Operant Conditioning-inspired Neural Adaptive Control

With the increasing complexity of system and its environment,model-and-rule based classical control methods can hardly satisfy the performance requirements.Therefore,novel and advanced control methods are of both theoretical significance and practical interest for long-term safe and stable operation of systems.This dissertation focuses on developing high performance,cost-effective and user-friendly control for nonlinear systems in the presence of unknown floating structure,abrupt external disturbances,mechanical wear,sub-system faults,and functional failures of critical components(e.g.,sensors or actuators).From the angle of biological facts,a series of brain-inspired neural adaptive controllers are presented to prevent the deterioration of control performance caused by serious nonlinearities and uncertainties.By integrating the neuroscience with control theories,the running mechanism of traditional feed-forward neural network(NN)is optimized,and its learning and reasoning abilities are further enhanced.As a result,the overall system performance can be improved including control precision,convergence rate,computation efficiency,and operation stability.The main contents and contributions of the dissertation are summarized as follows:1)Inspired by the biological concept of operant conditioning,the dissertation presents a novel type of NN,called operant conditioning bionic model(OCBM),and then applies it to a class of high-order non-affine systems.By developing reward strategy and neuron adaptive units,the OCBM has online adaptive weights,auto-tuning structural parameters of basis functions and self-growing number of neurons.The OCBM based control scheme derived from Lyapunov stability analysis can ensure the closed-loop system semi-global uniformly ultimately bounded.Numerical simulation results validate the effectiveness of the proposed method in dealing with unknown floating system structure and uncertain external disturbances.Comparative studies also show that the proposed control has higher control precision,faster convergent rate and less computational burden than some typical methods.2)By employing locally weighted learning framework,the OCBM is improved systematically.To account for uncertain/unknown nonlinearities,a finite neuron self-growing(FNSG)strategy is proposed to guide the process for adding new neurons,resulting in a prototype of self-adjustable NN structure for better learning capabilities.Another improvement is the use of barrier Lyapunov function(BLF)rather than switch control technique,which avoids discontinuous control actions and ensures the NN inputs to remain bounded during the entire system operation.Furthermore,the control action is also smooth almost everywhere except for a few number of time instants at which new neurons are produced.It benefits from the use of the smooth saturation function,the continuous weights updating law,and Gaussian-type weighting functions.Compared with NN with fixed number of neurons and self-organizing structure,simulation results illustrate that the FNSG based control can avoid generating redundant neurons,and helps to save computational resources.3)Motivated by the discoveries on the structure and regulatory mechanism of the brain-nervous system,a multiple self-adjusting elements based NN(MSAE-NN)is constructed with time-varying ideal weights,diversified basis functions,and self-growing and pruning neurons.The strategy of neuron self-growing and pruning is derived from FNSG and modified to include smooth operators to prevent neurons from being added or eliminated abruptly.Meanwhile,an MSAE-NN based control approach is presented for high-order non-affine system with discontinuities.By combining robust adaptive technique and BLF stability analysis,the MSAE-NN unit is enabled to be fully functional in the control loop during the entire process of system operation and thus ensures more reliable and more effective control performance.4)In this dissertation,the model of MSAE-NN is further applied to a class of multi-input multi-output non-affine systems.Combining with two specific tasks in robot manipulator control,two brain-learning associated control(BLAC)schemes for joint-space and Cartesian-space tracking are presented respectively to deal with discontinuous model,unknown disturbances and subsystem faults.The proposed control approaches are capable to avoid parameter estimation of basis functions and thus reduces the tedious process of tuning parameters manually.Furthermore,the MSAE-NN can be ensured effective during the period of system operation because of the use of BLF based rigorous theoretical design.It is worth mentioning that the resultant controllers do not require precise system parameters and inverse operation,which makes it a simple structure to implement,cost-effective and easy to use.