Does completely redoing vector observations or doing heavy revision break the training? What about for small changes (I had to create a new environment in the past). I know that the same order of vector observations for replacing the vectors does not effect the agent. What if I need to make changes millions of steps down, will it break my ai?
Edit: Specifically, after changing a lot of observations, adding new and removing old observations, will I be able to resume training in a new environment or will I need to restart training in a new environment?
(In a few days, when I get a bit of time between regular dev, I will just run this on my own end and come back with the answer if no one knows, as I currently don’t have my ml trainning computer on hand.)
Generally, yes, but it depends quite what you are changing. Imagine you as a real person learned to recognise flowers by what colour they are, but then your eyes suddenly changed over night so that all the colours looked different. You would think you were seeing one flower, but actually be seeing a different one.