I'm reminded of a paper about an AI system with curiosity. Basically, the AI would learn to predict the effect of its actions, and would prefer to take actions

I'm reminded of a paper about an AI system with curiosity. Basically, the AI would learn to predict the effect of its actions, and would prefer to take actions where its prediction error was large. That way it would gather more data about stuff it didn't know yet. This works well to learn to play video games without defining an extrinsic reward. However, there is a problem when the system encounters the equivalent of TV. It then gets stuck watching the unpredictable stream of events. paper and code: https://pathak22.github.io/large-scale-curiosity/ https://pathak22.github.io/large-scale-curiosity/ Coverage by "2 minute papers" youtube channel: https://youtu.be/fzuYEStsQxc https://youtu.be/fzuYEStsQxc