Neuro-Fuzzy Controller Based on Model Predictive Control for a Nonlinear Underactuated Mechanical System

Fuzzy Logic Control (FLC) has been viewed as an effective feedback approach that simplifies the mathematical burden when developing complex control systems. The rationale of FLC is the digital implementation of "human-like" control rules provided either by experience or by an expert agent. Unfortunately, for high-precision control applications, FLC has proved difficult to identify an appropriate set of such rules. This drawback was compensated by the emergence of NeuroFLC techniques based on artificial neural networks. Neuro-FLC has therefore focused on self-learning inference methodologies in which the required rules (and fuzzy sets) have been determined automatically. Despite some attractive features, many self-learning approaches present significant challenges when the hardware implementation is resource-constrained. One such challenge relates to what has been called the rule explosion problem: this describes the fact that FLC inference methodologies tend to create very large rule sets for multivariable control systems. Therefore, this paper proposes a Neuro-FLC based on a sub-cluster rule reduction in order to implement the algorithm on a resource constrained embedded system (e.g., ARM Microntroller). The Neuro-FLC technique learns from a Model Predictive Algorithm (MPC) and it is implemented for controlling an underacuated nonlinear mechanical system. The algorithm proposed is capable of deploying the optimal energy to the system that guarantees stability while the performance related to the time-response can be safely chosen by the user.