Deep Bio-inspired Neural Networks Based Adaptive Control

With the increases of the complexity of control system,the realtime of control objective and the uncertainty of working environment,the traditional control theory and method based on the mathematical model of control system can hardly solve various control problems associated with uncertain nonlinear systems.Because of its continuous nonlinear function approximation ability,adaptive learning ability,and higher parallel ability,neural network has been widely adopted in dealing with the modeling uncertainties and coupling nonlinearities in complex systems.As the artificial imitation of biological neural network,the learning ability of neural network mainly depends on the imitation degren of biological characteristics.Therefor,it is of great significance to design neural network with more traits of real biological neuron system,and then study such kind of neural network based control theory and method.By combining the existing results of neuroscience and complex network theory,in this paper,several biologically plausible neural networks,which are focused on neural plasticity or/and neural network structure,are constructed.Furthermore,their applications in nonlinear system control are studied.The main contents of this paper are summarized as follows:(1)As the radial basis function neural networks are widely used in nonlinear control,in this paper,a new intrinsic plasticity computational model for radial basis function is derived and an intrinsic plasticity based RBFNN is proposed.Then an intrinsic plasticity based RBFNN adaptive control method is proposed for a class of nonlinear uncertain systems.The rigorous stability analysis and proof are given by applying the Lyapunov stability theory,through which all signals in the closed-loop systems are uniformly ultimately bounded.Numerical simulations confirm that,compared with the traditional RBFNN based control,the proposed control scheme can improve the control precision evidently.(2)Instead of only considering the intrinsic plasticity,here the interplay between intrinsic and synaptic plasticity is investigated.Furthermore,a new type of recurrent neural network —— Echo state network,not the RBFNN,is utilized.Thus an intrinsic plasticity and synaptic plasticity based ESN is obtained,in which the intrinsic plasticity is used to adjusting the activation functions and the synaptic plasticity is applied to neural weights.Then an intrinsic and synaptic plasticity based neuroadaptive tracking control scheme is developed for a class of uncertain MIMO systems.The rigorous stability analysis and proof are given by applying the Lyapunov stability theory,through which all signals in the closed-loop systems are uniformly ultimately bounded.Numerical simulations confirm that,compared with traditional RBFNN based control method and intrinsic plasticity based ESN control method,the control performance of the proposed method is evidently enhanced.(3)Inspired by the clusters structure in mammalian brain networks,a multi-cluster structure is generated off-line and then a diversited multi-clusted reservoir based ESN is constructed,in which the neurons in the same cluster have the same function(share the same type of activation function)and the connections are more intensive,while the neurons in different cluster are equipped with different activation functions and the connections become more sparse.Thus,on the one hand,the online computing burden is reduced,on the other hand,the coupling between the neuron is weakened and the dynamics characteristic becomes richer.Then a diversited multi-clusted reservoir based ESN adaptive control for a class of uncertain MIMO systems is developed,which is able to allow the tracking error to achieve asymptotic convergence via rigorous theoretical analysis.The effectiveness of the proposed method is also confirmed by numerical simulation via the comparison with the traditional control method,admitting better tracking performance of the proposed control.(4)Inspired by the success of deep learning,a deep ESN with multilayer reservoirs is constructed,in which each reservoir bears a diversited multi-clusted structure instead of the random structure in traditional ESN.Then a neuroadaptive prescribed performance tracking control scheme for a class of uncertain MIMO systems is developed.The rigorous stability analysis and proof are given by applying the Lyapunov stability theory,through which the transient behavior and steady state performance are ensured.Numerical simulations confirm that compared with the shallow reservoir based ESN control,the proposed one is able to increase the convergence speed and improve the control precision.(5)By combining the neural plasticity and compex network structure,a deep ESN with intrinsic plasticity,synaptic plasticity,and the diversited multi-clusted reservoir is proposed.Then a neuroadaptive tracking control scheme for a class of uncertain robot systems is developed.The rigorous stability analysis and proof are given by using the Lyapunov stability theory,through which all signals in the closed-loop systems are uniformly ultimately bounded.Numerical simulations confirm that compared with the neural plasticity based neural network control or the improved structure-based neural network control,the proposed neuroadaptive tracking control scheme with neural plasticity and improved structure can achieve the best control performance.Furthermore,it is shown that neural plasticity based neural network control method has higher control precision cpmpared with the improved structure based neural network control method.