the problem of flat gradients in RL

[Blimpyway]

Hi,

I just read your article about this repository. I think key to (animal, human, artificial) intelligence is finding/learning/evolving good strategies to generate local or partial goals from sparse, higher level goals. Less pattern matching problems like "what I am looking at" and more of "what should I do next"

The following paper is an interesting example about building colored reward maps out of smell gradients

https://www.biorxiv.org/content/10.1101/2021.09.24.461751v1