For differentiable problems, there’s backpropagation. For everything else, there’s RL.
17
58
6
450
Not quite right. A more accurate statement would be "for everything else, there is gradient-free (zeroth-order) optimization." RL is when there is a sequential decision process and what you see depends on previous actions you took.
8
24
296
Replying to @ylecun
We use different definitions of RL. In mine, any problem can be phrased as a one-step MDP (such as in arxiv.org/abs/1611.01578), and zeroth-order optimization is a special case. Can debate definitions, but I use mine because algos like PPO are doing RL regardless of MDP used.

Neural Architecture Search with Reinforcement Learning

Neural networks are powerful and flexible models that work well for many difficult learning tasks in image, speech and natural language understanding. Despite their success, neural networks are...

arxiv.org

Feb 3, 2019 · 5:59 PM UTC

3
28
Replying to @gdb @ylecun
Someone should really clear up the definition of RL. For some it’s sequential decision MDPs, for some it’s a collection of algorithms, for some it’s anything that involves an agent/environment loop, …
1
2
15
Replying to @gdb @ylecun
What about AGI, is it differentiable? Many people agree is not RL since we don't have the luxury of so many attempts
1
1
Replying to @gdb @ylecun
Please do it.