Mixed Varying-parameter Dynamic Learning Network Solution Method For Time-varying Problems

Complex time-varying dynamic systems are a very important type of problem in the fields of science and engineering applications.In addition,due to the practical application process,an ideal noise-free system is very rare,and the presence of noise in practical applications will affect the performance of the model,and resulting in low solution accuracy.Therefore,the solution of time-varying dynamic systems with noise interference is worth exploring.Due to the traditional fixed-parameter convergence differential neural network cannot exceed the upper limit of the parameter change due to parameter selection,its convergence effect and convergence speed will be limited by the initial limit value.Therefore,according to the design idea of adaptive time-varying parameters,this paper proposes a new mixed varying-parameter dynamic learning network(MVP-DLN)is applied to solve time-varying dynamic systems with noise(time-varying convex quadratic programming and time-varying nonlinear inequalities).The main work of this paper includes the following two parts:1.Due to the convergence speed is not fast enough and robustness is weak when using the existing neural networks to solve the time-varying convex quadratic programming problem with equality constraints,the paper proposes a new mixed varying-parameter dynamic learning network.First,the theoretical analysis prove that the proposed mixed varying-parameter dynamic learning network has better convergence performance and stronger robustness performance than the modified zeroing neural network when using the monotonically increasing odd activation function by constructing and deriving the Lyapunov function.Secondly,in order to further illustrate the superiority of the performance of the proposed MVP-DLN,a simulation comparison experiment was carried out with the traditional fixed parameter differential convergence neural network(i.e.,gradient neural network,annihilation neural network,and improved annihilation neural network).The simulation results verify that the proposed mixed varying-parameter dynamic learning network has faster convergence speed and stronger robustness than the traditional fixed-parameter differential neural network when using different activation functions to solve the time-varying convex quadratic programming.2.Nonlinear inequalities are widely used in the fields of science calculation and practical engineering.The traditional zeroing neural network has a slow convergence speed and weak robustness when solving this problem.Therefore,the proposed mixed varying-parameter dynamic learning network is further used to solve time-varying vector-type nonlinear inequality problems.Firstly,by discussing the relationship between the initial state value of the variable and the time-varying initial solution set,the theoretical analysis prove that whether the initial state value of the variable is inside or outside the time-varying initial solution set,the MVP-DLN has global convergence and strong robustness with different activation functions.Secondly,in order to further illustrate the performance superiority of MVP-DLN,a simulation comparison experiment with the traditional zeroing neural network is conducted.The simulation results prove that the mixed varying-parameter dynamic learning network has faster convergence speed and stronger robustness than the traditional zero neural network when it uses different activation functions to solve the time-varying vector nonlinear inequality.