Adapting to Non-stationarity with Growing Expert Ensembles