Timeline for Is RL just a less rigorous version of stochastic approximation theory?
Current License: CC BY-SA 4.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| May 22, 2020 at 16:51 | history | edited | nbro | CC BY-SA 4.0 |
added 82 characters in body
|
| May 22, 2020 at 13:46 | comment | added | nbro | @FourierFlux Well, we could say that all matter (including our brains that discovered and developed chemistry, RL, and SAT) is made of atoms, so, in the end, everything is an atom or a bunch of atoms. How productive is that reasoning for the purposes of distinguishing RL from SAT? | |
| Apr 15, 2020 at 14:40 | comment | added | FourierFlux | Chemistry has its own theory/assumptions which predated modern physics. I'm not seeing that in RL, which is why I said it was a less rigorous version of SAT. | |
| Apr 15, 2020 at 7:09 | comment | added | Neil Slater | @FourierFlux: Your claims are similar to "chemistry is just physics". Whilst true at some level, a researcher who only studies and practices physics will not be an expert chemist. | |
| Apr 15, 2020 at 4:11 | history | edited | nbro | CC BY-SA 4.0 |
added 58 characters in body
|
| Apr 15, 2020 at 4:04 | comment | added | nbro | They are distinct in the sense that there are notions in RL (e.g. MDPs) that do not necessarily exist in SAT. As I say in my answer, if you want to study the convergence properties of RL algorithms, you probably need SAT, yes, because RL algorithms are essentially iterative algorithms and the goal in RL is to find extrema of functions. However, note that RL wouldn't exist without actions (or decisions). In fact, RL is sometimes also called sequential decision-making. | |
| Apr 15, 2020 at 4:01 | comment | added | FourierFlux | I wouldn't call it distinct, maybe a subfield. But can you really do any work in a theoretical sense in RL without a background in SAT? | |
| Apr 15, 2020 at 4:00 | comment | added | nbro | @FourierFlux I've updated my answer to further explain that, yes, they overlap, but they are distinct. | |
| Apr 15, 2020 at 3:57 | history | edited | nbro | CC BY-SA 4.0 |
added 107 characters in body
|
| Apr 15, 2020 at 3:51 | history | edited | nbro | CC BY-SA 4.0 |
added 156 characters in body
|
| Apr 15, 2020 at 3:49 | comment | added | FourierFlux | Sure, but at it's core RL is essentially stochastic approximation with some stuff added ontop. The Robins Monroe theorem is pretty much what everything is built on and without it RL wouldn't exist(at least in any rigorous sense). | |
| Apr 15, 2020 at 3:34 | history | answered | nbro | CC BY-SA 4.0 |