Initialization of neural networks by means of decision trees

Performance of neural-networks learning is known to be sensitive to the initial weight setting and architecture|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 by decision-tree generators. The system TBNN (tree-based neural net) compares very favourably to other learners in terms of classiication accuracy on unseen data and is also computationally less demanding than the backpropagation algorithm applied to a randomly initialized multilayer perceptron. The behavior of the system is rst studied on specially designed artiicial data. Then, its performance is demonstrated by a real-world application.