I do think this is a step towards an algorithmic way to make the right graph.
It’s an interesting question where the wrong graph would ever come from in the first place, given that we can not observe causation directly. If we are to run a bunch of copies of AIXI, for example, connected to a bunch of robotic arms, and let it observe arms moving in unison, each will learn that it controls all the arms. Representation of all the arms motions as independent would require extra data.
It’s an interesting question where the wrong graph would ever come from in the first place, given that we can not observe causation directly. If we are to run a bunch of copies of AIXI, for example, connected to a bunch of robotic arms, and let it observe arms moving in unison, each will learn that it controls all the arms. Representation of all the arms motions as independent would require extra data.