An existing concept for sequence planning in production planning and control was extended by a global decision instance based on neural networks. Therefore, information regarding the state of the production and available orders were normalized and analyzed by one agent. In contrast to a partially observable Markow Decision Problem one single agent was allowed and used to process all available information. Feasibility and problems were examined and compared with a concept for decentralized decisions. The implementation consists of two parts, which continuously interact with each other. One part is a simulation of a job shop, including multiple machines. The other parts tackle the Markow Decision Problem with the use of double Q reinforcement learning in order to estimate the best sequence at any given time. Later, problems due to scaling and comparisons to the usage of multiple agents are given.