提出了一种新颖的具有自构筑能力的神经网络结构,称之为Modular-tree和两个相应的自构筑算法。在此结构中,任何现存的前馈神经网络均可以作为子网。对于一个给定的学习任务,利用提出的生成算法通过对输入空间递归地划分,自动生成一树状的模块神经网络,从而避免了网络结构预置问题。由于使用了“分治”原理,Modular-tree具有良好的性能及快速训练的能力。此结构已用于多个监督学习问题(包括:标准测试及现实世界问题)并取得令人满意的实验结果。 A novel self-construction neural network structure is proposed, which is called Modular-tree and two corresponding self-building algorithms. In this structure, any existing feedforward neural network can be used as a subnet. For a given learning task, a tree-like module neural network is automatically generated by recursively dividing the input space using the proposed generation algorithm, thereby avoiding the problem of network structure pre-configuration. Due to the “divide and conquer” principle, Modular-tree has good performance and quick training ability. This structure has been used for a number of supervised learning problems (including: standard testing and real-world problems) and satisfactory experimental results.