Research On Nonlinear Adaptive Inverse Control Using Dynamic Neural Networks

Adaptive inverse control (AIC) is a new method for the control problems. Itsrealization is based on adaptive filtering technology. Dynamical neural networks areimportant tools for design and control of nonlinear adaptive inverse control systems(NAICSs). In this dissertation, the architecture and learning algorithms of a certaindynamical neural networks and NAICSs based on that are studied. The main contents andresults are concluded as follow:A novel dynamical neuron model is proposed for deal with temproal and spatialsignals. It is a kind of IIR filter synapses neuron, but the weights of the filters are adaptiveand the architecture is simpler than traditional ones. Feedforward neural networks based onsuch neurons, called DAFNNs, are locally recurrent neural networks, a kind of dynamicalneural networks. Compared with traditional recurrent neural networks, DAFNNs haveadaptive temporal depth and resolution. So they are good solutions for NAICSs.A modified PSO (Particle Swarm Optimization) algorithm is proposed for nonlinearfuntions optimization. It has high convergence velocity and great convergence precision byimprove the locally searcher ability on best position of the swarm. Such PSO algorithm isused as a off-line training algorithm for neural networks. It can converge to global bestsolution in theory. Simulation results also show the neural networks trained with it havebetter generalization ability than that of BPTT algorithm.A kind of on-line gradient learning algorithms based on signal flow graphs isproposed for neural networks on-line training. The signal flow graph and adjoint one of aneural network are used to compute the its gradients. The adaptive learning rate is designedby Lypunov theory for such on-line training algorithms.A modified NAICS is proposed for improve the output disturbance cancellationability of the system. Such systems can cancel the disturbance immediately withoutaffecting other adaptive processings. Another NAICS is proposed in the paper. A set ofadaptive LMS filters are used as identifiers and controllers in the systems. The architectureof the systems is similar to a linear AIC. So the adaptive algorithms of the systems aresimpler than that of typical NAICSs.