Abstract
Abstract Ore dispatch in open-pit mining is a multidimensional stochastic optimization problem that requires dynamic decision-making to balance competing objectives: maximizing throughput, maintaining grade consistency, and reducing equipment queuing. This study formulates and evaluates a comparative reinforcement learning (RL) framework that learns adaptive dispatch policies in a simulated Internet of Things (IoT)-enabled open-pit mine. Three deep RL algorithms are implemented and compared: value-based dueling double DQN, on-policy proximal policy optimization (PPO), and maximum-entropy soft actor-critic (SAC). The results show that, although all algorithms can learn viable policies, SAC exhibits higher sample efficiency and more stable convergence, reaching strong performance in fewer simulated episodes. A detailed policy analysis reveals that the SAC agent learns an advanced non-intuitive strategy that improves grade control and yields balanced performance across objectives. Still, it does not consistently outperform a simple shortest-queue heuristic on all key performance indicators (KPIs), particularly throughput and queue minimization. This discrepancy exposes an agent–objective alignment problem and indicates that the hand-crafted reward function does not fully capture the underlying business priorities. The proposed framework, therefore, serves not only as an optimization tool but also as a diagnostic mechanism for exploring multi-objective trade-offs and revealing misalignment between engineered rewards and high-level KPIs, providing a practical basis for designing and validating autonomous dispatch systems prior to real-world deployment.