Re:Deep reinforcement learning is a waste of time (2019)
by freeone3000 on Saturday July 10, 2021 @01:54PM
(#27795173)
It's refuting the premise that supervised RL becomes less supervised because you put the feedback in a handcrafted function and use a neural network. Deep RL in its current state should be grouped with Supervised RL, in other words (which is why I personally think that imitation learning is a great way forward, in contrast with the author). The issue is the amount of interactive tweaking and lack of a natural reward function that prevents DeepRL from being unsupervised.
[ Parent
| on HN ]
Re:Deep reinforcement learning is a waste of time (2019)
by rich_sasha on Saturday July 10, 2021 @03:09PM
(#27795683)
AlphaZero is not supervised, in the sense that it learned from known correct actions (earlier versions of AlphGo did learn from online games). So although it needed human supervision, sure, it didn’t need us to provide correct answers.
The authors point does somewhat stand that you don’t have the problem of reward engineering in board games so they are a dead end from that point of view - they skirt around the core problem instead of tackling it.
Re:Deep reinforcement learning is a waste of time (2019)
by freeone3000 on Sunday July 11, 2021 @12:22AM
(#27798452)
AlphaZero only works on video games. If you remove its ability to judge progress by game score, which is a reward function (though not the one used, due to delayed reward issues), then it's not capable of finding its feedback. It only works in constructed environments where the environment provides the reward function implicitly. Maybe we can video-game-ify laundry folding sufficiently? I'm doubtful.