11 IV April 2023

https://doi.org/10.22214/ijraset.2023.50822
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

Software Effort Estimation Based on Use Case

Reuse (Back Propagation)
Venkatesh A1, Pavan2, Santosh3, Yugandhar G4, Sunil Manoli5
School of Computer Science and Engineering, REVA University, Bengaluru-560064, Karnataka, India

Abstract: Project managers often use effort estimating strategies to manage the human resources of current or upcoming
software projects. Prior to project implementation, cost, time, and personnel estimation are basically necessary. For every project
of software, getting accuracy in Effort Estimation has always been difficult. In this study, the estimation of software
development effort was determined using a back propagation model. This model's goal is to investigate the capabilities and
potential uses of Utilizing artificial neural networks (ANN) as a tool for forecasting the effort required for software development.
In order to estimate the software work, we are attempting to implement a machine learning technique in this research. Out of all
machine learning methods, we are applying an algorithm based on Artificial Neural Networks that is Back propagation. The
Desharnais dataset, a well-known publicly available dataset for estimation of software effort, is used to test the approach. The
performance and accuracy of the tested model have been evaluated using three metrics: MMRE, MRE, and Pred (0.25). In the
sections below that follows, I explain the algorithm and its results.
Index Terms: Artificial Neural Network, Software Effort Estimation (SEE), Machine Learning, Back propagation.

I. INTRODUCTION
The action of estimating a time necessary to create software iscalled effort estimation. Estimating the work is a critical job in
the software industry. To produce accurate estimates, manycomputational models have been developed. Initial estimates without
a clear understanding of the needs are inaccurate, but as the project advances, estimate accuracy increases. Choosing the right
estimating technique is crucial as a result. Estimates of the effort can be utilized as input in budgets, investment plans, iteration
plans, and project plans evaluations, pricing methods, and bidding rounds. Since at least the 1960s, The problem of software
development effort estimation has been addressed byresearchers and practitioners in the field of software. project. The biggest
difficulty in project scheduling in the software industry is deciding how much of the project's resources should go toward the
testing phase. It has been discovered that the testingphase often uses between 40% and 50% of the resources.
Estimating the specific amount of work that has to be put into the testing phase is quite difficult, though. As a result, the project
planning is flawed. Inadequate testing of a project could cause the company to suffer severe losses. The study's primary focus has
been on creating formal models for estimating software effort. Software effort estimation has been investigated using a range of
methodologies, supporting vector regression (SVR) [4], radial basis function (RBF) neural networks [1], bagging predictors [9], and
more modern machine learning techniques, such as the COCOMO [12] and COCOMO [12]. Machine learning techniques create
models using data from previous work., which are then used to predict how much work will go into upcoming ventures. The vast
majority of techniques for calculating software effort only offer estimates [1][4][7][9][10]. However, in addition to the estimate, it
would be important to include estimates accuracy measures [10]. As a result, an estimation technique would be able to give a range
of accuracy for where the effort would fall.

Figure 1.Outcome of Poor Project Management

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2858
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

II. LITERATURE SURVEY

A review of several recent studies on software project effort estimation is provided in this section. Numerous publications have
lately been published in this area of current study. We concentrate our review in this section on a few significant article that used in
machine learning to calculate software projects effort. Despite extensive research over the past 20 years, the software industry still
faces significant challenges when it comes to accurate resource estimation. Authors have occasionally discussed different
approaches to the problem. Jorgensen provided a detailed examination of numerous studies on the development effort. 10. Wavelet
neural networks are used by K. Vinay Kumaret al.11 to estimate effort. An improved FLANN algorithm for software work
prediction was presented by B. Tirimula Rao et al. For the purpose of forecasting development costs and timelines, Multilayer
perceptron and back propagation are used by G. Witting and G. Finnie4 as learning methods. algorithms. N. Karunanitthi et al.
shown how to use ann to gauge the dependability of software that makes use of several feed- forward and Jordan network
techniques. The training method for ann with back propagation is used by N. Tadayon14. A training set of data and a validation set
of data were created from the same data set, however this was not stated explicitly in the literature. Our goal in this vision is to use
artificial neural networkapproaches to improve the five scaling variables for the COCOMO II estimate model as well as a number of
dependent variables, including the cost drivers. It is difficult to comprehend the findings because neural networks, which act as
"black boxes," have dispersed knowledge [5]. A neuro-fuzzy was suggested by Huang et al. to get past this restriction. Utilizing a
cost model that is constructive and easy to understand to estimate software work [12]. A neuro-fuzzy technique, which has a great
generalisation capacity and performs well with ambiguous and incorrect inputs, is the basis of this model. In simulations, the neuro-
fuzzy model performed better for estimating software effort than the conventional COCOMO method [12].As we can see, there has
been a lot of study on using decision tree-based models to improve SDEE accuracy. An algorithm for back-propagation learning is
a type of artificial neural network. An ANN with a back-propagation algorithm is one of the most important learning algorithms in
use right now. The issue of multiclass classification can also be solved with it. The hidden, output, and input layers make up an
majority of ANNs. The nonlinear activation function argument, or sigmoid function, is specified by the weighted sum of the input
neurons (Baareh et al., 2006). Let y1(p), y2(p),, yn(p) be the network's needed output and x1(p), x2(p),..., xn(p) be its inputs. P
establishes the iteration number. El-Sayyad et al. (2015) provide the following illustration of the back-propagation neural network's
function.
III. PROPOSED METHOD
The dataset is preprocessed so that effort may be calculated using a specific method and a limited number of equations. The
estimated effort is then compared to the anticipated effort that was acquired by applying: Backpropagation, as well as the actual
effort that was attained (Backpropagation). The created data mining models and method are then evaluated using MMRE, MRE and
PRED Performance Evaluation Metrics (x).

A. Dataset (Desharnais)
Jean-Marc Desharnais produced this dataset in 1988 [4]. It is one of the earliest datasets for SDEE. It has so been utilized in
numerous empirical research, including [10] [7] [9]. Dataset with 81 rows and 12 attributes which is of real software projects data
from a software company called canadian make up this dataset. Based on their technological environments, these 81 projects have
been divided into three subgroups: the conventional environment (46 projects), the "improved" traditional environment (25
projects), and the micro environment (10 projects). There are 12 features total in each project, nine of which (Team Experience,
Manager Experience, Length(months), Entities, Transactions, Adjustment, Points Non Adjust, , Points Adjust, and Language(1,2,3))
are independent, and one of which (Effort) is dependent.

Figure2. Attributes from the Dataset

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2859
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

B. Pre-processing of Data and Labelling

Preparing raw data to be used with a machine learning model is known as data pre-processing. In order to build a machine learning
model, it is the first and most important stage. It is necessary to perform these actions in order to clean the data and prepare it for a
machine learning model, which also improves the model's efficacy and accuracy.
There are ten attributes total in the dataset, one of which is effort, which is a dependent attribute. The remaining nine qualities are
independent. First, the dataset is divided in half, with training data going into testing data in a ratio of 70: 30. The training datamake
up 30% of the dataset, while testing data make up 70% of it.
While testing the new instance, the effort attribute is removed from the dataset. The effort is then generated by the algorithm and
verified against the actual effort values to ascertain the algorithm's correctness and the possible range of values for a particular set of
project-related attributes.

Figure3. Flowchart for Effort estimation process

C. Back Propagation Algorithm

Feedforward neural networks are trained using the popular back propagation algorithm. Instead of explicitly computing the gradient
with respect to each individual weight as would be the case naively, Using the network weights as input, it calculates the gradient
of the loss function and is incredibly effective. Due to their effectiveness, gradient methods—including the popular gradient
descent and stochastic gradient descent—can be used to update weights and train multi-layer networks to minimize loss. When
using back propagation algorithm, the loss function's gradient with respect to each weight is computed using the chain rule, layers
by layers, and reiterating the previous iteration layer in order to prevent the computation of intermediate terms in the chain rule
from being repeated.

Mathematical model
1) Equation 1 shows the results of the hidden layer's calculations:

( ) = [∑ ( ) ( )θj
=1
The quantity of input neurons is depicted by the variables which, n, j stands for number of the hidden layer, and wij, which stands
for the estimated mapped weights between the variables Journal of Computer Science, 2019, Volume 15(3), Page 321.331, by
Abdel Karim Baareh. A threshold value is a value that determines whether data is considered to be valid the output layer is reached
after passing through the hidden layer and the input layer.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2860
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

2) Equation 2 shows how the sigmoid function was put into practice:

3) Equation 3 shows the calculated output of an output layers: yk(p)=sigmoid[[∑j=1xjk(p)wjk(p)−θj] where m is the quantity of
input neurons in back propagation and k is the output layer

4) The Gradient of Errors obtained coming from the output layer is represented by Equation 4:
δk(p)=yk(p)[1−yk(p)]ek(p)

5) Equation 5 shows the weights determined by the ANN.:

∆wjk (p)=αyj(p)+δk(p)

6) Use Equation 6 to modify the ANN weights.

wjk(p+1)=wjk(p)+∆wjk(p)

7) Equation 7 represents the hidden layer's estimated gradienterror.: δj(p)=yj(p)[1−yj(p)]∑δk(p)wjk(p)

8) Using Eq. 8, the weights are further computed..
∆wij(p)=αxi(p)+δj(p)

9) The revised weights are represented in Equation 10:

wij(p+1) =wij(p)+∆wij(p)
They entirely overlap on other subspaces and only separate on the two-dimensional space bounded by B+ and B-. Finding this
separation from the N-dimensional data is the objective.

Figure4.Workflow Diagram of Back propagation Algorithm

Supervised learning is a technique used by neural networks to produce output vectors from input vectors on which the network is
based. If the result does not match the created output vector, it compares the generated output to the favoured results, and creates
an errors report. It is the weights then made adjust in accordance with an bug report to produce the results you want.

Algorithm of Back propagation

 Step1: First, X must enter through the pre connected path. Step2: Using true weights W, the input is simulated. Typically,
weights are selected at random.
 Step3: Determine each neuron's output through the hidden layer from the input layer and out towards the output layer.
 Step4: Determine the errors in the layer of output Back propagation Error(diff)= Actual Output - Desired Output
 Step5: Return between the hidden layer and the output layer and change the weights to lower the error.
 Step6: Continue the procedure until the intended result isobtained.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2861
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

Figure5. Flow Chart of Back propagation

D. Multiple Linear Regression model

The statistical approach known as the use of multiple linear regression to forecast the result depends on the values of two or more
other variables other variables. It’s a development of multiple linear regression and is occasionally just referred to as multiple
regression. The factors We employ independent or explanatory factors to forecast the value of the dependent variable,
meanwhile the variable we are trying to forecast isreferred to as the dependent variable.

Formula for multiple linear regression model

Yi=β0+β1xi1+β2xi2+….+βpxip+€
Where as:
 The predicted or dependent variable in this case is yi.
 The value of y where x1 and x2 are both 0 is known as the y- intercept, or β0.
 Y's change in relation to a shift of one unit in either xi[1] or xi[2] is represented by the regression coefficients β1 and β2,
respectively.
 With regard to each independent variable, βp represents the coefficient of slope.
 The model's random error (residual) term is ϵ.

Multiple linear regression model is a statistical tool that uses two or more independent variables to predict the outcome of a
dependent variable. Analysts can use the technique to determine the total variance of the model as well as the proportional
contributions of each independent variable.
This regression analysis model enables higher variance and accuracy when it comes to outcome prediction as well as analyzing the
impact of each explanatory variable on the model's overall variance due to the many variables, which can be linear or nonlinear.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2862
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

Figure 6. Flow Chart of Multiple Linear Regression model

When the Multiple linear regression model involves two or more independent variables and one dependent variable. is used to
assess the connection between them. When seeking information, you can utilize the multiple linear regression model to:
1) How closely two or more independent variables are related. are positively correlated with one dependent variable (for example,
the amount of rainfall, temperature, and Adding fertilizer impacts crop growth).
2) Value of the dependent variable in relation to a specific of the independent variables' values (For instance, the crop's
anticipated yield at specific levels of temperature, rainfall, and fertilizer application).
3) Considering the most relevant attributes, the effort values are calculated. The predicted value of Effort is compared with the
actual effort value for each of the output of the algorithmic model corresponding to its input. Each time the weight vector, is
updated and the output is computed again and compared. The graph of predicted vs actual values
4) The estimated effort is mainly dependent on the attributes which are present in the dataset. Whenever we calculate the
estimated effort we will compare it with the actual effort which is present in the dataset the difference is MRE here.

IV. RESULT
Firstly, only the attributes impacting the effort are considered and rest of the attributes are dropped based on the correlation matrix.
The heatmap of the values of attributes are-

Figure 7: heatmap

The results revealed that the NFR feature "Language" is the most statistically significant attribute, and that removing it from the
regression model's training leads in a 25 percent rise in MMRE.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2863
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

The next statistically significant attribute was "Envergure," which was followed by the less statistically significant attributes "Team
Exp" and "Manager Exp."
Software effort and the attribute "Team Exp" have a negative correlation, whereas "Manager Exp" has a positive correlation.
The findings also demonstrated that the MMRE error increases by100% when the quality standards attributes are disregarded, and
prediction models are created exclusively using software size asan independent.

Figure8.Training and Validation Accuracy Graph

Figure5. Performance Accuracy

A. Accuracy of the trained Model

Accuracy plays an important role to determine whether the trained model will be suitable or whether it provide us with desired
outputs or not. After splitting the datasets training 70% and testing 30% of datasets accuracy is obtained.
Figure9 shows the training and validation accuracy graph obtained by my proposed scheme. According to the observations,
validation accuracy occurs to be less than training accuracy. This can be improved by reshuffling the validation set. The model
should be trained more for validation samples as the datasets in that case are still rising and falling for some epochs.

Figure9.Training and Validation Loss Graph

In Figure 7, Representation of training and validation loss graph is shown. It is observed that there is higher validation loss than
training loss, it represents that my model is overfitting. It learns superstition which means that patterns in my datasets of training
data but in reality, it’s not the case and hence it is not true for validation data.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2864
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

Figure10. Comparative results

The MMRE and MDMRE accuracy of the trained model by thevalidated datasets are-

Figure 11. Training and Validation Accuracy

V. CONCLUSION
This system designed is based on various neural networktechniques that have been used to effort estimation. Every technique
aims to provide the most accurate software effort estimation. In this research, we suggest that a good method for calculating
software development effort is the back propagation algorithm, which is an ANN model. It was advised to adopt the back
propagation strategy, which will quickly propagate errors, for complex and computationally intensive tasks, where the outcomes of
the multilinear regression were contrasted however, it is necessary to assess the approaches' correctness since they are mostly
needed in software work estimation. We found that neuron-based models are more accurate estimators and can therefore be utilized
to determine software effort estimation for alltypes of projects.

VI. FUTURE ENHANCEMENT

Future editions of this model will incorporate the suggested model along with some tuning strategies to deal with values that are
accompanied by confusing and unclear facts. expanding it to a strategy that is likely to be a future study area Future synthetic data
generated using well-known approaches will be used to train and evaluate the suggested model, resulting in increased accuracy over
current methods. By fine-tuning hyper parameters and using optimization techniques, the performance of the model can be
improved. This will result in a successful software development project using estimating approaches, therefore in the future, the
estimation of the effort in development of software projects can be done by using other important attributes.

REFERENCES
[1] S.G MacDonell, “A comparison of modeling techniques for software development effort prediction” ,in :Proceedings of the International Conference on Neural
Information Processing and Intelligent Information Systems ,Dunedin, New Zealand, Springer, Berlin, 1997, pp. 869–872.
[2] K. Strike, K. El-Emam, “Software cost estimation with incomplete data”, IEEE Transactions on Software Engineering, 27(10) 2001.
[3] A. C. Hodgkinson, and P. W. Garratt, “A neuro fuzzy cost estimator”, in: Proceedings of the Third International Conference onSoftware Engineering and
Applications—SAE 1999, pp. 401–
[4] C. Kirsopp, M. J. Shepperd, and J. Hart, “Search heuristics, casebased reasoning and software project effort prediction",Genetic and Evolutionary
Computation Conference (GECCO), New York,AAAI, 2002.
[5] X. Huang, D. Ho, J. Ren and L. F. Capretz, "Improving the COCOMO model using a neuro-fuzzy approach". Applied Soft Computing, vol. 7, pp, 29-40,
January 2007.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2865
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue IV Apr 2023- Available at www.ijraset.com

[6] P. L. Braga, A. L. I. Oliveira, G. H. T. Ribeiro and S. R. L. Meira. “Bagging predictors for estimation of software project effort”. In: IEEE/INNS
International Joint Conference on Neural Networks, IJCNN’2007, Orlando Florida.
[7] R. Babuska ,Fuzzy Modeling for Control, Kluwer AcademicPublishers, Dordrecht, 1999
[8] A. A. Porter and R. W. Selby. Evaluating techniques for generating metric-based classification trees. J. Syst. Softw., 12(3):209–218, July1990
[9] Krishnamoorthy Srinivasan and Douglas Fisher. Machine learning approaches to estimating software development effort. IEEE Trans. Softw. Eng., 21(2):126–
137, February 1995.
[10] Sutton S Informed projection Proceedings of the 2018 InternationalConference on Software and System Process, (76-85).
[11] Tanveer B, Vollmer A and Braun S A hybrid methodology for effort estimation in Agile development Proceedings of the 2018 International Conference on
Software and System Process.
[12] Boehm B Software cost estimation meets software diversity Proceedings of the 39th International Conference on Software Engineering Companion. A. A.
Porter and R. W. Selby. Evaluating techniques for generating metric-based classification trees. J. Syst. Softw., 12(3):209–218, July1990.
[13] Barry W. Boehm, “ Software Engineering Economics”, prentice– Hall , Inc., Englewood Cliffs, New Jersey 07632, 1981, ISBN0 – 13 –822122 – 7.[7]
[14] S. H. Lee, H. Goëau, P. Bonnet, and A. Joly, "New perspectives on plant disease characterization based on deep learning," Comput. Electron. Agricult., vol. 170,
Mar. 2020, Art. no. 105220.
[15] Z.-W. Hu, H. Yang, J.-M. Huang, and Q.-Q. Xie, "Fine-grained tomato disease recognition based on attention residual mechanism," J. South China Agricult.
Univ., vol. 40, no. 6, pp. 124132, Jul. 2019.
[16] Rahul Kumar Yadav,Dr. S. Niranjan,Software Effort Estimation Using Fuzzy Logic: A Review, International Journal of Engineering Research & Technology
(IJERT) Vol. 2 Issue 5, pg.1377-1384, May2013. ISSN: 2278-0181.
[17] ImanAttarzadeh and Siew Hock Ow, Software Development Effort Estimation based on a New Fuzzy Logic Model,International Journal of Computer
Theory and Engineering, Vol 1. No. 4 October2009.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2866