The Feasibility of Using Reccurent Neural Networks as a Tool for Improving the Scrum Sprint Planning Process
О. О. Гудь, Nataliia E. Kunanets · Vìsnik Nacìonalʹnogo unìversitetu Lʹvìvsʹka polìtehnìka Serìâ Ìnformacìjnì sistemi ta merežì · 2024
The study substantiates the feasibility of using machine learning technology to improve the iteration planning process in IT projects implemented using the Scrum methodology. The problem of productivity planning in teams is set. The subject and object of the research are formulated. The expected scientific novelty and practical significance of the research results are described. A range of potential issues related to task planning in IT projects, particularly the accuracy of team productivity forecasting, is considered. Key factors influencing the planning process are identified, and possible solutions are analyzed. The success of applying machine learning technologies in project management is analyzed. An evaluation of machine learning technologies for forecasting the implementation of tasks in Scrum project iterations is conducted. The focus is on the use of recurrent neural networks in these processes. The Long Short-Term Memory (LSTM) model is selected for predicting the productivity of IT project teams. The goals, objectives, and tasks of the research are formulated. Historical project performance data is collected and analyzed. The performance of the developed model is analyzed depending on the specified parameters and input data format. Three model variants with different numbers of complete algorithm training cycles are proposed. The data is normalized to optimize the model. It is established that Long Short-Term Memory models can accurately predict future performance based on normalized historical data from previous sprints. The prediction results are analyzed. Ways to further improve the model are identified. The feasibility of using the recurrent neural network approach in sprint planning is proven. Methods of using recurrent neural networks for IT project task planning are proposed. The limitations of this approach are identified. An alternative option for applying recurrent neural networks in case of non-compliance with the limitations is proposed. Prospects for further research are outlined. Conclusions were drawn regarding the course and results of the conducted research.