A hybrid MINLP and two-layer reinforcement learning framework for real-time fleet optimisation in open-pit mining

This study presents a hybrid framework combining a multi-objective mixed-integer nonlinear programming model with a two-layer reinforcement learning (RL) system for real-time truck dispatching. The upper Relaxation-Induced Neighbourhood Search (RINS) layer optimises material destinations, while the lower RL layer dynamically assigns trucks based on system states. Applied to a copper mine, the approach outperformed a random base case representing uncoordinated dispatching and achieved consistent improvements over the RINS benchmark. Relative to the base case, truck waiting time decreased by 46.3%, processing capacity utilisation increased by 38.8%, constraint violations dropped by 73.1%, and haulage costs were reduced by 34.8%.