N0 | Method | Type | Strengths | Weakness | Accuracy | Data | Validation |
1 | COCOMO- II | AM | Simple to carryout estimations; takes less time and effort to estimate; Useful in large projects | Details of the past projects needed to estimate; May leave out hidden costs hence lead to more expensive estimation; Calibration is required; Cannot meet current software standards | MA | ED | EV |
2 | FPA | AM | Easy to estimate development costs at the requirements gathering stage/initial stage; Independent of languages and tools used | Quality attributes, development time and man power are not considered | MA | LD | EV |
3 | PM | AM | Estimation depends on only time and size which are key to cost estimation; Needs fewer parameters compared to COCOMO 81 and COCOMO II | Does not consider other important aspects of the software development | MA | LD | EV |
4 | EJ | NA | Best technique where limited data is available; Experience of experts makes the estimation more accurate and realistic | Biased opinions of the experts; Difficult to document the parameters used by the experts to estimate; Experts require experience of similar projects | LA | LD | EV |
5 | TDE | NA | Only few details are required to estimate; Simple and less time consuming | Difficult to identify low level problems which causes under estimation; Less details may overlook important attributes of the project | LA | LD | n/a |
6 | BUE | NA | More stable compared to top down approach; Errors are estimated and very stable | Development time and system-level activities are not considered | LA | LD | n/a |
7 | PTWE | NA | It depends only on the customer budget | Costs do not accurately reflect the work required; Low quality system is developed due to client budget constraint | LA | ND | n/a |
8 | NN | LM | Very good in capturing non-linear relationships; Deep neural Nets do not require a lot of features, they come up on their own; There is a lot of flexibility, you could choose different architectures; RNN’s, LSTM, etc. | They tend to over fit; They require enormous amount of computation power | VHA | ED | NI, DCS, CVM |
9 | GA | LM | Need not be supervised | Too many hyper parameters | HA | ED | CVM, VAF, RWS |
10 | FL | LM | Considers real valued states instead of binary | Does not perform well on complex datasets like COCOMO | VHA | ED | JM, DCS, CVM |
11 | BN | LM | Bayesian network encodes all variables; missing data entries can be handled | Difficult to model | HA | ED | CVM |
12 | SVR | LM | It works best with text data: string kernel | It takes a lot of memory (RAM); Doesn’t scale well when dataset is huge | HA | ED | LOM |
13 | RT | LM | more capable of handling noisy datasets | Models are unstable at times, suffer with high variance, low bias. (Keyword: bias-variance tradeoff) | MA | LD | CVM, DCS, LOM |
14 | ABE | NA | Simple and easy to use; No bootstrap cost | They do not work with categorical data | MA | LD | EV, ED |