Interpolated regression for on-line local modeling in feedforward learning control

This paper proposes a technique of on-line modeling for feedforward (FF) learning control. For an unknown nonlinear multi-input multi-output (MIMO) plant which is free of zero dynamics, we construct a bank of filters each of which corresponds to a local model of inverse dynamics. In real time we select one such filter corresponding to the current operating point (called scheduler) and accumulate input-output (i/o) data to the filter while we derive FF control signal via regression from the accumulated data. The number of filters is finite but the plant operates continuously, hence we need to discretize the scheduler for classification. In a conventional scheme, however, we have merely truncated the value and resulted in a large approximation error. This paper proposes yet another scheme that uses an interpolation technique for regression in local modeling, thereby improving accuracy of response shaping. Numerical simulation is carried out to verify effectiveness of the proposed scheme.