Due to diverse functional aspects in software systems, proper metric systems remain to be a major concern in software engineering. Therefore, software engineers focus on measuring the functionality size in software development projects. In reducing the complex task of software metrics in terms of functional size, functional point analysis method was developed [13] . Functional point analysis refers to the standardized methods of determining software sizes by using functional constraints which determine the key features to be designed [14] . Significantly, this method is universal because its application is not limited to programming languages and technologies. In function point analysis, two major components are measured. The most crucial aspects of software application measure in function point analysis comprise the data functionality and transaction functionality attributes [13] . Precisely, the metrics of the two basic features are determined by evaluating the scope of the system product, quality indicators, productivity, and the system performance.

Function point analysis (FPA) evaluates the system’s metrics from a functional perspective, thereby resolving issues associated with technology dependency in the development lifecycle [13] . The efficiency of FPA in software engineering is achieved through a comprehensive analysis of applications in three stages [13] . The first stage of function point analysis concerns identifying the forms of transactions to be made in the software applications. Secondly, the engineers evaluate and appraise the components of the software system. Lastly, the process involves evaluations of the general system characteristics. Fundamentally, general system characteristics are categorized using into 14 main features [13] . These are data processing, system performance, hardware configurations, transaction rates, data entry, end-user efficiency, online updates, reusability, ease of use, support of multiples sites and change facilitations [13] . There has also been tremendous research in enhancing function point analysis with non-functional requirements.

The Putnam’s model is a dynamic multivariate model which was developed by Larry Putnam in the late 1970s for effort estimation [15] . The model functions by examining the many software projects and analyzing the distribution of manpower. The relationship between the size and effort is non-linear and is highly sensitive to delivery time. It provides a simple and computationally plausible way of predicting software costs. It is used to calculate both effort and time required to complete a software project based upon the specified size of the project. The Putnam’s model equation is given as [15] :

B 1 / 3 ∗ Size Productivity = Effort 1 / 3 ∗ Time 4 / 3 (3)

With this model, SLIM is the tool that is useful when it comes to cost estimation and allows workforce scheduling. Thus, it achieves the objectives of cost estimation as it provides the schedule and cost of the essential resources. However, it’s capacity in estimating the total manpower requirements and development time at an early stage is still not satisfactory [16] . It has the setback of not accounting for other aspects of the software project [16] . There are a series of aspects relating to software development especially in its life cycle that needs to be considered. The uncertainty in the size of the software may lead to the inaccurate cost estimation. According to [17] SLIM’s error percentage is said to be 772.87%. On the contrary, its advantage is that the model is based on two variables that are keys to cost estimation which are time and size [16] and it needs fewer parameters compared to COCOMO 81 and COCOMO II.

Expert judgment is one of the traditional techniques that are used in the software cost estimation in the early phases of the software development [3] . The technique is such that it relies heavily on the expertise and the experience of an expert at cost estimating. It depends on the domain knowledge of the expert rather than the historical data [3] . The experienced estimator is tasked with the responsibility of estimating the cost of software based on the fact that they have sufficient knowledge that ensures cost estimation is as accurate as possible [3] . On the same note, the expert judgement used to estimate cost on a given project is often limited to a specific filed of expertise of the estimator. An expert who has knowledge and experience on a specific field is more likely to have majored in the given project putting them in a better position to estimate the cost. Expert judgement comes in handy especially when there are limitations that limit effective and efficient data collection [18] . In doing so, expert judgement is required to make decisions based on the limitations and the stringent factors. Delphi technique is one of the examples that follow the expert judgement approach. An expert can be able to offer an honest and experienced opinion on the best course of action when it comes to an understanding of the impacts of a system being incorporated. There is also a setback of tedious processes that are undertaken in documenting the factors that an expert requires to make the judgement [3] . Most of the factors in many software projects are many and documenting them is tedious as well as difficult. There is no specific validation for this approach as the estimation depends solely on the domain knowledge and previous experiences of the expert.

Additionally, there is the disadvantage of obtaining cost estimates that are biased and optimistic. Expert judgement is given by experts who have human emotions and is very likely that their emotions influence the judgement process [3] [18] . Therefore, there is the possibility that pessimism, optimism, as well as bias might influence the judgement process.

Top Down cost estimation approach focuses on estimating the cost of a project from the global properties of the overall project and using either algorithmic such as Putnam model or non-algorithmic methods [19] . The estimation is then split into various components in proportion [19] . This method can be followed when there is limited historical data available about the similar project. This technique is more beneficial while the project is still in its early stages. This is because, at this stage, there is no need for detailed information about the project [19] .

Top Down estimation is used for high-level decisions when the planning horizon is quite long. This technique is used when there is very little specific project information where we can get a ballpark estimate. Unlike other cost estimation techniques, this approach focuses on activities like management and integration which are overlooked in other techniques. The major disadvantage is that the ballpark estimate is very inaccurate [19] .

This is the exact opposite of the top-down estimation methodology. In this particular technique, the cost of every component of the software is derived and then the final result is obtained by combining these elements to get the total estimated cost of the project [19] . The aim of this technique is to obtain a proper estimate that will be an accumulation of the estimates of the smaller components of the software. Depending upon the variety of the projects both the methodologies are useful [19] . The estimation methodology is best suited for small projects to estimate the cost.

In the bottom-up estimation, we take each work package or each activity in a schedule and assign a dollar cost to that and then add them up to get the overall cost of the project [19] . This can create a very detailed estimate that will be very accurate. However, there are some disadvantages of the bottom-up estimate.

1) If we add up each activity, we may be lacking any coordination between the activities such as resource overlap or if some of the dependencies that one activity has to be finished before another can start.

2) The accuracy is high, but it can also take a lot of time to create this level of estimate, which means that it can be expensive to create.

So, to overcome these disadvantages it is preferable to do a high-level estimate i.e. top-down estimate for the whole project and then do some detailed estimates for the activities that are coming up in the short term.

In this approach, the estimation of the software project is directly proportional to the budget of the customer. The customer’s budget is more focused rather than the functionality of the software and the project costs whatever the customer has to spend on it [19] . This approach is not so recommended as instead of focusing on software functionality it focuses more on the client’s budget and capacity. Accuracy varies drastically based on the client’s budget and therefore it is rated as low accuracy. This approach does require extensive data or no previous data at all as the client will convey the requirements and does not depend on the historical data [19] . It is not a good practice as it may cause a delay in the development and delivery and may also force the development team to work overtime. The validation of this approach is based on the customer’s budget and person-month factor.

On the contrary, the major advantage of the price-to-win approach is that it the estimation doesn’t exceed the customer’s budget. And as the functionality of the software is restricted with the customer’s budget the quality of the software is compromised [19] .

ANN is one of the major approaches that are exploited in the sector of machine learning models. As the name suggests, it is usually inspired by the neural part of the brain system with an intention of imitating an intelligent living organism [27] . It is composed of two layers, that is the input and output layers, within the layers there is a hidden layer which constitutes of units whose main purpose is to assign weights to the data that is fed from the input. These weights are assigned randomly to the data [28] . They are among the exploited tools in cost estimation due to their excellent performance [1] . Neural networks have existed for long as they can be traced back to the year 1940 although the best way to exploit them was the missing link which has been currently accomplished. There are different types of neural networks that exist with each of them having a defined level of complexity coupled with use. The most common and general type of neural network (NN) is the feed forward neural network as the name suggests data travels in only one direction which is from input to output [27] . Other types include Recurrent Neural Network (RN), Convolution Neural Networks (CNN), LTSM Recurrent Neural Networks, etc. The ANN approach is used in cost estimation where the pattern of certain data requires to be understood prior to estimating the project cost. In the cost estimating field, it is used to classify data into predefined classes, clustering, and prediction [28] . For instance, in projects such as the stock exchange market, it is used to make predictions in the market through the exploitation of the previous data.

Artificial Neural Networks are usually excellent in capturing non-linear relationships which makes it an iconic advantage of the model. The model has a deep neural net which requires fewer features [27] . The neural net has the ability to develop its own features which are also an advantage when limited features are available in the dataset for training the model. It is also flexible in the sense that there are a plethora of features one can choose from which include CNN’s, RNN’s, LSTM RNN’s, etc. (Figure 3).

It has also some limitations whereby sometimes they tend to over-fit the data in the software cost estimation process [2] . Additionally, the model also requires an enormous amount of power for computation which may not be available at all times. It was better if the model is in a position to efficiently operate even in places where there is less consumption power as putting up the high computation power is costly to the involving firm.

Several Artificial Neural Networks models have been used in the software cost estimation processes which are mostly common according to the investigation [9] :

a) Feed-forward neural network

b) Recurrent Neural Network

c) Radial basis function (RBF) network

d) Neuro-fuzzy networks

Hamza et al. [9] concluded that choosing the right artificial neural network model is essential to get the accurate estimations. In their study [9] they have also concluded that feed-forward neural network works better than other models in ANN’s but, the technique needs data filtering to prevent noise. In the case of noisy data, the radial basis function is more suitable.

They are defined as adaptive and heuristic search algorithms which are a subject to the theory of natural selection by Darwin. In a current, study they are described as one of the most active areas of research which have been designed through nature-inspired metaheuristics [29] . Genetic Algorithm (GA) forms one of the soft computing techniques in software cost estimation process whereby its main role is to change certain parameters of classical methods such as COCOMO approach to predict software cost in a more accurate manner [29] . GA has widely been exploited in different fields of cost estimation such as correcting the identification system, path-searching problems within a project [29] . Additionally, GA has been used to solve a variety of NP-hard computational problems [29] . This model takes inspiration from nature such as bullet train design based on fish. Used for optimization problems (Np-problem) few of them are firefly algorithm particle swarm optimization and cuckoo search just to mention a few.

GA model usually exploits the optimization problem by using an evolutionary process. The first benefit of the model is the fact that it is easy to set up than neural networks but is less flexible. Once the algorithm begins it is on its own. It learns its own features; hence we don’t have to supervise the process. On the other hand, it has a disadvantage which includes less flexibility, many hyper parameters which includes preference of functions, reproduction rates, the percentage of elitism and cross over, dealing with out of bound conditions, creating a strategy and setting the required tree sizes and depths within the model [30] . The model has no guarantee of finding an optimal solution, infinite time because it has asymptotic convergence, containing a number of parameters, sometimes the result is highly dependent on the parameters set. It has also self-adaptive parameters. It is computationally expensive and Meta models of functions are used in this process too.

The original use of the model was to establish manpower required to complete a certain project [29] . It is also used in scheduling different tasks within the field of cost estimation which implies that it has the ability to determine and evaluate a variety of tasks within a single project. It has also been used in mining data within the same scope which is considered as a complex exercise in case traditional methods are exploited in place. On the extreme end GA has been used in optimizing distributed tasks within the software cost estimation process. It has been used to solve distributed queries which imply that all relevant queries about a certain project can be accessed during the initial steps of its development. It is also easy to make assumption and predictions to the software being developed while using this method. It is somehow time conservative especially when operated by experienced personnel.

The model is a computing approach that is based on the degrees of truth rather than the unusual true or false normally referred to as Boolean logic which most of the modern computers are based on [31] . Additionally, it can be said to be an approach to computing based on many-valued function for instance, instead of a task completed or not, one can say 50% is completed. Fuzzy Logic (FL) approach gives an acceptable but definite output which is usually in response to inaccurate (fuzzy), distorted, incomplete and ambiguous input. FL was developed in 1965 by Loft Zadeh from a concept of fuzzy set theory. The first system is the one that is exploited in the estimation sector as most of the systems produce crisps data as input and expect the same type of data as output. There are three steps that have to be followed while using FL; the first is the Fuzzification which converts a crisp into a fuzzy set [31] . The second is the Fuzzy Rule-based System, at the step after all the crisp input has been fuzzified into their respective linguistic values the inference engine then derives their linguistic values [32] . Defuzzification is the final step which involves the conversion of fuzzy output into crisp output.

Fuzzy logic is deployed for decision making whereby it can be implemented with various sizes and abilities ranging from small microcontrollers to large workstation-station based software development. In the sector of software cost estimation, FL has been used to give acceptable reasoning although it does not guarantee accurate reasoning [32] . This approach is very easy to use despite the fact that it has numerous functions that it can execute when it comes to cost estimation of the software development [32] . Also, with the approach, it is advantageous as estimators of software are in a position to estimate and give an anticipation of the different aspects of the project even before the initiation level. The method has also a fast learning ability compared to other methods under the same scope [31] . The major disadvantage that the approach lacks guarantee of accuracy which is a vital part of the software cost estimation as to avoid incurring financial losses. Through the defuzzification process, the model is able to interact the output into a numerical value as per the desire.

There are a number of three applied approaches under this method which include; 2FA-kmodes it is used in clustering where numerical datasets are shown through fuzzy sets however the fuzz k-modes algorithm is used to maintain categorical attributes. The second approach is 2FA-kprototype it is used in clustering also where project data are clustered into homogeneous sets [31] . The final approach is the classical analogy it is used to predict the effort of the target project using information from former similar projects. The first two models 2FA-knodes, 2FA-kprototypes perform same and are both better than classical analogy. The similarity in the functionality of the two models is technically for the purposes of increasing efficiency within the model [32] . The major disadvantage associated with the two models is that they are not very good in COCOMO dataset and classical analogy doesn’t recognize categories as “less risky”, “medium”, and “high”. Thus, there is a need for implementations to be conducted for the model so as to enable future proper operation in case they are deployed in COCOMO and other datasets.

The model is also referred to as a probabilistic directed acyclic graphical model. It exploits graphical models to represent sets of variables coupled with their conditional dependencies through a directed acyclic graph (DAG) [33] . In other terms it uses Bayesian inference to perform probability computations they aim at modeling conditional dependence while developing software cost estimation which is usually represented by the edges within a directed graph. In addition to that, the model uses three main inference tasks: inferring unobserved variables, parameter learning and structure learning [33] . Through developers of software understanding the different relationships that exist in the model, they can efficiently conduct inference on random variables within software [33] . Each edge in the model corresponds to a conditional dependency while at the same time each node corresponds to a unique random variable. Thus, a specific pattern is usually followed in the model so as to attain its functionality within the software cost estimation process.

In a recent study, results have indicated that the inclusion of Bayesian Network, a machine-learning model in software cost estimation models it could improve on the levels of accuracy in the project [4] . This approach is able to break down project data about the cost into a form that is easy to analyze while at the same time it offers cost items coupled with the probability which can capture the uncertainty of different items within the software to be developed [33] . In software cost estimation sector, it is exploited in initial planning also as it has the ability to evaluate different parameters of planning such as time, cost coupled with resources to be exploited in the development of software. This model only allows exactly tested cost estimation information to be integrated with natural master feelings.

The best advantage with exploiting this model is that it gives a better assurance on accuracy and it is also able to improve on the overall quality of software that is being developed. Most of the software cost estimation models have some difficulties in offering assurance which makes the model outstanding compared to the rest. It is also a model that saves time which is of great importance when it comes to software project development [4] . Since the model encodes all variables, then it is also able to handle different types of missing data. In case it is deployed learning casual relationships, they help better understand a problem domain as well as forecast the consequences should there be an intervention [4] . Through probabilistic and casual semantics, it is ideal to use the model for representing prior data and knowledge.

Despite it being a complex approach in software cost estimation it is easy to use even with staffs who have less experience in the field [33] . On the other hand, there are some disadvantages while using this model such as Information theoretically infeasible it turns out that specifying a prior is extremely difficult challenge. Thus, being in a position to efficiently exploit this model, then users have to be well familiarized with the model’s mode of learning. Being in a position to understand different languages, the model is able to execute more functions which are advantageous to the project running. Acquiring this knowledge takes some significant effort, this implies a lot of time will have to be consumed during the training period. Despite a lot of time being spent on training, also on the other hand it has to be the required training hence improper training would ruin the entire model. The models also do not always tell the truth as, they don’t specify their actual prior, but they give credentials to the convenient one.

It is also referred to as support vector networks or machines which are supervised models which are capable of learning algorithms and finally analyzing the same data that will be exploited for regression and classification analysis [34] . In other description, it is said to be a concept a set of related supervised learning methods usually used to analyze data coupled with recognized patterns. This model when applied in software cost estimation it takes a set of input data then it gives a prediction of each input [4] . The input must be a member of the support vector machine (SVM) which makes the model to be a non-probabilistic binary linear classifier [34] . In software cost estimation, it is used to solve problems related to pattern classification within the software. So as to apply it effectively in software cost estimation developers have to be in a position to design questions based on a problem and the design involved within the software.

The model has over the year’s utilized regression and classification as the main mode of analyzing data within software. In China, the model was famously used in predicting software development efforts [34] . Support vector reasoning has been found to be helpful in hypertext and text categorization within the software. This model is usually utilized in software cost estimation so as to improve on the levels of accuracy compared to the classical query refinement schemes. It is also exploited in image segmentation in different software depending on the role that the software is designed to perform [34] . With the model in operation, the software development process automatically avoids over fitting of data. Thus, we can technically say that the model is conservative when it comes to financial and time while planning for a project.

A major advantage of the model is that it has the ability to work best with text data known as string kernel [4] . On the other hand, there it has been found to be occupying a huge memory and it does not operate well when the data set is huge. Occupying a huge memory concerning storage of data sets makes it inefficient a factor that they need to reconsider. The fact that it is occupying a huge data set for storage makes is not being able to meet the cost-effective criterion which should be a model for each model. In this model, there are implementations and improvements which have been done on it such as dimensionality reduction such as PCA and ICA. The implementation has been done on the model so as to improve on the functionality of the model.

It is also referred to as a decision tree, is a logical model most exploited for decision analysis in software cost estimation. It is a common approach exploited in software cost estimation based on a number of factors and their consequences [35] . In other descriptions, the model is perceived as a procedure that is used for classification and regression in the sector of software cost estimation process [35] . The model is usually in the form of a tree structure with different internal nodes that stands for a test on an attribute. Individual branch on the tree usually represents an outcome of the test and the class label is held on the leaf node. Looking at the tree from a technical perspective, it is evident that all the branches stand for a specific function within the software development process. The model has usually trained through top-down training method which is specific to a certain direction [35] . However, current research shows there are efforts that are aimed at training the model so as to operate in a multidirectional purpose. This will be a breakthrough within the sector of software cost estimation as it means the model will be able to execute more functions than it does as for now [35] . Over the years it has been used to predict the amounts of effort required to develop a software system which has proved to be a good model as it saves on time and budgets that have been planned on. Thus, we can technically say that the model is cost efficient as it saves on the factors that are likely to jeopardize a developing project.

The regression tree has been widely used in inductive learning, in turn, it has proved to be good in terms of predictive accuracy especially in software cost estimation process despite its complexity [35] . One of the main advantages is the fact that it is fast compared to neural networks and support vector reasoning. The efficiency that is instilled in the model is because it has a simple structure which executes diverse roles is simple steps. Also, while exploiting the model they tend to be very interpretable thus one can almost make an anticipation of what is to be formed in the final project. This implies that there are possibilities of making assumptions on the outcome of the project that is being worked on.

The analogy was used in the past as a method of supporting or supporting explanations of a particular natural phenomenon which was later incorporated in the software development sector of software development of for the purposes of software cost estimations [36] . Analogy Based is one of the most efficient approaches in software cost estimation due to its outstanding performance and capability of handling complex datasets [37] . The model exploits comparison as the main form of a subject to compare software project under considerations with past historical projects which have prior known characteristics, schedule and efforts. Thus, this model has to rely on previous projects similar to the ones that are to be developed thus the levels of accuracy are usually very high [37] .

Conventional analogy-based models usually deploy the same number of analogies in all the projects that it is involved with similar data sets which improves on the estimation approximations. There are some researchers that claim using the same number of analogies could only improve on the data sets only and not the entire project. Thus, when exploiting this model, it is important to understand the

characteristics of each data sets so as to be able to discover the optimum set of analogies for each project [38] . Accuracy levels can only be increased when users of the model are aware of the different data characteristics instilled by each data set. Through the comparison of various datasets, the model is used to approximate the time that a project would take for it to be completed. In cost estimation, the model is relied on to give empirical evidence on the probability of certain software features being accurate.

It is also flexible and intuitive in nature as it can be applied in a variety of circumstances where other algorithmic modeling and estimating techniques don’t operate [36] . As much as it is said to be dependent upon past history so as to execute its functions, it can still operate even where past data is not available. However, there are still some setbacks while using the model as it lacks appropriate analogs [38] . For instance, despite it being widely used in the industries it has not stipulated standards on how it should be exploited for expert opinion-based opinion. The common advantage that most users of the method are familiar with is the ability of the model to avoid bootstrapping of cost [38] . On the extreme end, it is accompanied by a major disadvantage which is its inability to operate with categorical data. Hence more implementations need to be done on the model so as to enable to operate with all types of data (Table 1).