A hybrid intelligent active force controller for robot arms using evolutionary neural networks

In this paper, we propose a hybrid intelligent parameter estimator for the active force control (AFC) scheme which utilizes evolutionary computation (EC) and artificial neural networks (ANN) to control a rigid robot arm. The EC part of the algorithm composes of a hybrid genetic algorithm (GA) and an evolutionary program (EP). The development of the controller is divided into two stages. In the first stage, which is performed off-line, the proposed EC algorithm is employed to evolve a pool of ANN structures until they converge to an optimum structure. The population is divided into different groups according to their fitness. The elitist group will not undergo any operation, while the second group, i.e. stronger group, undergoes the EP operation. Hence, the behavioral link between the parent and their offspring can be maintained. The weaker group undergoes a GA operation since their behaviors need to be changed more effectively in order to produce better offspring. In the second stage, the evolved ANN obtained from the first stage, which represent the optimum ANN structural design, is used to design the on-line intelligent parameter estimator to estimate the inertia matrix of the robot arm for the AFC controller. In this on-line stage, the ANN parameters, i.e. the weights and biases, are further trained using live data and back-propagation until a satisfactory result is obtained. The effectiveness of the proposed scheme is demonstrated through a simulation study performed on a two link planar manipulator operating in a horizontal plane. An external load is introduced to the manipulator to study the effectiveness of the proposed scheme.