A new constructive learning algorithm is proposed in this paper.By allowing to add one subnet each time to existing network during the course of training

it can automatically divide a complex task into a series of subtasks which are processed by their corresponding subnets.The divide and conquer strategy can arrive at a solution to a given task.Its strength lies in its automatic decompostion of given task and its modular training of network.Each of them leads to a great reduction of optimal search effort and an improved training.The simulation result shows that it not only achieves a better perfomance than BP network on two spiral classification problem

but also achieves a better perfomance than Cascade Correlation Learning Network in generalization ability.