Improving modified cocomo ii artificial neural network using hyperbolic tangent activation function
Software cost estimation is a complex and critical issue in software industry but it is an inevitable activity in the software development process. It is one of important factors for projects failure due to the ambiguity and uncertainty of software attributes at the early stages of software devel...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| التنسيق: | أطروحة |
| اللغة: | English English English |
| منشور في: |
2017
|
| الموضوعات: | |
| الوصول للمادة أونلاين: | http://eprints.uthm.edu.my/859/1/24p%20SARAH%20ABDULKAREM%20ABDULAZIZ%20AL-SHALIF.pdf http://eprints.uthm.edu.my/859/2/SARAH%20ABDULKAREM%20ABDULAZIZ%20AL-SHALIF%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/859/3/SARAH%20ABDULKAREM%20ABDULAZIZ%20AL-SHALIF%20WATERMARK.pdf |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | Software cost estimation is a complex and critical issue in software industry but it is
an inevitable activity in the software development process. It is one of important
factors for projects failure due to the ambiguity and uncertainty of software attributes
at the early stages of software development. The estimation of effort in COCOMO II
depends on several software attributes namely software size (SS), scale factors (SFs)
and effort multipliers (EMs). Several researchers integrate COCOMO II with Artificial
Neural Network (ANN) to overcome the ambiguous and uncertain of these attributes.
However, ANN contributes to slow convergence caused by sigmoid function. Thus,
this research proposes Hyperbolic Tangent activation function (Tanh) to be used in the
hidden layer of the ANN architecture to produce faster convergence. Back-propagation
learning algorithm is applied to the multilayer neural network for training and testing.
The proposed activation function has been trained and tested using two different
architectures of NN which are basic COCOMO II-NN and modified COCOMO II-NN
that uses COCOMO II NASA93 dataset. The result has been compared to different
activation functions namely Uni-polar sigmoid, Bi-polar sigmoid, Gaussian and
Softsign. The experiment results indicate that Tanh with modified COCOMO II-NN
architecture achieved 23.2780 % Mean Magnitude Relative Error (MMRE) for 19
testing projects and 9.8948 % MMRE for 9 testing projects which is the lowest MMRE
among other activation functions. In conclusion, Tanh with modified architecture of
COCOMO II-NN provides much better estimation results than other methods and can
lead to improvement of software estimates. |
|---|