Automated generation of assembly schedules for precast building projects under uncertainty using reinforcement learning and Monte Carlo sampling

Efficient onsite assembly sequence planning and scheduling (ASPS) is crucial for the successful delivery of precast building projects. The manual ASPS process is tedious, error-prone, and sub-optimal. Existing research on its automation lacks in considering real-world constraints and on-site uncertainties, and suffers from high computational burdens. To address this challenge, this paper proposes a novel reinforcement learning (RL) and Monte Carlo sampling (MCS)-based method for automated ASPS. The method utilizes a temporal graph network to create state embeddings, which are then employed by a Proximal Policy Optimization algorithm-based agent to learn the ASPS policy. The agent learns the policy over a distribution of uncertain variables using MCS, with their values randomly sampled at the start of each episode. Validation on a real-world precast building project demonstrates that the proposed method outperforms traditional methods, yielding dominant solutions in 60% of test cases in deterministic and stochastic conditions, while requiring only about one-third of the training time. Future research can explore Pareto front generation and reward engineering to enhance the practical applicability of the proposed method.