one of my favourite ICLR2019 talks. Learn to explore by using learning progress as reward. The idea makes so much sense, I don't know why I don't see it applied often? Maybe the devil is in the details? They demonstrate impressive results on many things I've wondered about like: learning to use tools, language development, effect of body morphology, action parametrization

https://www.youtube.com/watch?v=L8li6vWPhwc

https://scholar.google.com/citations?hl=en&user=gCqGj4sAAAAJ&view_op=list_works&sortby=pubdate . See more here in these iclr2019 notes: https://david-abel.github.io/notes/iclr_2019.pdf

Intrinsic Motivation Systems for Autonomous Mental Development

Information-seeking, curiosity, and attention: computational and neural mechanisms

Learning to play Mario by Curiosity

Trying to visit parts of state/action space that are hard to predict, or that haven't been visited before (work by Lillycrap on iclr2019)

Curiosity in RL

PowerPlay

https://www.sciencedirect.com/science/article/pii/S1364661313002052

Relations to Metalearning