Evolutionary algorithms for adaptive predictive control

A nonlinear adaptive model predictive control strategy based on evolutionary algorithms (EAs) is proposed. An EA was employed as a robust online tuner of the weights of a neural network used to identify the mismatch between the real plant and the nominal model caused by disturbances and unmodeled dynamics. A second EA, was used as a constrained optimizer to online plan optimal input policies over a defined prediction horizon basing on the identified model. The effectiveness of the proposed control strategy was tested to control the liquid level of a two tanks nonlinear time varying simulated system. Some considerations about algorithm complexity and online computational requirements are discussed.