A Surrogate-Assisted Evolutionary Framework with Verified Policy Feedback for Resource-Constrained Multi-Project Scheduling

Abstract Engineering portfolios require coordinated scheduling of projects that compete for shared resources. Evolutionary optimisation offers flexible search, but repeated schedule generation is costly and fixed search decisions respond poorly to changing population conditions. This study develops a surrogate-assisted evolutionary framework that coordinates hierarchical search control with budgeted schedule verification. A policy selects genetic-algorithm and differential-evolution branches, their operators and coupled parameter settings. Multiple actions generate offspring from the same population state, and verified relative quality supplies within-generation learning feedback. An iteratively refitted gradient-boosted tree model prioritises candidates for exact schedule generation and resource realignment. Verified schedules determine survival, model labels and policy rewards, while explicit accounting charges every scheduling pass to the run-level budget. Experiments across three benchmark libraries demonstrate competitive scheduling quality against general-purpose and problem-specific baselines. Surrogate assistance reduces schedule-generation work by 40% and mean runtime by 27.5% relative to the search-only variant, while maintaining closely comparable makespans. The framework balances portfolio schedule quality and computational effort through coordinated learning and verified search.