An evolutionary reinforcement learning framework for joint work package sizing and scheduling with uncertainties

Work package sizing and scheduling are two core components in project management, often modeled as optimization problems to improve productivity, resource utilization, and risk control. In practice, however, projects face uncertainties such as machine breakdowns and labor absences, making static package structures and fixed schedules inefficient or even infeasible. This highlights the need for adaptive approaches that can dynamically adjust both work package structures and schedules to ensure timely completion and manage resource fluctuations. In this study, we formulate the joint work package sizing and scheduling problem as a Markov Decision Process, explicitly capturing concurrent uncertainties stemming from resource disruptions as well as variability in work content and task durations. To solve it, we propose a novel evolutionary reinforcement learning framework comprising three synergistic modules: (1) a graph-based reinforcement learning algorithm to learn adaptive and generalizable policies, (2) the Cross-Entropy Method to efficiently explore the policy space, and (3) an asynchronous-synchronous mechanism to balance exploration and exploitation. Extensive experiments demonstrate that our method outperforms existing baselines across diverse JWSSP instances, offering a promising solution for real-world, uncertainty-aware project scheduling.