10h30 - 10h55
Self-learning Hyper-heuristics for the Optimization of Industrial Mining Complexes
The simultaneous stochastic optimization of mining complexes (SSOMC) is a large-scale stochastic combinatorial optimization problem that incorporates all components of the mineral value chain, simultaneously managing the extraction of materials from multiple mines and their processing through interconnected facilities to generate a set of final products while accounting for geological uncertainty to manage the associated risk. While simulated annealing has been shown to outperform comparing methods for solving the SSOMC, early performance may dominate recent performance in that a combination of the heuristics' performance is used to determine which perturbations to apply. A new generation of self-learning stochastic optimizers will be presented. The hyper-heuristics select the heuristic (perturbation) to be applied using reinforcement learning to efficiently explore the local search best suited for a particular search point. By learning from data describing the performance of the heuristics, a problem-specific ordering of heuristics is obtained, collectively finding better solutions faster. Beyond the data-driven heuristic selection, a tree structure, similar to the branch-and-bound algorithm, is proposed to aggregate the steps, bound certain decision variables, and backtrack if necessary. The proposed hyper-heuristics are tested on several real-world industrial mining complexes, showing a major reduction in the number of iterations and computational time.
10h55 - 11h20
Actor-Critic Reinforcement Learning Applied to Adaptive Shovel Fleet Management in Industrial Mining Complexes
This work presents an Actor-Critic approach that allocates shovels in an industrial mining complex to improve the overall financial performance of the related operation while accounting for the operational requirements of these pieces of equipment and adapting to incoming new data. Additionally, the model updates stochastic grade-control decisions by introducing a hierarchical clustering approach. The generated clusters are subjected to spatial constraints that minimize the profit-loss function associated with processing the whole cluster at a single location. Incoming blasthole data from grade-control activities are used to update the uncertainty associated with geological attributes that characterize the mineral deposit using the ensemble Kalman filter method. Next, the grade-control clusters are updated in order to adapt to this new information. Both orebody models and grade-control updating happen internally on a simulation of the mining complex operation, which also provides the material flow from the mining faces to the processors. Thus, an Actor-Critic agent is trained to perform fast shovel allocation decisions under the above-mentioned framework. A case study at a copper mining complex, composed of two open-pits and multiple destinations, showcases the ability to adapt to new information and make quick decisions improving the financial performance of the mining complex by 15%.
11h20 - 11h45
An actor-critic reinforcement learning framework for simultaneously optimizing a short-term production schedule in a mining complex with an uncertain supply
A mining complex is a large engineering system where valuable minerals are extracted, transported, processed, and delivered to customers and the market. A new framework for short-term production scheduling of a mining complex is proposed that combines actor-critic reinforcement learning with stochastic mathematical programming to simultaneously optimize the short-term production schedule of a mining complex with uncertain supply. The optimization integrates coarse preconcentration decisions to manage waste production prior to processing using screening facilities.