Classification based cost estimation model for embedded software

Abstract

Software development companies face many problems in order to complete their projects successfully: on time, within budget and with no defects. Scheduling and resource allocation directly affect financial performance and market position of a software company. The challenge is how the project managers will decide what level of skill set, for how long and at what cost they will need for a given project. Therefore, practitioners increasingly need intelligent oracles to help them make these decisions. These oracles can be defined as the learning based prediction models for effort and cost estimation. Such predictive models prevent project managers to take wrong decisions due to inaccurate estimations. In this research, we focus on building learning based predictive models for cost estimation in embedded systems domain. Our proposed model tackles the prediction accuracy problem from both data usage and model development aspects. Firstly, we focus on data usage and investigate what kind of and how much training data should be used for software cost estimation in embedded systems. Secondly, we focus on model development and propose a new cost estimation model for embedded software. We believe that our results would assist the software managers while selecting the data to train the cost models and allocating available resources more efficiently by using more accurate analysis. In literature, there has not been any study that focused on embedded software cost estimation yet. We aim to fill in this gap for embedded systems domain. In addition, we present a new cost estimation model which achieves high accuracy rates. In our empirical work, we utilize a wide range of machine learning techniques in order to make our results be independent from the techniques used. Also, we used datasets from three different sources in order to be able to generalize our results under different set-ups.