Neural Nets and Diversity

Although the issue of reliability is extensively discussed in the software engineering literature, it has received only limited attention in the Neural Computing community. In this paper, the software engineering concept of diversity is made use of to improve the performance of a neural net system solution to a problem of fault diagnosis in a marine diesel engine. Essentially the aim here was to find methods of creating a set of solutions which are diverse in the sense that they each fail on different inputs. A truly diverse set of solutions can be combined by means of a majority voter to yield 100% generalisation performance. The issue of identifying the best ways to promote diversity in neural nets was investigated in terms of a fault diagnosis problem in a marine engine. It was concluded that two effective methods for creating diverse solutions were (i) to take data from two different sensors, or (ii) to create new data sets by subjecting the set of inputs to non-linear transformations. These conclusions have far reaching implications for other Neural Net applications.