An Approach for Software Effort Estimation using Artificial Neural Network

Abstract

Effort Estimation of Software is one of the important task to determine number of persons, time and cost required to develop the project. Various scientists and industrialists elaborated different approaches to find out effort estimation of a project. Recent methods can be categorised into following section:( 1) Paramet ric Models: Constructive Cost Model-II (COCOMO-II)[1], Software Evaluation and Estimation of Resources Software Estimation model(SEER-SEM)[2] (2) Ex pert judgement (3) Various machine learning methods and analogy based (4) Hy brid methods[3]. After studying the various approaches, current work proposes by developing a model based on artificial neural network using the mathemati cal approach proposed by the COCOMO-II. Calibration of constant value “A” in COCOMO-II approach is done by taking average of first n input data set. The new calibrated constant value of "A" is used to train and test the model. Input data set are divided into three part horizontally. First fifteen attributes are effort multipliers and sixteenth attribute is the kilo lines of code (KLOC) and seven teenth parameter is the actual result in person months. First two sets of input data are passing to two set of neurons of artificial neural network(ANN). Input to one set of neuron is given by fifteen attributes of effort multiplier and input to second neuron is given by KLOCattribute. Forwardprop agation and backward propagation are applied to the ANN to reduce the error and assign the appropriate weight to each input attribute of the neural network and hidden layers. The input data set Cocomo81.csv uses from PROMISE Soft ware Engineering Repository. ANN with calibrated constant value of "A" , ANN without calibrated constant value of "A", RandomForestRegression,SupportVec tor Machine(SVM) and K-Nearest Neighbour models are trained and result are compared. The result suggest that model trained using calculated new constant value of “A” gives more accuracy than the model trained using original constant value of “A” proposed by the COCOMO-II.