I’m more interested in games that don’t have explicit rewards. A fun project might be to train an agent to play, say, Minecraft or SimCity, and optimise for novelty and curiosity, to see what it is capable of producing under game constraints such as resource limits and survival.https://twitter.com/mark_riedl/status/1031328324876988416 …
-
I have some initial results on pure exploration for SNES games. Even outside of not-quite-games like Mario Paint, algorithms still need to traverse menus to find and opt-into the various the core gameplay modes.
I very much like the "curiosity driven exploration" approach but I wonder how it would perform in games where one would get rewards in certain situations purely based on luck. I think an AI might get addicted to gambling in such a setting, very similar to human behavior.
