Study And Application Of Dynamic Tunneling Technique In Training Of BP Neural Network

The dynamic tunneling algorithm for global optimization has been proposed recently. The dynamic tunneling algorithm is composed of a sequence of cycles. Each cycle includes solving two kinds of dynamical systems: a dynamic optimization system by which a local minimum is found and a dynamic tunneling system by which a new initial solution in a lower valley is determined. The two dynamic systems work alternately to approach a global optimal solution of an objective function.BP (Back Propagation) algorithm is the most popular training algorithm in applications for its non-linear mapping approach capability and robustness. However, it has some defects, such as converging slowly and immersing in local vibration frequently. So an algorithm using dynamic tunneling technique to train BP neural networks has been proposed and has been proved to have good performances.Based on the conventional dynamical tunneling technique, the multi-trajectory dynamic tunneling algorithm has recently been proposed. This algorithm improves search efficiency of the conventional dynamic tunneling system by increasing the trajectories and introducing the interaction between each trajectory of the tunneling system.This paper proposes a new algorithm adopting the multi-trajectory dynamic tunneling technique and the error-limitation dynamic changing technique to train the BP neural networks. The simulation results are provided for three different examples to demonstrate the performance of the proposed method in overcoming the problems of initialization and searching efficiency. The performance of the conventional dynamic tunneling technique and the multi-trajectory dynamic tunneling technique in training BP neural networks are also given and compared in this paper.Moreover, this paper constructs a parallel compute platform to apply the new algorithm by using PC group and PVM virtual system and LINUX OS. On this platform this paper adopts the Master/Slave modal and assigns the data to every computing node to compute. The performance has been provided in the paper.