Abstract: Underground mine production scheduling is the core component of intelligent mining production. Its decision
optimization faces complex challenges of multiple constraints and dynamic environments. Based on the study progress in the industry
and the team′s study accumulation,the system has proposed three key scientific issues in this field. The difficulty of
multi-objective optimization lies in the efficient trade-off between conflicting objectives and the search for the Pareto frontier.
Uncertainty handling mainly involves shifting from preset probability models to data-driven robust modeling to cope with the
strong randomness of geology,equipment,and market. Real-time response requires the scheduling system to strike a balance between
decision speed and quality to quickly respond to production disturbances. Focusing on these issues,the study progress of
methods such as mathematical programming,meta-heuristic algorithms,and simulation optimization has been reviewed. The core
contributions and inherent limitations of these methods in addressing the three issues have been analyzed. Mathematical programming
provides a rigorous modeling framework,but the "dimension disaster" and parameter sensitivity reveal the structural
contradiction between it and the high-dimensional nonlinearity of underground mining. Meta-heuristic algorithms and simulation
optimization expand the solution boundaries for complex problems,but they are respectively constrained by the lack of convergence
theory guarantee and high modeling costs. On this basis,the transformative potential of machine learning technologies has
been explored. Deep learning provides a data-driven new paradigm for uncertainty modeling through high-precision prediction.
Reinforcement learning opens up an end-to-end decision-making path for real-time dynamic scheduling through continuous interaction
with the environment. Further analysis reveals the shortcomings of current study. Multi-objective optimization lacks an
effective decision preference guidance mechanism,uncertainty handling presents a disconnected state between prediction and
optimization,real-time response technology is still far from practical application,and systematic exploration of multi-method integration
is lacking. Correspondingly,the study focus in this field during the "15th Five-Year Plan" and beyond is to build a
full life cycle multi-objective collaborative optimization framework,develop a decision-making paradigm deeply integrated with
prediction and optimization",break through the real-time scheduling technology of safe and trustworthy reinforcement learning,
and promote the intelligent decision-making closed loop driven by digital twins,in order to push underground mine production
scheduling towards an intelligent and autonomous process.