Direct versus indirect neural control based on radial basis function networks

This work presents a comparison between direct and indirect neural control methods based on the radial basis function (RBF) architecture. As far as direct control schemes are concerned, a novel direct inverse neural RBF controller taking into account the applicability domain criterion (INCAD) is utilized. ? model predictive control (MPC) formulation based on RBF networks is tested as an example of indirect method. The performances of the two control schemes are evaluated and compared on a highly nonlinear control problem, namely control of a continuous stirred tank reactor (CSTR) with multiple stable and unstable steady states. Results show that the INCAD controller is able to provide satisfactory performance, while performing almost instant calculation of the control actions. MPC on the other hand, outperforms the INCAD in terms of speed of responses, due to the built-in optimization capability; however, the lengthy procedure of solving online the optimization problem impedes the practical use of MPC on systems with fast dynamics.