Abstract
An operating industrial mining complex is a data-rich environment that collects new data while monitoring several different activities. Such new information increases the knowledge about the operation and can be used to update uncertainty models, thus guiding updated decision-making. This work considers incoming blasthole data that update uncertainty and variability associated with the orebody models, which then is used to provide final destination decisions that minimize misclassification losses. A new stochastic grade-control model based on a spatially constrained clustering approach that minimizes the profit-loss function is presented. These grade-control cluster decisions are considered feasible allocation where the shovel equipment can be allocated. Next, an actor-critic (AC) reinforcement learning approach is introduced to provide equipment allocation and production scheduling decisions given the new information acquired throughout grade control. A discrete-event simulator of the mining complex’s operations is used to forecast the material flow resulting from mining the material from benches and sending them to the related processor. The AC method interacts with the simulator of the mining operations, improving the production scheduling decisions over time. A case study at a mining complex demonstrates the improvements in destination decisions from a better understanding of the uncertainty that characterizes the orebody models. The AC approach is compared with a baseline approach where the shovel paths were previously optimized. Results show that the adaptation of the production schedule increases considerably metal production by using the AC approach, which leads to a 15% increase in cash flow for 30 days of operation.