to use for the evaluation of both methods. The first key insight and main premise for the embedding model to work is a three-step process of creating three types of vectors. The first step is to train vectors (word vectors) for each individual word in the data warehouse; this is called word embeddings. For instance, the word `bike' will have a vector. The next step is when the word vectors are averaged for each unique column value (column vectors) in the data warehouse, thus leaving an entry like `mountain bike' with one vector which is the average of the vectors for `mountain' and `bike'. Lastly, the utterance by the user is averaged (utterance vectors) by using the word vectors created in step one, and then, by using cosine similarity, the utterance vector is matched to the closest column vectors in order to identify data warehouse concepts in the utterance. The second key insight was to train word vectors firstly for location, then separately for item - in other words, per dimension (one set for location, and one set for item). Removing stop words was the third key insight, and the last key insight was to use Global Vectors to instantiate the training of the word vectors. The results of the evaluation of the models indicated that the embedding model was ten times faster than the heuristics model. In terms of accuracy, the embedding algorithm (95.6% accurate) also outperformed the heuristics model (70.1% accurate). The practical application of the research is that these models can be used as a component in a chatbot on data warehouses. Combined with a Structured Query Language query generation component, and building Application Programming Interfaces on top of it, this facilitates the quick and easy distribution of data; no knowledge of a programming language such as Structured Query Language is needed to query the data.
Open Access
Optimisation of complex simulation models
(2013) Bezuidenhoudt,Cecile Margaret; Durbach, Ian; Stewart, Theodor
Computer simulation models are widely and frequently used to model real systems to predict output responses under specified input conditions. Choosing optimal simulation parameters leads to improved operation of the model but it is still a challenge as to how to go about optimally selecting these parameter values. The aim of this thesis was to see if a method could be found to optimise a simulation model provided by a client. This thesis provides a review of the literature of various simulation optimisation techniques that exist. Five of these simulation optimisation techniques - Simulated Annealing, Genetic Algorithms, Nested Partitions, Ordinal Optimisation and the Nelson-Matejcik Method - were selected and applied to a test case stochastic simulation model to gain an understanding into the techniques for their use in optimising the test model. These techniques were then used and applied to optimise a real life simulation model provided by a client. A technique combining the Ordinal Optimisation and Simulated Annealing optimisation methods provided the best results. This technique was provided to the client as a strategy to implement into their simulation model.
Open Access
Player performance evaluation in rugby using Stochastic Multi-Criteria Acceptability Analysis with simplified uncertainty formats
(2013) Calder, Jon Matthew; Durbach, Ian
This dissertation considers problem contexts in which decision makers are unable or unwilling to assess trade-off information precisely. The primary aim is to investigate the use of Stochastic Multi-criteria Acceptability Analysis (SMAA) with simplified representations of uncertainty as a decision support tool for prescriptive modelling in 'low involvement' contexts such as these. A simulation experiment is used to assess (a) how closely a rank order of alternatives based on partial information and SMAA can approximate results obtained using full-information multi-attribute utility theory (MAUT), (b) whether a number of 'simplified' SMAA models which make use of summarised measures of uncertainty instead of a full probability distribution might also be suitable in certain contexts, and (c) which characteristics of the decision problem influence the accuracy of this approximation.
Open Access
Resource constraints in an epidemic: a goal programming and mathematical modelling framework for optimal resource shifting in South Africa
(2021) Mayet, Saadiyah; Silal, Sheetal; Durbach, Ian
The COVID-19 pandemic has had devastating consequences across the globe, and has led many governments into completely new decision making territory. Developing models which are capable of producing realistic projections of disease spread under extreme uncertainty has been paramount for supporting decision making by many levels of government. In South Africa, this role has been fulfilled by the South African COVID-19 Modelling Consortium's generalised Susceptible-ExposedInfectious-Removed compartmental model, known as the National COVID-19 Epi Model. This thesis adapted and contributed to the Model; its primary contribution has been to incorporate the feature that resources available to the health system are limited. Building capacity constraints into the Model allowed it to be used in the resource-scarce context of a pandemic. This thesis further designed and implemented a goal programming framework to shift ICU beds between districts intra-provincially in a way that aimed to minimise deaths caused by the non-availability of ICU beds. The results showed a 15% to 99% decrease in lives lost when ICU beds were shifted, depending on the scenario considered. Although there are limitations to the scope and assumptions of this thesis, it demonstrates that it is possible to combine mathematical modelling with optimisation in a way that may save lives through optimal resource allocation.
Open Access
Simplified approaches for portfolio decision analysis
(2022) Kantu, Dieudonne Kabongo; Durbach, Ian
Traditional choice decisions involve selecting a single, best alternative from a larger set of potential options. In contrast, portfolio decisions involve selecting the best subset of alternatives — alternatives that together maximize some measure of value to the decision maker and are within their available resources to implement. Examples include capital investment, R&D project selection, and maintenance planning. Portfolio decisions involve a combinatorial aspect that makes them more theoretically and computationally challenging than choice problems, particularly when there are interactions between alternatives. Several portfolio decision analysis methods have been developed over the years and an increasing interest has been noted in the field of portfolio decision analysis. These methods are typically called “exact” methods, but can also be called prescriptive methods. These are generally computationally-intensive algorithms that require substantial amounts of information from the decision maker, and in return yield portfolios that are provably optimal or optimal within certain bounds. These methods have proved popular for choice decisions — for example, those based on multiattribute value or utility theory. But whereas information and computational requirements for choice problems are probably manageable for the majority of diligent decision makers, it is much less clear that this is true of portfolio decisions. That is, for portfolio decisions it may be more common that decision makers do not have the time, expertise and ability to exert the effort to assess all the information required of an exact method. Heuristics are simple, psychologically plausible rules for decision making that limit the amount of information required and the computation effort needed to turn this information into decisions. Previous work has shown that people often use heuristics when confronted with traditional choice problems in unfacilitated contexts, and that these can often return good results, in the sense of selecting alternatives that are also ranked highly by exact methods. This suggests that heuristics may also be useful for portfolio decisions. Moreover, while the lower information demands made by choice problems mean that heuristics have not generally been seen as prescriptive options, the more substantial demands made by portfolio decisions make a priori case for considering their use not just descriptively, but as tools for decision aid. Very little work exists on the use of heuristics for portfolio decision making, the subject of this thesis. Durbach et al. (2020) proposed a family of portfolio selection heuristics known collectively as add-the-best. These construct portfolios by adding, at every step, the alternative that is best in a greedy sense, with different definitions of what “best” is. This thesis extends knowledge on portfolio heuristics in three main respects. Firstly, we show that people use certain of the add-the-best heuristics when selecting portfolios without facilitation, in a context where there are interactions between alternatives. We run an experiment involving actual portfolio decision making behaviour, administered to participants who had the opportunity to choose as many alternatives as they wanted, but under the constraint of a limited budget. This experiment, parts of which were reported in Durbach et al. (2020), provides the first demonstration of the use of heuristics in portfolio selections. Secondly, we use a simulation experiment to test the performance of the heuristics in two novel environments: those involving multiple criteria, and those in which interactions between projects may be positive (the value of selecting two alternatives is more than the sum of their individual values) or negative (the opposite). This extends the results in Durbach et al. (2020), who considered only environments involving a single criterion and positive interactions between alternatives. In doing so we differentiate between heuristics that guide the selection of alternatives, called selection heuristics, and heuristics for aggregating performance across criteria, which we call scoring heuristics. We combine various selection and scoring heuristics and test their performance on a range of simulated decision problems. We found that certain portfolio heuristics continued to perform well in the presence of negative interactions and multiple criteria, and that performance depended more on the approach used to build portfolios (selection heuristics) than on the method of aggregation across criteria (scoring heuristics). We also found that in these extended conditions heuristics continued to provide outcomes that were competitive with optimal models, but that heuristics that ignored interactions led to potentially poor results. Finally, we complement behavioral and simulation experimental studies with an application of both exact methods and portfolio heuristics in a real-world portfolio decision problem involving the selection of the best subset of research proposals out of a pool of proposals submitted by researchers applying for grants from a research institution. We provide a decision support system to this institution in the form of a web-based application to assist with portfolio decisions involving interactions. The decision support system implements exact methods, namely the linear-additive portfolio value model and the robust portfolio model, as well as two portfolio heuristics found to perform well in simulations.
Open Access
Species distribution modelling of Aloidendron dichotomum (quiver tree)
(2018) Dube, Qobo; Durbach, Ian
A variety of species distribution models (SDMs) were fit to data collected by a 15,000km road-side visual survey of Aloidendron dichotomum populations in the Northern Cape region of South Africa, and Namibia. We fit traditional presence/absence SDMs as well as SDMs on how proportions are distributed across three species stage classes (juvenile, adult, dead). Using five candidate machine learning methods and an ensemble model, we compared a number of approaches, including the role of balanced class (presence/absence) datasets in species distribution modelling. Secondary to this was whether or not the addition of species’ absences, generated where the species is known not to exist have an impact on findings. The goal of the analysis was to map the distribution of Aloidendron dichotomum under different scenarios. Precipitation-based variables were generally more deterministic of species presence or lack thereof. Visual interpretation of the estimated Aloidendron dichotomum population under current climate conditions, suggested a reasonably well fit model, having a large overlap with the sampled area. There however were some conditions estimated to be suitable for species incidence outside of the sampled range, where Aloidendron dichotomum are not known to occur. Habitat suitability for juvenile individuals was largely decreasing in concentration towards Windhoek. The largest proportion of dead individuals was estimated to be on the northern edge of the Riemvasmaak Conservancy, along the South African/Namibian boarder, reaching up to a 60% composition of the population. The adult stage class maintained overall proportional dominance. Under future climate scenarios, despite maintaining a bulk of the currently habitable conditions, a noticeable negative shift in habitat suitability for the species was observed. A temporal analysis of Aloidendron dichotomum’s latitudinal and longitudinal range revealed a potential south-easterly shift in suitable species conditions. Results were however met with some uncertainty as SDMs were uncovered to be extrapolating into a substantial amount of the study area. We found that balancing response class frequencies within the data proved not to be an effective error reduction technique overall, having no considerable impact on species detection accuracy. Balancing the classes however did improve the accuracy on the presence class, at the cost of accuracy of the observed absence class. Furthermore, overall model accuracy increased as more absences from outside the study area were added, only because these generated absences were predicted well. The resulting models had lower estimated suitability outside of the survey area and noticeably different suitability distributions within the survey area. This made the addition of the generated absences undesirable. Results highlighted the potential vulnerability of Aloidendron dichotomum given the pessimistic, yet likely future climate scenarios.