This paper deals with the machine learning based controller design for realizing nonlinear model predictive control (MPC) with low computational cost, and the application for a diesel engine air path system is shown. Since the solution of an optimal control problem considered in MPC depends on several variables at each time, the relationship between the variables and the solution, that is, the control law could be approximated by a neural network. In the case of high dimensional inputs, however, some efficient sampling methods for the approximation are needed to avoid a combinatorial explosion. To reduce the sampling dimension, we newly applied an efficient sampling method which is combined with a design of experiment. Using the method, the neural network structured optimal controller that operates the valves in the air path system was designed, and its tracking capability to the reference value was demonstrated in the simulation. The computational time of the controller was approximately 0.022ms at each cycle, indicating that fast computation of MPC was achieved.
