Geometallurgy as a multi-disciplinary field has been applied at various levels in different operations. By linking the ore performance in mineral beneficiation processes to the ore block model, it supports estimating the value of a block before it is mined. Efforts in the classification of the ore into geometallurgical classes have led to a better understanding of the entire value chain. While classification provides a convenient tool for forecasting and visualization purposes, it simplifies the actual complexity of an ore body. In mining and process planning, sequential decisions are made to maximize an objective function or equivalently minimize a regret function. Using available information from geology or metallurgical test work, an optimal strategy can be found using tools from the machine learning community.

In this study, a framework based on machine learning to maximize the use of such classifications for sequential decision-making is proposed. The concepts of reinforcement learning and bandit algorithms, offer powerful tools to explore and exploit different optimization strategies. In certain cases, theoretical guarantees about the performance of given methods can be obtained by regret bounds.

Based on existing models of a porphyry copper deposit and an iron ore deposit, this study presents a methodology and different available algorithms to maximize an objective function that depends on a high number of variables and in the presence of noise or uncertainty in the models. Different numerical experiments provide a basis for discussion and comparison to human decisions. The hypotheses relative to each algorithm are discussed in relation to the mineral processing models.