Causal Entropic Force based AGI

[BrendanBenshoof]

There have been a LOT of agi architecture proposals over the years, and so far none of them have panned out. The purpose of this proposal is not to purport to be some final solution but rather to explore the viability of an approach.

the concept of Causal Entropic Forces purports that behavior that we consider intelligent in fact maximizes the entropy of possible future states. My largely incorrect metaphor for this process is "prevent the universe from becoming boring or entering into a terminal state (terminal cycles might be ok).

In order for this to be a viable approach, we require two parts:

• The model of the universe is an "imagination" for the agi, that would allow it to consider consequences of given actions.
• Entropy maximizer, which strives to select actions that maximize future entropy

such an AGI model would learn on two fronts, it would need to maintain it's model of the universe in reference to it's observations and it would seek and record heuristic strategies based on it's model of the universe (refining these strategies based observed results)

[BrendanBenshoof]

The hard part of using this type of method, is a scalable mechanism for approximating the future entropy of a given action/state.

I think we could map states + actions as a transition graph, then if we embed the graph (particularly into hyperbolic or other spaces without translation) then the states closest to the epicenter of the space would have the most entropy.