News for hackers. Stuff that matters.

A comment on Generally capable agents emerge from open-ended play. Parent comment

Re:Generally capable agents emerge from open-ended play by eutectic on Tuesday July 27, 2021 @12:55PM (#27974669)
It's proven that if you want polynomial sample complexity in the size of the state space you need directed exploration. The algorithms 'flail' because they are initialized with random policies.
[ Parent | on HN ]
Re:Generally capable agents emerge from open-ended play by whimsicalism on Tuesday July 27, 2021 @09:44PM (#27979547)
Couldn't any directed exploration also be produced as the result of random flailing?

Or is this average sample complexity?

[ Parent | on HN ]
Re:Generally capable agents emerge from open-ended play by eutectic on Thursday July 29, 2021 @04:11PM (#28000877)
An example of a bad case for random exploration would be a narrow ridge where you die if you fall off and you only receive reward if you get to the end.

So it's a worst case result with respect to the MDP, but expected time/high probability with wrt random chance.

[ Parent | on HN ]