11 comments

[ 6.1 ms ] story [ 92.6 ms ] thread
Neat, I don't understand what they mean by having embedded a reward video into the set. Is that a video where copying the behaviour will deliver victory?
Yes, they take the state of the video every 16 frames and look at its embedding. These were made into checkpoints.

The AI is rewarded if at each checkpoint the state vector its produced is sufficiently aligned with the videos.

I guess that's the initial training to deal with sparse rewards.

This should probably say "ML" or "AI" or whatever, I was slightly disappointed to realize it was not a funny paper about… I don't know to be fair.
I can see where you're coming from, the title definitely made me initially feel like it was going to be about getting satisfaction from watching other people play a game or something like that.
This is really cool. A step in the right direction towards general learning through observation.
Are audio cues also analyzed here? ie: "We observe that use of the audio signal in CMC results in more emphasis being placed on key items and their location in the inventory"
Yes. s3.2 suggests they use audio cues to help them align the video frames from different videos. (I guess it's easier to correlate audio than video.)
I can imagine video quality on youtube varies more than audio, or that audio is easier to hash / make signatures of.
Hmm I was more under the impression it was for context, in other words creating/executing strategies based on what audio cues were received (like when a key or coin is acquired) and keeping tabs of what actions the user performed after that point.
This is actually quite human. I also watch Let's plays if I struggle with a quest (or game in general).

Also interesting assumption to say "harder = fewer rewards". Probably doesn't always apply but is a good generalization.