The IRMA Community
Newsletters
Research IRM
Click a keyword to search titles using our InfoSci-OnDemand powered search:
|
Building Optimal Back-Propogation Trained Neural Networks for Firm Bankruption Predictions
Abstract
There have been considerable research activities in developing optimal neural networks during the past couple of decades. Optimal neural networks have good generalization ability and incur less computational cost due to their simple structures. A simpler neural network may approximate human judgement more closely [12] and may provide fewer “rules” or simpler formulae describing the network’s behavior than complex neural networks. Several techniques have been developed to obtain optimal neural network structures. They include pruning techniques [15,11], weight decay techniques [3], network construction techniques such as upstart [2] and tiling algorithm [7], and network selection techniques [4,8]. Although each method has given encouraging results both in terms of generalization and finding efficient architectures on simple test problems, it is not yet clear which of the methods described is best for a given decision-making problem [5]. Frean [2] conducted a comparative test in which the upstart algorithm used fewer units than the tiling algorithm. The weight decay method is easy to implement and easy to use because of its close relation to the back-propagation algorithm [3]. However, the pruning technique requires the user to pay constant attention to each hidden unit’s output for each input pattern [15]. There is a need for more effective and robust pruning techniques. This study focuses on the design of an optimal neural network model for firm bankrupting prediction task. A new direct pruning algorithm will be proposed. The article is organized as follows. A brief literature review on related optimal neural network techniques will be provided in the next section. The new pruning technique for developing simplified back-propagation trained networks will be discussed. Then, the evaluations of the proposed method will be discussed. The performance of the neural network is compared to that of multivariate discriminant analysis models for matched bankruptcy samples. The article concludes with summary and future research opportunities.
|
|