Multiresolution Controller Based on Window Function Networks for a Quanser Helicopter

To improve neural network (NN) performance, new activation functions, such as ReLU, GELU, and SELU, to name a few, have been proposed. Windows-based functions, such as flat-top or atomic functions, used in processing signals have begun to be used in NNs for dynamical systems. Although wavelet functions are the most popular, some additional functions with similar properties must be evaluated. This paper presents a window function neural network (WFNN) for identification tasks. Using the identification by WFNN, the self-tuning gains of the proportional multi-resolution (PMR) controller are carried out. To show the performance of the proposed approach for different window functions, numerical simulations of the Quanser helicopter of 2 degrees of freedom are presented.