Just subscribed! I love the level you teach at in your videos. It’s slightly above the level of Statsquest but not too dense that I need to mentally prepare before watching. (No shade to Statsquest, two random events can be independently great).
Can you please make a full playlist on reinforcement learning. No one explains the math stuff as simply as you do. Also please do a separate video going into greater mathematical detail proving the theorem, kind of like numberphile2
Thanks for the great video 💪👏 will there be more RL videos coming? E.g I would like to understand more how to set up reward functions. How do I weight rewards from different actions against each other? Also how would we set up the environment in the model based approach? And more
You mean causal inference? I suggest you to refer to econometrics textbooks, as in Economics we are pretty strong on that. - "Mostly harmless econometrics" by Pishke for a graduate level in applied stats (pure stats would find it undergrad level) - "Econometric analysis of cross-section and panel data" by Wooldridge
@@matteogirelli1023 what do you suggest for someone who want to causal inference to my domain of problem taht is climate science and earth science i just got started by reading causality by judea pearl
Great video! Thanks for sharing. I have a question though. I am new to the subject, so I am having trouble to understand the last step in your derivation (29:00). I am speaking about the "very very easy thing to do". Would you be kind enough to point me to where I can find more information about that? Thanks!
Thanks Ritvik. Still having some trouble with why Reward (R) does not depend on Theta, mathematically - in your tree diagram, all the rewards are +-1/p, dependent (in their absolute quantity) only on the state, but also dependent (in their sign, which seems non-negligible in a reward system) on a Theta-based choice (H or L). Are you able to describe in a different way the intuition behind why d/dTheta logP(R,S|S,A) = 0?
Maybe better said is: what's the nuance (or obvious principle) that allows consideration of an explicit variable in a derivative wrt that variable, but disallows consideration of an implicit variable (ie further back in the causal chain) in a derivative wrt to that variable? Thanks again, your channel rocks!
Hey! Excellent question and it isn’t obvious in any sense. The key lies in the fact that this is a conditional probability rather than an unconditional one. If we removed the conditions on P(R,S | S,A) so it is just P(R,S) then this absolutely does depend on the policy theta and we can measure this dependency by tracing through the causal diagram. However by using a conditional probability we assume that the previous state and previous action are taken as given, at which point the probabilities for the next state and next reward are fixed and do not depend on the policy. Please let me know if that helps!
Hi, I've trying to use multiple sources to look for the proof of this theorem. However, none of them use product rule (for derivative, time: 20.22). Can you please share with me if you know of a resource which does include the product rule.