Recurrent neural network learning and neural network learning controller.

摘要

A set of efficient learning algorithms for time-invariant recurrent neural networks are developed based on quasi-Newton optimization techniques. Simulations show that these algorithms improve the learning rate and accuracy by at least two to three orders of magnitude when compared to the State-of-the-Art in Recurrent Neural Network Learning. Simulation and experiment further validate the usefulness of the recurrent networks in modeling nonlinear mechanical systems, which include emulating the step response of a robot arm and identifying the model of a single-screw compressor.; A model of time-varying recurrent neural networks is presented for modeling time-varying systems. The learning problem of the recurrent network is formulated as one of functional optimization. Dynamic optimization is used to derive necessary conditions for optimal weight functions. A learning algorithm for finding the weight functions is subsequently developed based on a function-space quasi-Newton method. Simulations in modeling nonlinear systems are carried out to demonstrate the ability of the recurrent network and the efficiency of the new learning algorithm.; A direct neural network learning controller is developed which is capable of improving its performance in the control of an unknown dynamic plant. Training this controller is based on Powell's gradient-free learning algorithm. Thus, no plant model but only the output of the plant is required. Simulations show that the controller not only learns to follow a trajectory better but also has a faster learning rate when compared to conventional linear controllers.; The time-invariant recurrent networks are further used to identify and control robot arms. First, the recurrent networks are trained to identify the forward and inverse dynamics of a robot arm excited by white noise and color noise respectively. Then, a controller consisting of the inverse model and a linear controller is established for trajectory control. Furthermore, a learning control scheme is developed to further improve trajectory tracking performance for repetitive processes. Finally, simulations are carried out to show that the recurrent networks can identify the forward and inverse dynamics of a nonlinear dynamic system and that the learning controller can achieve very high tracking accuracy.