Automating and optimising pushback selection using reinforcement learning

Pushback optimisation, an early phase of strategic open pit planning, shapes mining phases that guide extraction for the mine's life and strongly influences operational feasibility and economic value. The task unfolds in two steps: first, a set of nested pits is generated; second, a subset of these pits is chosen, defining the pushbacks. Selecting pushbacks is hard because it must meet several operational criteria and has an impact on the production schedule, and thus the project's NPV. Existing tools offer partial help but still depend on manual judgment and rough schedules, often preventing optimal solutions. This study presents an automated pushback-selection method built on reinforcement learning (RL). An RL agent learns, through interactions with the nested-pit environment, to select the pushback set that maximises NPV while respecting key constraints, minimum mining width (MMW), waste-to-ore ratio, balanced tonnage swings, and full use of mining and processing capacities. The framework is tested on the publicly available McLaughlin mine dataset, with emphasis on maintaining the MMW constraint. Results show the RL approach is efficient and economically superior: it produces pushbacks whose NPV is 9% higher than mining every nested pit sequentially. These outcomes underscore RL's promise for automated, constraint-aware pushback optimisation.