Computing and Information Systems - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemDynamic Multi-objective Optimisation Using Evolutionary Algorithms( 2022)Dynamic Multi-objective Optimization Problems (DMOPs) offer an opportunity to examine and solve challenging real world scenarios where trade-off solutions between conflicting objectives change over time. Definition of benchmark problems allows modelling of industry scenarios across transport, power and communications networks, manufacturing and logistics. Recently, significant progress has been made in the variety and complexity of DMOP benchmarks and the incorporation of realistic dynamic characteristics. However, significant gaps still exist in standardised methodology for DMOPs, specific problem domain examples and in the understanding of the impacts and explanations of dynamic characteristics. This thesis provides major contributions on these three topics within evolutionary dynamic multi-objective optimization. Firstly, experimental protocols for DMOPs are varied. This limits the applicability and relevance of results produced and conclusions made in the field. A major source of the inconsistency lies in the parameters used to define specific problem instances being examined. The uninformed selection of these has historically held back understanding of their impacts and standardisation in experimental approach to these parameters in the multi-objective problem domain. Using the frequency and severity (or magnitude) of change events, a more informed approach to DMOP experimentation is conceptualized, implemented and evaluated. Establishment of a baseline performance expectation across a comprehensive range of dynamic instances for well-studied DMOP benchmarks is analyzed. To maximize relevance, these profiles are composed from the performance of evolutionary algorithms commonly used for baseline comparisons and those with simple dynamic responses. Comparison and contrast with the coverage of parameter combinations in the sampled literature highlights the importance of these contributions. Secondly, the provision of useful and realistic DMOPs in the combinatorial domain is limited in previous literature. A novel dynamic benchmark problem is presented by the extension of the Travelling Thief Problem (TTP) to include a variety of realistic and contextually justified dynamic changes. Investigation of problem information exploitation and it's potential application as a dynamic response is a key output of these results and context is provided through comparison to results obtained by adapting existing TTP heuristics. Observation driven iterative development prompted the investigation of multi-population island model strategies, together with improvements in the approaches to accurately describe and compare the performance of algorithm models for DMOPs, a contribution which is applicable beyond the dynamic TTP. Thirdly, the purpose of DMOPs is to reconstruct realistic scenarios, or features from them, to allow for experimentation and development of better optimization algorithms. However, numerous important characteristics from real systems still require implementation and will drive research and development of algorithms and mechanisms to handle these industrially relevant problem classes. The novel challenges associated with these implementations are significant and diverse, even for a simple development such as consideration of DMOPs with multiple time dependencies. Real world systems with dynamics are likely to contain multiple temporally changing aspects, particularly in energy and transport domains. Problems with more than one dynamic problem component allow for asynchronous changes and a differing severity between components that leads to an explosion in the size of the possible dynamic instance space. Both continuous and combinatorial problem domains require structured investigation into the best practices for experimental design, algorithm application and performance measurement, comparison and visualization. Highlighting the challenges, the key requirements for effective progress and recommendations on experimentation are explored here.
-
ItemUnderstanding role of provenance in bioinformatics workflows and enabling interoperable computational analysis sharing( 2018)The automation of computational analyses in data-intensive domains such as genomics through scientific workflows is a widely adopted practice in many fields of research nowadays. Computationally driven data-intensive experiments using workflows enable Automation, Scaling, Adaption and Provenance support (ASAP). Provenance data collection is an essential factor for any computational workflow-centric research to achieve reproducibility, transparency and support trust in the published results. At present capture of provenance information across the plethora of workflow management systems and custom software platforms in the bioinformatics domain is not well supported and as such, there exist numerous challenges associated with the effective sharing, publication, understandability, reproducibility and repeatability of scientific workflows. This thesis focuses on providing a unified, interoperable and systematised view of provenance with specific focus on workflow environments in the bioinformatics domain. We identify and overcome the current disconnect between various workflows systems and their existing provenance representations. Through empirical analysis of complex genomic data analysis workflows using three exemplar workflow systems, we identify implicit assumptions that arise. These assumptions produce an incomplete view of provenance resulting in insufficient details that impact on workflow enactment requirements and ultimately on the reproducibility of the given analysis. We propose a set of recommendations to mitigate against such assumptions and enable workflow systems to document and capture complete provenance information that can subsequently be used for re-enacting workflows in other contexts and potentially using other workflow platforms. Based on this empirical case study and pragmatic analysis of related literature, we define a hierarchical provenance framework offering `Levels of Provenance and Resource Sharing''. Each level of this framework addresses specific provenance recommendations and supports the capture of rich provenance information, with the topmost layer enabling the sharing of comprehensive and executable workflows utilising retrospective provenance. To realise this framework, we leverage community-driven, domain-neutral, platform-independent and open-source standards to implement ``CWLProv'' - a format for the methodical representation of provenance supporting workflow enactment aggregating resources specific to the given enactment and associated workflow configuration settings. We realise CWLProv through the Common Workflow Language (CWL) for workflow definition and utilise Research Objects (ROs) for resource aggregation and PROV-Data Model (PROV-DM) to support the capture of retrospective provenance information as required for subsequent workflow enactments. To demonstrate the applicability of CWLProv, we extend an existing workflow executor (cwltool) to provide a reference implementation that generates metadata and provenance-rich interoperable workflow-centric ROs. This approach aggregates and preserves data and methods needed to support the coherent sharing of computational analyses and experiments. Evaluation of CWLProv using real-life bioinformatics pipelines is demonstrated to highlight the utility of the approach demonstrating the interoperability of workflow analyses and the benefits to research reproducibility more generally.