>It's a shame how data hungry DRL is (even when compared to DL), but the DRL framework encompasses all the standard classification/regression tasks and also inc

>It's a shame how data hungry DRL is (even when compared to DL), but the DRL framework encompasses all the standard classification/regression tasks and also includes decision making, planning and pretty much anything else you can think of. Sure, and a Turing machine can express any program you care to write. RL seems like the Turing tar pit of machine learning: theoretically able to express everything, practically convenient mostly just for trivial examples. >Model generality and data efficiency are in an inverse relationship and a lot of research has been in moving up or down this hyperbola. Sure, basic Curse of Dimensionality, but the whole success of hierarchical modeling in the brain, hierarchical Bayesian methods, and deep neural networks has been that hierarchical modeling seems to ameliorate or even defeat the Curse of Dimensionality. The question is: well, why can't it do that in reinforcement learning? In interesting terms: why does the teaching signal have more information in supervised learning than reinforcement learning, relative to the inherent uncertainty of the task?