Sensitivity-Based Dimensionality Reduction for Surrogate-Assisted Many-Objective Large-Scale Project Scheduling

In large-scale stochastic many-objective time–cost–environment trade-off problems, influence is unevenly distributed: each objective is governed by a reduced subset of the decision variables that drive most of the controllable variation, while the remaining variables enlarge the Pareto search space—and the set of levers managers must deliberate over—without materially affecting performance. This study integrates Monte Carlo simulation, surrogate modeling, Sobol global sensitivity analysis, and many-objective evolutionary optimization to identify that influential subspace and confine the search to it, demonstrated on a 30-task construction project with 68 input variables, five conflicting objectives, and 150,000 training scenarios. Multi-criteria benchmarking of eleven surrogate architectures and eight evolutionary algorithms selected a Multilayer Perceptron (R2 = 0.993, MAE = 3.962 × 10−3) and AGE-MOEA (HV = 0.741, IGD = 0.064, SP = 0.036). Sobol screening reduced the decision space from 60 to 40 task-interpretable variables; at equal budget, the reduced space retains 91.6–95.5% of the full-space hypervolume, a front-level cost that exceeds the discarded sensitivity mass (0.3–2.2%) and quantifies, for this instance, the behaviour of the Factor Fixing criterion at the front level. A space budget analysis showed that both spaces converge well before the full budget (2.0–2.2× speedup at no more than 1.7% additional hypervolume loss): the efficiency gain stems from the budget, while the screening contribution is structural: it identifies which scheduling decisions can be fixed at nominal values, at the cost measured above. An a posteriori re-evaluation with the exact model bounded the optimistic surrogate error (MAPE 1.26–4.38%; 118 of 126 solutions feasible), and the CRITIC-weighted selection yielded a 36.46-day schedule at USD 283,189 with a sustainability index of 0.924.