Research On Modeling And Control Of Robots Using Recurrent Neural Networks

Recurrent neural networks (RNN) is a kind of neural networks (NN) with feedback layer,which makes the network changed from a single static mapping into a dynamic input/outputmapping. RNN is more suitable for nonlinear dynamic system identification and controlrelative to feed forward neural networks (FFNN). However, the difficulties in its trainingalgorithm by and large have prevented its practical applications because it has to adjust all theweights in the network, echo state networks (ESN) is a novel RNN which has strong dynamicapproximation ability and fast convergence rate because of its peculiar state reservoir (SR),futhermore, only the network-to-output connection weights have to be trained, manydynamical systems, which were diffcult to learn with the existing methods, have been easilylearnt by ESN.For this reason, after study the basic ESN and its corresponding algorithm, an improvedESN learning algorithm based on the augmented strategy is proposed according to themodeling and control of the complex nonlinear dynamic system including robot system. Thenthe improved ESN is applied to the modeling and control of nonlinear dynamic system andcompared with other methods such as ESN, basic RNN as well as support vector machines(SVM) et al, experiment results show the effectiveness and superiority of the proposedmethod. The main contents of this thesis are as follows:(1) To study the basic structure and algorithm implementation of RNN. Further study thebasic principle of ESN and its learning algorithms, analyse the offline algorithm and therecursive least squares (RLS) algorithm of online learning. On this bases, give an improvedlearning algorithm for ESN based on augmented strategy, the output weight matrix of theimproved ESN is calculated by its augmented state vectors, which can enhance the nonlineardynamical property of the network.(2) Aiming at the application of nonlinear dynamic system identification based on ESN.It includes: the application of the single input single output system modeling in a higher orderNARMA model instances and hydraulic robot actuator; the application of the multiple inputmultiple output system modeling in a multidimensional nonlinear dynamic system and sevendegrees of freedom SARCOS anthropomorphic robot arm. Simulation results show a rapidand stable learning speed and a good performance of ESN rather than other methods.(3) To study the application of nonlinear control of dynamic system based on ESN. Thatis the application of adaptive control in second-order as well as third-order nonlinear dynamicsystem benchmark examples respectively. Firstly, the uncertain parts of the nonlinear systemis identified by an ESN, and then the control law is designed based on the ESN in order tomake sure that output of the closed-loop system is able to track the output of the reference model. The results show good performance of ESN rather than other methods such as SVMand FFNN et al. Moreover, ESN don’t need to estimate the system delay order and show verygood robustness.(4) The ESN is applied to the trajectory tracking control of a two degrees of freedom ofthe manipulator. Firstly the inverse model of uncertain part of the manipulator is identifiedwith an improved ESN, and a PD feedback controller is added to compense the modelingerror. Then the dynamic controller is designed based on the improved ESN. Finallyexperiments are carried on and compared with ESN as well as other RNN methods,Simulation results show the good tracking performance of the control scheme.