This paper compares different neural computing approaches to estimating the
number of changes needed in a software product based on assessments of sof
tware quality. Two forms of neural computation are considered: fuzzy neural
computation based on fuzzy sets and rough neural computation based on roug
h sets. Both forms of neural computation are defined in the context of the
McCall software quality evaluation framework, which is hierarchical. This h
ierarchy has three levels: factors (highest-level based user views of softw
are quality), criteria (mid-level based on characteristics of software), an
d metrics (lowest level based on quantification of software quality). The i
ntroduction of a neural approach to estimating the number of changes needed
to achieve software quality according to a project requirement is motivate
d by the need to harness the complexities inherent in the relationships bet
ween factors, criteria and metrics. The architecture of both types of netwo
rks is given. The results of calibrating both types of networks are also gi
ven. The performance of the two forms of neural computation is compared. (C
) 2001 Elsevier Science B.V. All rights reserved.