Adaptive Regression Trees for Nonlinear Adaptive Filtering
Jarren T. Worthen, Todd K. Moon, Jacob H. Gunther · 2022
Decision trees provide interperable parametric nonlinear models with effective algorithms to train the models on a fixed set of training data. In this paper we devise a method of making these models adaptive, so they adjust to data with changing statistics. This is accomplished by partitioning the trees in the model into sets called groves, then aging out groves, similar to the aging of errors in a recursive least-squares models. As old groves age out, new groves are introduced which compensate both for the changes in statistics of the data and the aging of the data. The adaptive algorithm is tested on a prediction problem on a simulated nonlinear signal.