Estimation of inverse model based on ANN and PSO with adaptively varying acceleration coefficients

Abstract

We consider estimation of an inverse model for the uncertain systems. Multi-layered neural network (NN) can approximate any continuous nonlinear mapping, then we use NN in order to represent the inverse model. The the back-propagation rule, which is commonly used for updating the weights, requires the sensitivity function of the system. However, we can not calculate this function because of uncertainty of the systems. Hence, we apply particle swarm optimization (PSO) to update the weights. PSO is suitable for learning NN, because it dose not require the derivative of the objective function. This paper introduces a novel parameter automation strategy in order to overcome the premature convergence. The acceleration coefficients are adaptively varies with the distance between the particle and the gbest which is the best solution ever found by the swarm. We can maintain the diversity of the swarm and search a better solution around the gbest efficiently. We compare the proposed method with the basic PSO and TVAC PSO, which is proposed by Ratnaweera, through the simulation of the inverse kinematics problem. The proposed method can obtain the accurate inverse model than the basic PSO and TVAC PSO.