CONNECTING LANE'S THEORY WITH REINFORCEMENT LEARNING: THE FIRST AI AGENT FOR OPTIMAL NONRENEWABLE RESOURCE EXTRACTION *J. Lozano1, M. Amado2 1Genius Mining AI, Lima Perú, (*Presenting author: jorge.lozano@bee3tech.com) ABSTRACT The strategic planning of open-pit mines has traditionally relied on deterministic models and metaheuristics to maximize Net Present Value (NPV). However, these methods often fail to adapt to the dynamic stochasticity of geological and economic variables without exhaustive recalculation. This paper introduces a novel approach using Reinforcement Learning (RL) to develop an autonomous AI agent capable of optimizing the cut-off grade policy and extraction sequences. By mapping Lane’s Theory of Cut-off Grades into a Markov Decision Process (MDP) and utilizing the Bellman Equation for value iteration, the agent learns to maximize long-term rewards through interaction with a simulated mining environment. Results demonstrate that the RL agent not only converges to the optimal NPV faster than traditional iterative methods but also exhibits high adaptability to fluctuating operational costs and market prices. This research lays the foundation for Mining 4.0, where autonomous agents provide a resilient framework for strategic decisionmaking in complex mineral resource management. KEYWORDS Reinforcement Learning, Strategic Mine Planning, Lane’s Theory, NPV Optimization, AI Agents. 1. INTRODUCCION Strategic mine planning is an NP-hard optimization problem where the objective is to determine the optimal extraction sequence and cut-off grade policy over the life of a mine (LOM). Traditional methods, such as Mixed-Integer Linear Programming (MILP), face scalability issues when considering multi-destination and multi-period constraints under uncertainty. The mining industry currently faces a "decision gap": despite having vast amounts of data, the ability to pivot strategic plans in real-time remains limited. This paper proposes the integration of Artificial Intelligence, specifically Reinforcement Learning, to transition from static planning to an adaptive, agent-based optimization model. Our goal is not merely to solve the problem, but to teach a machine how to solve the problem. In the Figure 1 show an overview of our solution.
RkJQdWJsaXNoZXIy MTM0Mzk2