Output Prediction in Presence of Systematic Model Mismatch

In system identification for output prediction, unknown systematic errors in the measurement model and systematic drifts in the parameters have conventionally been treated with ad-hoc methods like exponential forgetting and moving windows. Such devices, although useful for zero-mean random errors and slow drifts, yield "arbitrary" tuning-dependent parameter estimates and output predictions in the realistic case of systematic model mismatch. This work presents a new method that yields optimal output predictions in face of systematic errors and drifts, without involving the arbitrariness of any user-specified tuning knobs.