Evaluation of an Interval-Based Fitness Function within SOMA for COCOMO-Based Effort Estimation

This study introduces a fitness function, Predictions Within Individuals’ Intervals (PII [Lo, Hi] ), designed to optimize software effort estimation (SEE) tasks within the Self-Organizing Migrating Algorithm (SOMA). The effectiveness of the PII [Lo, Hi] fitness function is evaluated and compared against well-known fitness functions, including MAE, MSE, MMRE, PRED(0.25), and MMER. In addition to fitness-function-level comparisons, the study further benchmarks models calibrated using the PII [Lo, Hi] fitness function against Quantile Regression (QR) and Quantile Regression with Gradient Boosting (QRGB). The evaluation uses the Constructive Cost Model (COCOMO) on the NASA18, Kemerer, Miyazaki94, and ISBSG datasets. The study employs evaluation metrics, statistical testing, residual analysis, and convergence analysis to validate the results. The findings show that MAE and PII [-10, 25] produced stable and accurate predictions, while MSE and MMER demonstrated higher variability. Residual and convergence analyses further support the robustness of PII [-10, 25] , though its performance declined in one scenario (TEST 3), warranting further refinement. The results were also supported by the Wilcoxon signed rank test and Win-tie-loss performance analyses. Future research will focus on extending the applicability of the PII [Lo, Hi] fitness function and exploring its interaction with different optimization algorithms, as well as investigating the potential of automated parameter tuning.