Adaptive Classifiers for Nonstationary Environments
Cesare Alippi, Giacomo Boracchi, Manuel Roveri, Gregory Ditzler, Robi Polikar · 2015
Deployment of sensor/actuator networks designed to address industrial and environmental applications makes possible the online acquisition of a large set of data from a monitored environment. Predictive models developed to work in non-stationary environments must then be able to detect changes, characterize and isolate faults and, if requested, adapt itself to track the change to maximize performance. This chapter expands the adaptive classifier study by presenting methods to learn the concept drift. It discusses the presentation of strategies for adapting the classifier to the new working conditions. The chapter summarizes the most common change detection tests present in the literature to address variation in the process X, while it describes the use of these tests to inspect the output of adaptive classifiers. The chapter presents an overview of a new breed of algorithms, commonly referred as concept drift algorithms, which are designed to work in such non-stationary environments.