Reinforcement Learning Applied to Fleet Allocation and Informed Short-Term Production Planning of Industrial Mining Complexes

An actor-critic reinforcement learning approach is presented to improve the production of an industrial mining complex by defining shovel allocation and adapting the short-term plan given grade-control decisions. Blasthole data updates the simulated orebody models, which are input into a stochastic grade-control method based on a spatially constrained clustering approach that minimizes the profit-loss function. These aspects are embedded in a discrete-event simulator that defines the material flow from faces to processors. A case study at a copper mining complex shows the methods’ ability to adapt to new information, improving the quality of the material feed and increasing cash flow by 23% compared to a baseline case.