Optimizing neural networks for pneumatic muscle actuator system identification: a cooperative coevolutionary approach

Abstract

This paper presents a cooperative coevolutionary optimization algorithm to overcome issues in gradient descent-based neural network, such
as getting stuck in local minimum and slow convergence. The proposed method combines JAYA and a modified differential evolution (DE)
techniques to optimize neural network weights. It works by splitting the population into two subpopulations, each focusing on optimizing
different aspects of the network weights. The method's effectiveness is tested on two benchmark nonlinear dynamical systems and compared
with existing methods. Results show that the neural network optimized by this approach achieves high accuracy and robustness. Finally, the
practical applicability of this method is demonstrated by modeling the pneumatic muscle actuator (PMA) system using experimental data,
where the PMA system is made up of Festo's MAS-10 N220 pneumatic artificial muscles and controlled with a DAQ NI 6221 card.