This work investigates the applicability of a reinforcement learning (RL) approach, specifically AlphaGo Zero (AZ), for optimizing sheet-metal (SM) production schedules with respect to tardiness and material waste. SM production scheduling is a complex job shop scheduling problem (JSSP) with dynamic operation times, routing flexibility and supplementary constraints. SM production systems are capable of processing a large number of highly heterogeneous jobs simultaneously. While very large relative to the JSSP literature, the SM-JSSP instances investigated in this work are small relative to the SM production reality. Given the high dimensionality of the SM-JSSP, computation of an optimal schedule is not tractable. Simple heuristic solutions often deliver bad results. We use AZ to selectively search the solution space. To this end, a single player AZ version is pretrained using supervised learning on schedules generated by a heuristic, fine-tuned using RL and evaluated through comparison with a heuristic baseline and Monte Carlo Tree Search. It will be shown that AZ outperforms the other approaches. The work’s scientific contribution is twofold: On the one hand, a novel scheduling problem is formalized such that it can be tackled using RL approaches. On the other hand, it is proved that AZ can be successfully modified to provide a solution for the problem at hand, whereby a new line of research into real-world applications of AZ is opened.