Just came here to post this ;)

Obscuring parts of the whole picture. And having an agent attempt to "fill in the blanks". Seems to be the key to self-learning predictive systems.

In real world, high dimensional spaces, the unpredictability of "black swans" will certainly need to accounted for by some mechanism. Great work, and if you are not following @hardmaru on twitter you are certainly missing out!

Thanks! We still haven’t figured it a way to get this method to work directly for high dimensional obs space (just used a VAE to compress it for now)

This story sounds similar to the AI-plays-hide-and-seek one [1]. More abstract/general skills than detailed modeling.

[1] https://www.vox.com/future-perfect/2019/9/20/20872672/ai-lea...

Here’s the link to the paper referenced in that article: https://arxiv.org/abs/1909.07528