Initialization of neural networks by means of decision trees

摘要

The performance of neural networks is known to be sensitive to the initial weight setting and architecture (the number of hidden layers and neurons in these layers). This shortcoming can be alleviated if some approximation of the target concept in terms of a logical description is available. The paper reports a successful attempt to initialize neural networks using decision-tree generators. The TBNN (tree-based neural net) system compares very favourably with other learners in terms of classification accuracy for unseen data, and it is also computationally less demanding than the back propagation algorithm applied to a randomly initialized multilayer perceptron. The behavior of the system is first studied for specially designed artificial data. Then, its performance is demonstrated by a real-world application.