A hybrid approach combining the genetic algorithm and reinforcement learning for long-term production scheduling at a cement quarry

The urban proportion of the global population is expected to rise to two-thirds by 2050, thus, increasing the demand for more urban infrastructure. Cement will play a critical role in ensuring accelerated construction of such infrastructure. A key objective of cement quarry production scheduling is to minimise the cost of extraction over the long term, subject to production constraints and variability in operating parameters as encountered in actual mining practice. Unlike previous studies that have applied genetic algorithm (GA) or reinforcement learning (RL) independently to production scheduling problems, this study proposes a hybrid GA–RL framework to combine complementary strengths of the two methods. In this framework, the GA's population-based search is used to generate candidate raw-material mixes, while an RL agent operating under a Deep Q-Network policy sequentially refines these mixes to improve the scheduling performance. This integration allows the approach to simultaneously explore a broad solution space and adapt to a decade of escalating costs and exchange-rate variability, a capability not previously demonstrated in long-term cement quarry scheduling. The hybrid GA-RL approach reduced raw material costs by approximately 11.8% (KES 2.251 billion) while meeting quality constraints, outperforming GA-only, RL-only, and deterministic manual scheduling methods used in current planning practice.