Abstract
This paper explores the issue of modifying machine learning algorithms when they are used with streams of data that have changing statistical properties and changing concepts as the data changes. Traditional methods of using static, stable datasets do not provide accurate results for dynamically changing datasets, so new dynamic methods for adapting to these changes are needed. As methodology the article proposed a state-driven processing architecture that combines the use of monitoring mechanisms to monitor the system for changes; detect changes in the system; and include decision-making processes into one single entity. This approach defines a model for the system state that provides an assessment of how stable the system is, and how far from its expected behaviour it has gone, so the appropriate adaptive methods can be selected. The results demonstrate the proposed approach provides more stable predictions; improved overall model performance; and greater model robustness in the face of changing conditions. The developed framework can be used in real-time systems, where system reliability, adaptability and rapid response to changes in underlying data are important, including defense-oriented systems.