At different stages of training different hyper parameters work best. For example learning rate should be decreased near the end of training. ml-agents has learning_rate_schedule and curriculum that partially address this but it is hard ti know what values work best and each stage. So I end up running my training and experiment with values at different stages of training. The problem here is that the initial parts of training get repeated each time I want to change hyper parameters for later stages of training. It wastes a lot of computation power and time. I tried using --initialize-from to continue training as a new training session but the result is different than running one longer training session with curriculum. I suspect that some stats carry over when using curriculum in the same training session versus starting a new session with --initialize-from. Can Unity ml-agents team shed some light on what would be the best way to change hyper parameters mid training? Hope my post makes sense.