I think I should mention that the lecturer is one of the researchers who proposed dqn, and his name was the first among all the researchers. I like how modest he is hahaha. This is actually one of my favourite lecture. So much insight. Thank you!
Isn't 0:25:29 a pseudo code for DDQN? We have Q and Q^ weights mentioned. On the other hand, the formula for y target is not the one of DDQN as far as I understand...
I believe it was about how frequently the weights of the Q net being learned is copied to the target network. Shouldn't be too frequently to avoid non-stationarity in target computation and again not too less frequently to avoid target network weights being too stale. Needs to be picked up through experimentation.
He doesn't seem like the greatest presenter, and while I guess it's hard to find people who excell at both machine learning AND presenting and I can certainly see his expertice on the topic, he might wanna work on the presentation part a litte :D He made it a bit hard for me to keep paying attention :/
![Seungmin "그렇게, 천천히, 우리(As we are)" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/kAzmhLHePqU/mqdefault.jpg)
I think I should mention that the lecturer is one of the researchers who proposed dqn, and his name was the first among all the researchers. I like how modest he is hahaha. This is actually one of my favourite lecture. So much insight. Thank you!
Thanks for sharing this lecture and the Deep RL Bootcamp 2017 playlist overall.
Isn't 0:25:29 a pseudo code for DDQN? We have Q and Q^ weights mentioned. On the other hand, the formula for y target is not the one of DDQN as far as I understand...
Why is the default video speed 0.5?
Thanks for that awesome lecture... You were very informative and insightful... :)
I did not expect that guy to sound like he does.
Some slides cover two or three points. One suggestion I'd give is to split one specific slide into multiple ones or add some animations.
Agreed. Sadly, this is true of so many presentations I've sat thru.
Very good explanations!
watch in 1.5x and thank me later
1.25 is perfect!
1.25 is better, everything seems natural xd
THANKS!!
I did 1.25 for this guy, and .75 for karpathy
LOL, yeah 1.25 or 1.5 speed, I can actually pay attention. This dude is.. sloooooooo...
Thanks for this lecture
great talk.
What's the question at 12:30?
I believe it was about how frequently the weights of the Q net being learned is copied to the target network. Shouldn't be too frequently to avoid non-stationarity in target computation and again not too less frequently to avoid target network weights being too stale. Needs to be picked up through experimentation.
nice !!
nice
He is opposite of Karpathy .
He doesn't seem like the greatest presenter, and while I guess it's hard to find people who excell at both machine learning AND presenting and I can certainly see his expertice on the topic, he might wanna work on the presentation part a litte :D He made it a bit hard for me to keep paying attention :/
He forgot to fill his belly before doing this
HE is AWKWARD AF
The guy is demotivating and uninterested in teaching. Please bring David Silver back, that guy makes the information more appealing in my eyes