<![CDATA[Purpose: Software development effort estimation (SDEE) frequently suffers from accuracy degradation due to redundant project attributes in medium-sized datasets such as NASA93. Most metaheuristic optimization studies focus solely on hyperparameter tuning without filtering input dimensions, leaving the model vulnerable to overfitting. Methods/Study design/approach: This study proposes a Hybrid Local Search-Particle Swarm Optimization framework integrated with Support Vector Regression (HPSO-SVR). Instead of relying on a sequential approach, the standard PSO algorithm is modified by embedding a Simulated Annealing (SA) mechanism as a local search operator. This allows the algorithm to evaluate the feasibility of binary feature subsets and continuous parameters (C, gamma, epsilon) concurrently within a single vector space. The model is tested using a nested 10-fold cross-validation protocol on the NASA93 dataset to prevent data leakage. Result/Findings: The HPSO-SVR model successfully identified and eliminated 7 disruptive attributes (including DOCU, RUSE, and SCED), reducing the dimensionality from 22 to 15 features. This approach yielded an MMRE of 22.15%, a Pred(25) of 63.44%, and an RMSE of 38.90 PM. Statistical validation using the Wilcoxon Signed-Rank Test proved that the addition of the local search mechanism for feature selection provided a significant improvement (p < 0.05) over standard PSO-SVR. Novelty/Originality/Value: The primary contribution lies in the application of a probability-based SA local search mechanism embedded within the PSO population structure. This successfully suppresses COCOMO II feature noise simultaneously with SVR tuning, yielding a more compact and accurate predictor configuration for early-stage project planning.]]>
Copyrights © 2026
