RL by DeepMind


#1

What do folks thin of DeepMind’s latest RL results?


#2

It is really good work. But why can they start from a more primitive? level. like
leaning to walk first, and then run, and then etc…
Seem like there staring a bit high up, on a bottom to top approach.


#3

keghn, yes, they need to learn to collect together a mass of experts walking, planning, talking, … I believe Google’s Jeff Dean is big into this. Google plus DeepMind will do impressive things over the next four years.


#4

A clear next step is to allow the agents on one team to communicate. This will form a good model of society of mind think of the agents as elements of a brain solving a problem.