Machine learning models for predicting software design effort

Abstract

Software design is a distinctive activity within the software development life cycle (SDLC). It is typically undertaken by an independent, specialized team, whose budget relies on the required person-hours prediction (i.e., effort). An over-prediction could cause project rejection before starting, whereas an under-prediction may result project cancellation before completion. A common practice for predicting the effort percentage by activity involves calculating it from the total SDLC effort. However, the reported design effort using this method varies widely from 0.62% to 50.35%. Another practice involves using prediction models; however, systematic literature reviews published up to 2025 indicate the absence of models specifically applied for software design effort prediction (SDEP) by existing only models addressing the total SDLC effort. Thus, the present study applies two models to SDEP reported as most accurate in the effort prediction field: support vector regression (SVR), and Multi-layer perceptron (MLP) neural network. Their parameters are optimized through genetic algorithms, and their performance is compared to that of a statistical regression model (SRM). All models were trained on seven data sets selected from an international public repository of software projects used in dozens of studies on software effort prediction. Results show that SVR performed statistically better than SRM in five data sets and equally on the remaining two. MLP outperformed SRM on three data sets and equally on the resting four. Consequently, both MLP and SVR can be used to SDEP.