Repository logo

Doctoral Degrees (Statistics)

Permanent URI for this collectionhttps://hdl.handle.net/10413/7126

Browse

Browsing Doctoral Degrees (Statistics) by Author "Bodhlyera, Oliver."

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Integrating artificial neural networks, simulation and optimisation techniques in improving public emergency ambulance preparedness for heterogeneous regions under stochastic environments.
    (2021) Mapuwei, Tichaona Wilbert.; Bodhlyera, Oliver.; Mwambi, Henry Godwell.
    The Bulawayo Emergency Medical Services (BEMS) department continues to rely on judgemental methods with limited use of historical data for future predictions, strategic, tactical and operational level decision making. The rural to urban migration trend has seen the sprouting of new residential areas, and this has put pressure to the limited health, housing and education resources. It is expected that as population increases, there is subsequent increase in demand for public emergency services. However, public emergency ambulance demand trends has been decreasing in Bulawayo over the years. This trend is a sign of limited capacity of the service rather than demand itself. The situation demanded for consolidated efforts across all sectors including research, to restore confidence among residents, reduce health risk and loss of lives. The key objective was to develop a framework that would assist in integrating forecasting, simulation and optimisation techniques for ambulance deployment to predefined locations with heterogeneous demand patterns under stochastic environments, using multiple performance indicators. Secondary data from the Bulawayo Municipality archives from 2010 to 2018 was used for model building and validation. A combination of methods based on mathematics, statistics, operations research and computer science were used for data analysis, model building, sensitivity analysis and numerical experiments. Results indicate that feed forward neural network (FFNN) models are superior to traditional SARIMA models in predicting ambulance demand, over a short-term forecasting horizon. The FFNN model is more inclined to value estimation as compared to SARIMA model, which is directional as depicted by the linear pattern over time. An ANN model with a 7-(4)-1 architecture was selected to forecast 2019 public emergency ambulance demand (PEAD). Peak PEAD is expected in January, March, September and December whilst lower demand is expected for April, June and July 2019. Simulation models developed mimicked the prevailing levels of service for BEMS with six(6) operational ambulances. However. the average response times were well above 15 minutes, with significantly high average queuing times and number of ambulances queuing for service. These performance outcomes were highly undesirable as they pose a great threat to human based outcomes of safety and satisfaction with regards to service delivery. Optimisation for simulation was conducted by simultaneously minimising the average response time and average queuing time, while maximising throughput ratios. Increasing the number of ambulances influenced the average response time below a certain threshold, beyond this threshold, the average response time remained constant rather than decreasing gradually. Ambulance utilisation inversely varied to increase in the feet size. Numerical experiments revealed that reducing the response time results in the reduction in number of ambulances required for optimal ambulance deployment. It is imperative to simultaneously consider multiple performance indicators in ambulance deployment as it balances resource allocation and capacity utilisation, while avoiding idleness of essential equipment and human resources. Management should lobby for de-congestion and resurfacing of old and dilapidated roads to increase access and speed when responding to emergency calls. Future research should investigate the influence of varying service time on optimum deployment plans and consider operational costs, wages and other budgetary constraints that influence the allocation of critical but scarce resources such as personnel, equipment and emergency ambulance response vehicles.
  • Thumbnail Image
    Item
    A statistical analysis of dissolving timber pulp properties using linear mixed models.
    (2017) Bodhlyera, Oliver.; Zewotir, Temesgen Tenaw.; Ramroop, Shaun.
    The main focus of the study was to understand the behaviour of seven timber genotypes based on seven chemical properties observed during the chemical pulping process with the prime objective of developing methods of grouping different timber genotypes into compatible groups of timber that can be optimally processed together. Four related statistical methods were used in analysing the data and each had a specific objective. The random coefficients model was used to investigate how the genotypes evolve over the processing stages and it was discovered that the rates of change of the chemical properties studied depended on their initial readings at the beginning of processing. This trend applied for all seven genotypes of pulping trees studied. The important results that came out of fitting the random coefficient model to the data is that the higher the raw stage readings (initial values) the higher the rates of change in the chemical properties over the processing stages. The changes were either increases or decreases in the chemical property studied. The random coefficient model was also used to suggest a rudimental mixing index for the different genotypes based on the average ranking of their slope parameters (rates of change) for the seven variables studied. It was found, for example, that the genotypes GUA and GUW are the least mixable ones. Piecewise linear regression models were used to identify important variables when classifying genotypes and it was generally found that viscosity is not a very useful variable in the classification of genotypes. Using piecewise linear regression models together with kernel density estimation a mixing index (scale) was developed that can be used to determine which genotypes are the most mixable for chemical processing. A coparison of the random coefficient and the piecewise linear regression models shows that the two models yielded very similar conclusions on what genotypes are most mixable during processing. Joint modelling was used to analysis the correlations between evolutions of different chemical properties studied. The various levels of correlations between these variables were discussed. The main limitation of the joint modelling method was its computational challenges because of the many parameters that need to be estimated at the same time.