The first version of this project was framed around reinforcement learning.
That framing was useful because it forced the project to ask a clear question: what would an intelligent scheduling agent need to observe, decide, and improve? It also made the risks visible quickly. Before an agent could learn anything useful, the environment needed to represent enough of the manufacturing world for its decisions to have meaning.
Early simulator work made the gap obvious. A cutting room is not just a queue, a machine, and a reward function. Jobs arrive with production context. Markers carry information about fabric, layout, quantity, and downstream expectations. Machines have states that are easy to summarize but harder to explain. Delays and exceptions often matter as much as the planned sequence.
Training an agent before modeling that behavior would have created a clean experiment, but not necessarily a useful one. The agent would have learned from whatever assumptions were easiest to encode. If those assumptions were shallow, the resulting policy could look coherent while optimizing the wrong world.
The project therefore stopped building reinforcement learning first. The immediate work moved toward production data, classification, investigation tools, and manufacturing analytics. Those pieces are less glamorous than an agent demo, but they define the environment the agent will eventually need to inhabit.
This does not mean reinforcement learning was abandoned. It was postponed until the foundation can support it responsibly. The current path is to make the manufacturing behavior more believable, validate the digital twin against observed production patterns, test simpler rule-based scheduling agents, and then return to reinforcement learning with a better environment and clearer evaluation criteria.
For this project, the engineering lesson was direct: an intelligent agent is only as credible as the world it learns from.