Reinforcement learning for sustainable construction scheduling: a many-objective framework with adaptive search and multi-criteria decision support

Purpose Delivering construction projects sustainably requires reconciling duration, cost, carbon emissions and the stability and utilization of site resources, yet these objectives conflict and the choice among candidate schedules is seldom made on an explicit, data-driven basis. This study aims to develop a data-driven decision-support approach that couples many-objective schedule optimization with multi-criteria selection so that project teams can expose and navigate these tradeoffs during delivery. Design/methodology/approach The scheduling decision is formulated as a five-objective multi-mode resource-constrained problem, decoded by a capacity-feasible serial schedule generation scheme. A Q-learning agent adapts mutation operators online; an entropy weight method and technique for order of preference by similarity to ideal solution (TOPSIS) stage evaluates non-dominated solutions matching or improving on current practice to select a compromise. Evaluation covers five benchmark projects of 15–120 activities, five optimizers, 30 seeds and six executed construction projects. Findings Adaptive operator control (AOC) raised mean hypervolume across all five optimizers (+0.6% to + 106.1%), delivering substantial performance gains on decomposition-based architectures (multi-objective evolutionary algorithm based on decomposition + 106.1%, reference vector guided evolutionary algorithm + 36.9%). Relative to baseline practice, selected compromise schedules improved duration up to + 18.5%, resource leveling + 37% to + 70%, underutilization up to + 50% and total cost up to + 6.7%. On executed projects, resource leveling improved by +20.4% to + 55.2%. Originality/value The framework unifies AOC, five-objective scheduling and dispersion-based selection into an integrated workflow, establishing the operational conditions where operator adaptation succeeds on discrete scheduling frontiers. Resource leveling and allocation are optimized as distinct measured objectives rather than simplified as constraints, providing decision support that preserves or improves on baseline practice in every objective.