NONLINEAR SYSTEM IDENTIFICATIONUSINGGENETICALGORITHM BASEDRECURRENT NEURAL NETWORKS
Yuqing Zhu, Wenfang Xie · 2006
Inthis paper, a new Genetic Algorithm (GA)isdeveloped tooptimize thearchitecture of a Recurrent Artificial Neural Network(RANN)withmultiple hidden layers. A new Direct Matrix MappingEncoding (DMME)methodisproposed to efficiently andeffectively represent thearchitecture ofa neural network. A modified Back-propagation (BP)algorithm is utilized totunetheweights andother parametersofRANNs. TheRANNoptimized bythis algorithm hasbeenapplied tothe identification of nonlinear dynamic systemswithunknown nonlinearities. ThreetypesofRANN-based nonlinear models areproposed todescribe thebehavior of nonlinear systems. Theeffectiveness ofthese models andidentification algorithms areextensively verified intheidentification ofseveral complex nonlinear systemssuchas smart actuatorpreceded by hysteresis, andfriction-plague harmonic drive.