A fuzzy adaptive controller using reinforcement learning neural networks

The authors describe an adaptive controller for complex processes which is capable of learning effective control using process data and improving its control through online adaptation. The controller is applicable to processes with multivariable states and with uncertain or nonlinear dynamics for which analytical models or standard control algorithms are either impractical or cannot be derived. This controller performs a fuzzy discretization of the process state and control variable spaces, and implements fuzzy logic control rules as a fuzzy relation. The membership functions of the fuzzy discretization are adjusted online and the fuzzy control rules are learned using a performance measure as feedback reinforcement. The fuzzy discretization procedure employs a data compression technique permitting multivariable state vector inputs. The controller is implemented with neural networks. Simulation results for the controller applied to a simple dynamical system demonstrate its effectiveness