|
|
A DYNAMIC REDUNDANCY BP ALGORITHM APPLIED IN THE FAULT TOLERANCE OF NEURAL NETWORKS |
Xu Liqin; Hu Dongcheng |
Department of Automation Tsinghua University Beijing 100084 China |
|
|
Abstract There are two main types of approaches in the research of fault tolerance of Multilayer Perceptrons(MLP): improvement in the learning algorithm and component redundancy after training. A dynamic redundancy BP algorithm is presented. In the training steps of the conventional adaptive BP algorithm with a momentum term, the most important weights are replicated based on their "significance". Applying the algorithm the fault tolerance of a network can be improved effectively. Compared with fault injection while training--a typical improved learning algorithm, although this dynamic redundancy algorithm gives no prominence in fault tolerance, the training time can be greatly reduced.
|
Received: 29 September 1999
|
|
|
|
|
|
|
|