DeepMind x UCL RL Lecture Series - Function Approximation [7/13]

Research Scientist Hado van Hasselt explains how to combine deep learning with reinforcement learning for "deep reinforcement learning".
Slides: dpmd.ai/functionapproximation
Full video lecture series: dpmd.ai/DeepMindxUCL21

[@haliteabudureyimu638]

You said at 1:36:05 that there are algorithms inspired by TD which guarantee to converge for non-linear cases. Can you provide the names and papers of these algorithms?

[@robinranabhat3125]

1:02:45 Convergence and Divergence
1:52:41 Deep Reinforcement Learning

[@imranqureshi4299]

Thank you for the lecture. Are there exercises available with these lecture series?

[@chrishermans]

I have a bit of an implementation question here... So what's the difference between implementing neural q-learning (as in DQN, but without the target network), and implementing fitted q-iteration (FQI), but with a neural net as a function approximator? Assuming we retain the weights from the last step, that is.
I am not finding many comparisons with FQI anywhere in the material provided here, and I'm just wondering if these are implementation details, or if there are some fundamental differences that I am missing somehow.
Please elaborate, anyone who can.
PS: and yes, I saw some response on this very question on stackoverflow. But it was not exactly a helpful response on there, imho. I'm hoping this audience will fare better. ;-)

[@josefbajada5106]

At 1:51:06 should it be n < t instead of n

[@entropist5251]

Yes.

[@starkt1554]

thank you

[@nuvembook505]

Gostei disto