Yes, this makes it very difficult to apply ML and RL in non-simulated scenarios. With simulated scenarios you can just replay and "sweep" across hyperparameter

Yes, this makes it very difficult to apply ML and RL in non-simulated scenarios. With simulated scenarios you can just replay and "sweep" across hyperparameters to find the best one. In a realworld scenario with limited information, fine tuning hyperparameters is much harder as you quickly find yourself in local maxima.