多层前馈神经网络的学习和综合算法

A LEARNING AND SYNTHESIS ALGORITHM OF MULTILAYERED FEEDFORWARD NEURAL NETWORKS

本文提出多层前馈网络的一种新的学习和综合算法──ＦＰ算法，并证明由此算法得到的网络作为通用联想记忆器时，具有如下优点：（１）每个样本都是吸引中心；（２）每个样本的吸引半径达到最大值；（３）网络没有假吸引中心；（４）网络具有最少的元件个数；（５）学习的复杂性达到最优（就其复杂性的阶而言）．故此网络在性能、结构、计算复杂性等方面均达到很好状态．

A new learning algorithm-forward propagation (FP) of multilayered feed-forward neural networks is presented in this paper. The authors show that as an associative memory the network constructed by the FP algorithm has several advantages. (1)Each training sample is an attractive center. (2) The attractive radius of each training sample reaches the maximum. (3) There is no spurious attractive center in the network.(4) The network has minimal number of elements. (5) The order of its learning complexity is optimal. The FP learning algorithm is also an effective synthesis tool, i. e., the network architecture can be constructed during its learning process.