A New Learning Algorithm Based on Trust Region Optimization Theory for Neural Networks

Neural network techniques have been widely applied to areas of such as data mining, information integration and grid computing. This paper proposes a new learning algorithm based on trust region optimization theory. In the paper, the Dogleg-algorithm to obtain the valid trust region steps is presented, and a self-adjustable method with variable coefficients is given to resolve the problem of oscillatory behaviors and low efficiency in the progress of tuning the trust region radius. We also prove the validity of the algorithm, and analyze experimentally the performance and characteristics of the algorithm.