Handling stochasticity in open-pit mining truck dispatch using deep reinforcement learning

The material haulage operation within an open-pit mine is one of the dominant contributors to total operating costs. Typical pre-scheduled truck allocation approaches can lead to schedule delays and under-performance owing to stochasticity, such as environmental uncertainty or equipment malfunction. This article implements an automatic dispatching solution using reinforcement learning to learn optimal policies, operate robustly under state uncertainty and make decisions in real time. A generalized simulation environment is developed, supporting stochastic variance in grade estimation, repair effectiveness, shovel loading times and breakdown risk. Six variations of deep Q-networks are designed and tuned using a novel parameter optimization method combining reward shaping and network tuning, and competitively tested against three heuristic baselines. The strongest production-prioritizing model increased average profits by 39.9%, while the most reliable model reduced breakdown events by 30%. These results are a step towards fully autonomous fleet operations, improving both mine safety and overall profits.