Perkiraan Biaya Perangkat Lunak dengan Jaringan Saraf Tiruan Propagasi Balik
MetadataShow full item record
In line with the advance of technology, human needs for software are also growing. Software is used to help people to lighten their work. This has caused the software becomes increasingly complex and the production cost also become increasingly expensive. No wonder if the software developers need a method of cost estimation that estimates accurately for the efficiency and effectiveness of the company. With this, the company can be protected from loss due to incorrect cost estimates. This study uses data from the COCOMO 81 derived from the Boehm journals in 1981, entitled Software engineering economics. This data group consists of 63 projects each have 17 main attributes, namely, a numerical type of 15 Cost Driver (CD), the size of the project in Kilo Lines of Code (KLOC), and the actual cost of the project in the Person of the Month (PM). Data are grouped into two parts, namely training data and test data in the form of 15 types of Cost Drivers and the size of the project (KLOC), this data is transformed into the range [0.1, 0.9] according to the sigmoid function of range. In this research, the neural network is used to do the software cost estimation. The network architecture used was the multilayer perceptron with a back propagation learning algorithm. The number of neuron in the hidden layer is determined from the optimal architecture which provides a best estimation value. Once found the best architecture which is has the optimal number of neuron, conducted the evaluation. The evaluation method used is the Mean Magnitude Error (MRE) and Pred (l). MRE is the error between the actual and estimated value while Pred (l) is prediction at the level L. Results of research indicate that there are relation between the types of project with the MRE. There is a tendency for this technique provides a high accuracy if the actual value of the test data is big. Poor accuracy obtained at the small value of actual test data. This research resulted in the value of Mean MRE (MMRE) of 76%, Median MRE (MRE Med) of 23% and Pred (0.25) of 52%. On this research, normalization of the target data has not been conducted so that the error generated in the process of evaluation is big.
- UT - Computer Science