Requirements Engineering for Machine Learning

Abstract

While there are many well-established Requirements Engineering practices for traditional, deterministic, software systems, the emerging field of Machine Learning introduces new challenges with Requirements Engineering. Current theoretical Software Engineering research has identified many challenges with RE for ML, but there is currently a lack of empirical evidence. Challenges are thought to arise for example because of the uncertain nature of ML, and the dependence on data. Innovative ML development is also highly creative, potentially introducing a trade-off between requirements and creative freedom. Through a case study, based on interviews, observations, documentation, and a combined focus group and questionnaire, this thesis provides insight into what challenges and success factors related to RE for ML that are present in an empirical setting, and compares them to literature in the research field. The thesis further recommends that practitioners in the field use a variation of Goal-Oriented Requirements Engineering, ML-GORE, together with practices aimed at understanding the domain and user, such as use case diagrams and scenario-based requirements elicitation. It is also recommended that the stakeholders are involved in the entire requirements and development process. However, it is suggested that these practices, and their impact on RE for ML, are validated in further research. Finally, the findings confirm that new challenges arise when applying RE to ML development. These challenges are to a great extent in line with previous theoretical research, with two of the major ones being data dependence and outcome uncertainty.