The divide-and-conquer neural network: its architecture and training

An efficient method of constructing mapping neural networks based on the divide-and-conquer principle is presented. The network thus constructed is referred to here as the divide-and-conquer neural network (DCNN). Instead of solving the whole difficult problem in one step, it is divided into much simpler subproblems, each of which can be learned locally by any kind of learning methods (in this paper, the method to solve linear equations has been used). Since the subproblems are easier, it could achieve less errors, better approximations, and better generalizations with reasonably small network size. Moreover, the number of hidden units and layers are determined as part of its operation and there is no concept about learning rates. Therefore, it will not suffer from trial and error like BP (backpropagation). Since any learning method can be used to learn the subproblems, it leads to extensive varieties of new construction algorithms in the future