|
The application of neural networks have been widely infiltrated into many other areas of life sciences and engineering sciences , gradually deepening with the theory of learning algorithm in intelligent control , pattern recognition , optimization computing, machine vision and biomedical made ??many achievements . Classic learning algorithm to converge , however , is slow , the algorithm is not complete , easy to fall into local minimum points and shortcomings , and it is difficult to deal with large - scale learning sample . In view of this, we will study the neural network learning algorithm based on analytical optimization method . First approximate intermediate value through the structural hidden layer , will based on the linear least squares fast learning algorithm is extended to multilayer neural networks from a single-layer neural networks , numerical experiments show that the new algorithm is suitable for large-scale sample of learning , has the iteration early learning error dropped quickly characteristics significantly improve the speed of training . In turn based on unconstrained optimization tone than by constructing a new self-adjusting tone scaling factor and collinear factor , respectively, to get a new class of self-tuning than BFGS algorithm and collinear tone than the BFGS algorithm . We give a global convergence of the new algorithm , neural network learning algorithm based on a new algorithm for the numerical experiments and analysis , and the results show that the new algorithm has better performance and numerical stability . Finally, given the algorithm using Matlab toolkit .
|