Browsing by Author "Koskinen, Jan"

Machine Learning Operations (MLOps) emerged as a practice for applying DevOps practices and culture for machine learning (ML) systems to increase the speed and reliability of deployments. These practices include advocating for automation and monitoring at all steps of the ML system construction, including integration, testing, deployment, and infrastructure management. In addition to continuous integration (CI) and continuous delivery (CD), MLOps introduces continuous training (CT), which is unique to ML systems and is concerned with automatically training and serving ML models. Operating ML systems in production requires continuously adapting to the evolving input data. This is especially evident in time series data, which can experience frequent drifts. Moreover, implementing CT in practice is challenging and heavily dependent on the task and available data. Depending on the complexity of the model and the amount of data, the training process can be computationally costly. Using a scheduled interval for retraining is inefficient if the model still performs adequately. We designed an ML pipeline capable of efficient continuous training using an error-based trigger for retraining the model. The ML pipeline is designed for a time series forecasting task, where the data is prone to frequent drifts. We applied the design science research methodology to identify the problem, design and develop a solution artifact, and evaluate its utility and efficacy. The resulting solution utilizes an open-source MLOps platform that runs on Kubernetes. The solution includes a custom retrainer component to enable CT. We demonstrated the efficacy of the solution using real energy demand data from a university property in Finland. Our evaluation shows that the system is capable of efficient continuous training.