Direct Neural-Adaptive Control with Quantifiable Bounds and Improved Performance

A new robust weight update method for use in direct neural-adaptive (or fuzzy-adaptive) control of oscillating systems can improve performance when compared to (e-modification. The proposed method is particularly applicable when using neural networks with local basis functions like the the Cerebellar Model Arithmetic Computer. The existence of ultimate bounds on the signals is established using a Lyapunov function. These ultimate bounds depend on the hounds of the nonlinear functions, whereas e-modification results in ultimate bounds that depend on a set of unknown ideal weights. Simulations demonstrate the new method performs very well in a situation where e-modification performs very poorly: when the input oscillates between two local basis functions.