Improving the Accuracy of COCOMO II Effort Estimation Based on Neural Network with Hyperbolic Tangent Activation Function

Abstract

Constructive Cost Model II (COCOMO II) is one of the best-known software cost estimation model. The estimation of the effort in COCOMO II depends on several attributes that categorized by software size (SS), scale factors (SFs) and effort multipliers (EMs). However, provide accurate estimation is still unsatisfactory in software management. Neural Network (NN) is one of several approaches developed to improve the accuracy of COCOMO II. From the literature, they found that the learning using sigmoid function has always mismatched and ill behaved. Thus, this research proposes Hyperbolic Tangent activation function (Tanh) to use in the hidden layer of the NN. Two different architectures of NN with COCOMO (the basic COCOMO-NN and the modified COCOMO-NN) are used. Back-propagation learning algorithm is applied to adjust the COCOMO II effort estimation parameters. NASA93 dataset is used in the experiments. Magnitude of Relative Error (MRE) and Mean Magnitude of Relative Error (MMRE) are used as evaluation criteria. This research attempts to compare the performance of Tanh activation function with several activation functions, namely Uni-polar sigmoid, Bi-polar sigmoid, Gaussian and Softsign activation functions. The experiment results indicate that the Tanh with the modified COCOMO-NN architecture produce better result comparing to other activation functions.