Applying deep reinforcement learning for construction labour dispatching in a union hall setting

This paper addresses the gap in dynamic resource allocation in construction by formalizing union hall dispatching, a stochastic process with uncertain job arrivals, durations, and worker turnover, as a Markov Decision Process (MDP) and developing a Deep Reinforcement Learning (DRL) agent to automate labour allocation. The agent incorporates a waiting-queue-based fairness metric and, in simulation, achieves consistent fairness improvements while maintaining competitive job assignments against heuristic baselines. Methodologically, the study tailors DRL formalization through careful state representation and reward design, supported by a narrative analogy that frames the agent as a novice dispatcher to enhance interpretability. While serving as a proof-of-concept study with acknowledged simplifications, the work (1) establishes an MDP foundation for DRL in dynamic resource allocation, and (2) demystifies DRL for construction applications. These contributions lay the groundwork for dynamic resource allocation as an emerging research direction in construction, with applicability extending beyond the union context. • Develops Deep Reinforcement Learning (DRL) agent for labour dispatching and resource allocation. • Formalizes dynamic union dispatching as a Markov decision process. • Designs a waiting-queue-based metric to measure labour dispatching fairness. • Demonstrates statistically superior performances by DRL agent over standard heuristics. • Lays a foundation for applying DRL in dynamic construction planning beyond union dispatching.