Optimal controller approximation using neural and fuzzy-neural networks

A feedforward neural network has been shown to possess the ability of approximating an optimal controller over a subspace of initial conditions. In addition to a feedforward neural network the fuzzy logic/neural network hybrid ANFIS (adaptive network based fuzzy inference system) structure has been shown to be a universal approximator. This paper discusses briefly the use of ANFIS to approximate an optimal feedback controller for the nontrivial problem of guiding a high-speed vehicle (glider) to a ground target and compares results to a three layer feedforward neural network controller. The controllers are trained using optimal open loop state and control data generated by VF02AD. Simulation results show that both controller types guide the vehicle to within 1000 ft of the target with practically no degradation in the optimal impact velocity over a range of initial conditions. The advantages and disadvantages of both controllers are discussed together with some practical limitations of the ANFIS optimal feedback controller.