A forward propagation learning algorithm(FP)of multilayered neural networks with feedback connections is presented in this paper．And the properties of cluster networks are discussed．A cluster with different grain-sizes can be obtained by applying FP to cluster．Its convergent speed is just a linear function of sample size．Its computational complexity is a bilinear function of simple size and the dimension of imput vectors．The network constructed by the algorithm uses a comparatively fewer number of elements and its weight simply has one of three values，i．e．，一1，0，1．Thus，it can be easily implemented into electronic circuits．The authors also discuss the properties of the network and show it is an ideal pattern classifier．