Estimation of inverse model by PSO and simultaneous perturbation method

Abstract

This paper describes estimation of the inverse model by multi-layered neural network. The back-propagation rule requires a sensitivity function of a system. If the system has uncertainly, then we can not calculate the sensitivity function. Hence, we propose a learning rule based on particle swarm optimization (PSO) combining with simultaneous perturbation. PSO and simultaneous perturbation are suitable for estimation of the inverse model with uncertainly, because they can update by only value of the objective function. PSO has a capability of finding a global minimum and simultaneous perturbation can search local area efficiently. We introduce two adaptation method of the combination ratio. One of them is to adapt it depending on the distance from gbest. The other is to adapt it depending on the value of the objective function. The proposed method are investigated using inverse kinematics problem. The simulation results show that the proposed methods obtain the more accurate inverse model.