Design Analysis And Application Of The Recurrent Neural Dynamic Models

A number of problems might be regarded as a class of system to investigated in the practical engineering and scientific computing.However,the traditional numerical algorithm of the digital computer serial processing may not be enough satisfied the requirements of real-time,on-line,speediness and accurate solution in dealing with them.Therefore,it is an urgent need for a kind of to use a parallel computing method different from serial processing to deal with these problems in the above fields.A kind of neoteric computational method based on the recurrent neural dynamics has emerged and been widely used in the above domain.The corresponding recurrent neural dynamic model of the problem is established,and the problem to be solved is to deal with as a kind of system in terms of working principle.Starting from its arbitrary initial state,it dynamically approaches the stable state of the problem to be solved,and finally obtains the stable state of the system.The recurrent neural dynamics not only show excellent dynamic behavior in dealing with problems.It also has the following advantages.First of all,the advantage in solving time-varying and time-invariant problems is the requirements of real-time,rapider and accurate solution.Then,it is easy to build software and hardware.A neural dynamic model relation to the corresponding problem,and appropriate electronic components are needed to build a circuit model,then the solution to the problem is obtained online from the circuit model.Finally,the reason is that the neural network model has the characteristics of biological nerve,which makes it have the ability of associative memory and self-adaptation.Therefore,it is of immense value to design the recurrent neural dynamic models,analyze the stability of the network models and broaden the application of related fields.In this paper,the recurrent neural dynamic models are used to solve the time-varying matrix inverse problems,time-varying convex quadratic programming problems and practical engineering application problems.The main research work is as follows:1.In order to solve the inverse problems of time-varying matrix on-line,an improved zeroing neural dynamic model is designed and constructed,which is compared with the traditional gradient neural dynamic model and the zeroing neural dynamic model.Under the condition of using five kinds of monotonically increasing activation functions,the improved zeroing neural dynamic model has faster convergence speed in solving the inverse problems of time-varying matrix,and the error between the state solution and the theoretical solution of the network is smaller.2.In order to solve the time-varying convex quadratic programming problems,a zeroing neural dynamic model with time-varying network design parameters is designed and constructed,which is compared with the traditional gradient neural dynamic model and the zeroing neural dynamic model.Under the condition of using the new Sign-bi-power activation function,the zeroing neural dynamics of time-varying network parameters has faster convergence speed in solving the time-varying convex quadratic programming problems,and the error between the state solution and the theoretical solution of the network is smaller.3.In order to solve the problem of practical engineering application,three kinds of the recurrent neural dynamic models are used to solve the matrix inversion problem of the MIMO system channel detection,and the results are better.From the point of the convergence speed,error precision and network complexity,the appropriate network models selection are given.