An Agile Software Development Life Cycle Model for Machine Learning Application Development

Abstract

Software development teams are often hampered when aligning machine learning production with standard software development processes. Iterative experimentation is needed to address the inherent complexities of data collection and preparation, model entanglement, and the technical debt of machine learning. The complexity of this process is compounded due to dependencies on the production environment and real- time data. We propose a unified framework which facilitates the planning, development, and deployment of a machine learning application through parallel processes for software and machine learning engineering. This allows for the risk of both the project and machine learning development to be significantly reduced through continuous integration, evaluation, and production. The framework, named MLASDLC, unifies concepts from standard software development life cycle methodologies (SDLC), development operations (DevOps) and machine learning operations (MLOps) to present a framework for the development of machine learning applications.