I was curious if you had any thoughts on the relationship with how this might change if you imagine that each agent also has a world model?
From the Active Inference perspective the reason why we have something like a self other boundary is because it is an artefact of trying to create stable representations of the world where you have a reference point (yourself) to plan around.
There should be a sort of intersection here where having a coherent identity in terms of action policies make it easier for you to plan and predict future action policies. If you imagine a temporal intrapersonal agent bargaining setup between different submodules the advantage of a consistent self model should creates more selection pressure towards predictability than the pure equilibria picture might suggest?
I’m also imagining that there might be specific subparts of action and cognition where keeping the strategies the same (updatelessness) is more beneficial than for other parts due to the compressed representation that is a self model.
My aim is to work towards a paradigm where allying under a coherent identity is a demonstrably “reasonable” development for sub-agents or TIs subject to selection pressures. I agree that one of these pressures is likely predictability or, almost equivalently, a compressable and simple self-model.
I hint at what this could look like in the section “Robust equilibria and updatelessness”. The shape of identity (e.g. how updateless the agent allows itself to be) may dictate not only how TIs interact with each other, but also how they will navigate strategic interactions with TIs belonging to other agents. This would ideally mean that a theory of intrapersonal games in an agent determines what solution concepts are valid for that agent in usual game theory.
This addresses your point in the specific case where the rest of the world consists of other agents; I’m not as clear on the relationship intrapersonal games have to “world models”. Food for thought, thanks.
I like this line of thinking.
I was curious if you had any thoughts on the relationship with how this might change if you imagine that each agent also has a world model?
From the Active Inference perspective the reason why we have something like a self other boundary is because it is an artefact of trying to create stable representations of the world where you have a reference point (yourself) to plan around.
There should be a sort of intersection here where having a coherent identity in terms of action policies make it easier for you to plan and predict future action policies. If you imagine a temporal intrapersonal agent bargaining setup between different submodules the advantage of a consistent self model should creates more selection pressure towards predictability than the pure equilibria picture might suggest?
I’m also imagining that there might be specific subparts of action and cognition where keeping the strategies the same (updatelessness) is more beneficial than for other parts due to the compressed representation that is a self model.
Fun read, thanks.
My aim is to work towards a paradigm where allying under a coherent identity is a demonstrably “reasonable” development for sub-agents or TIs subject to selection pressures. I agree that one of these pressures is likely predictability or, almost equivalently, a compressable and simple self-model.
I hint at what this could look like in the section “Robust equilibria and updatelessness”. The shape of identity (e.g. how updateless the agent allows itself to be) may dictate not only how TIs interact with each other, but also how they will navigate strategic interactions with TIs belonging to other agents. This would ideally mean that a theory of intrapersonal games in an agent determines what solution concepts are valid for that agent in usual game theory.
This addresses your point in the specific case where the rest of the world consists of other agents; I’m not as clear on the relationship intrapersonal games have to “world models”. Food for thought, thanks.