A joint work package sizing and scheduling problem considering resource constraints with a look-ahead heterogeneous reinforcement learning method

Effective work package sizing and project scheduling are crucial for construction project management. However, existing studies often address them as separate optimisation problems, neglecting the interactive nature of these processes. In practical scenarios with limited resources, there is an increasing demand for integrating work package sizing and project scheduling to enhance project management efficiency. This research aims to bridge this gap by developing an integer non-linear programming model that incorporates work package sizing and project scheduling while considering their interaction within a resource-constrained environment. Moreover, we introduce a novel look-ahead heterogeneous reinforcement learning approach that dynamically adjusts work package sizes and project schedules based on observed information at each decision step. A look-ahead sequential decision mechanism is proposed to effectively address the interdependencies and constraints inherent in the joint model. We extend the heterogeneous agent mirror learning technique to our problem to improve sample efficiency and ensure progressive enhancement of the joint policy. To evaluate our approach's effectiveness, we conduct experiments using datasets of 15, 30, 60, 90, and 120 tasks from Rangen and PSPLIB and validate our method in a real-world modular integrated construction project. The experimental results reveal that a joint model integrating work package sizing and project scheduling offers a more comprehensive understanding of their interplay, leading to better resource utilisation and improved project schedules. A comparative analysis against existing state-of-the-art reinforcement learning and priority rule-based methods further substantiates the superior effectiveness of our proposed approach, yielding an approximate 5% reduction in total project cost.