The paper is concerned with the development, optimization and assessment of a deep neural network based policy for dynamic truck-to-process allocation in an open-pit mining system. The study addresses the challenge of the need for real-time control in a highly uncertain environment. The policy was developed using a reinforcement learning (RL) method working in tandem with a discrete-event simulation model of the mining operations. The simulation model was based on the logistical and performance data collected from the Sungun mine in Iran. Key performance indicators, such as truck utilization, waiting time, and production rate, serve as the metrics for evaluating the model's effectiveness. The performance of the RL-based policy was optimized in a series of sensitivity analyses ranging key modeling parameters such as the structure of the neural network, and RL algorithm reward period. The performance of the optimized RL-based policy is compared to that of a common rule-of-thumb policy over a series of lengthy validation runs. The findings clearly show that the RL-based policy delivers substantial improvements over the rule-of-thumb approach, increasing the production rate by an average of 17% per cycle. In addition to technical performance, the paper also explores practical considerations essential for the RL-based policy’s successful deployment. These include its scalability, as well as the need for high-quality data and computational resources. A comparative critique of existing optimization methodologies further emphasizes the broader applicability of the RL-based policy approach in managing complex, dynamic optimization challenges. Importantly, the techniques and insights gained are not limited to mining; they have a potentially wide field of application in the construction industry as well. In summary, this research marks a significant step forward in the field of real-time control for both the mining and construction sectors.