The multi-slot framework: A formal model for multiple, copiable AIs

Abstract

Because advanced AI is likely in the future, so is the possibility of multiple advanced AIs. It is therefore also likely that such advanced AIs will be implemented in software that can be copied from hardware to hardware. The best existing theoretical framework for the rigorous formal treatment and prediction of such AIs are those based on the AIXI framework developed by Hutter [2]. Unfortunately, these single-agent frameworks do not allow formal treatment of multiple co-existing AIs. The current paper introduces a novel "multi-slot" framework for dealing with multiple intelligent agents, each of which can be duplicated or deleted at each step, in arbitrarily complex environments. The framework is a foundational first step in the analysis of environments that allow creation (by copying) and deletion of multiple agents. Even by focusing on the case where the agents do not interact, the notion of future of an agent is not straightforward anymore, so we propose several such definitions, leading to value functions and AIXI-like agents. Finally, the framework is shown to be sufficiently general to allow for the existence of a universal environment that can simulate all environments in parallel. A companion paper uses the multi-slot framework presented here to explore the notion of identity in man and machine. © 2014 Springer International Publishing.