Input data -> Modelling -> Choice of optimal response -> Output data
I propose a simpler design:
Data <-> Modelling
Specifically, an automaton capable of actions having a model of the world (including itself), and some prior beliefs about how things such as itself behave:
- It has memories of its past (which may be somewhat inaccurate or incomplete).
- In the absence of reasons to the contrary, it believes automatons such as itself will tend to end up in situations where they are happy and not in pain -- they are intelligent and thus will probably find a way to wangle things to their advantage.
- From its memories it has a set of plausible models for how it is likely to act in any given situation.
- From its models it estimates how it will act in future and, to the extent that its memories are fuzzy, how it has acted in the past.
This is related to my work on image restoration.
The automaton also has a simple device that monitors this self modelling and acts out one of the scenarios for what the machine might do next. There is no attempt to choose an optimal behaviour, it just conforms to how it estimates that it will act. It acts to avoid pain, etc, only because it has a prior belief that it will do so.
Thus the self-model is the self. One less entity, and entities should not be multiplied beyond necessity.
This seems to have properties we associate with consciousness: consciousness of self, quining. It also fits with the idea of "flow" states, where one stops thinking of oneself as separate from the system of which you are a part.