Lazy Twitter: Why do I think Reinforcement Learning is BS? What insight am I missing here?
All the heavy lifting appears to be performed by the neural networks. The only thing RL seems to be doing is performing an ad-hoc search method for finding new training data. RL is a glorified data collection procedure!
The self-play methods are indeed very compelling, but this idea is independent of RL. It's more similar to how you train a GAN where there are two networks that are cooperatively/competitively tuning their models.
There are also models that collect data to be replayed at a later time for further training. If RL is nothing but a glorified data collection mechanism, then I can't see where the hype is coming from.
The cult of RL comes in second after right after the cult of Bayes in terms of BS ideas about how the brain works.
If you look at the Bellman equation is everywhere in RL. It's the same damned Hamliton-Jacobi equation from dynamics. WTF then are we saying here? That the exploration method is the same as a ball rolling down a hill?!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
