Interpretable Deep Reinforcement Learning for Dynamic Truck Dispatching in Open-Pit Mining

Abstract

Dynamic resource allocation under uncertainty is a core challenge in operations research, particularly in capital-intensive sectors such as open-pit mining, where suboptimal dispatching decisions can lead to substantial productivity losses and increased operational costs. This paper proposes a Deep Reinforcement Learning framework to address the dynamic truck assignment problem in open-pit mining, with a dual focus on performance optimization and policy interpretability. The problem is formulated as a Markov Decision Process (MDP) and implemented within a discrete-event simulation environment that replicates the stochastic dynamics of truck-shovel operations. A Deep Q-Learning approach with neural network function approximation is employed to learn adaptive dispatching policies directly from simulated experience. To enhance transparency, we integrate Explainable Artificial Intelligence (XAI) methods to interpret and validate the learned decision strategies. Experimental results show that the proposed framework significantly outperforms heuristic and random dispatching baselines in terms of resource utilization and operational efficiency. Moreover, the incorporation of explainability provides actionable insights into policy behavior, bridging the gap between autonomous optimization and human-centered decision support in complex industrial systems