Self-Organized Middle-Out Abstraction

In this position paper we present a concept to automatically simplify computational processes in large-scale self-organizing multi-agent simulations. The fundamental idea is that groups of agents that exhibit predictable interaction patterns are temporarily subsumed by higher order agents with behaviours of lower computational costs. In this manner, hierarchies of meta-agents automatically abstract large-scale systems involving agents with in-depth behavioural descriptions, rendering the process of upfront simplification obsolete that is usually necessary in numerical approaches. Abstraction hierarchies are broken down again as soon as they become invalid, so that the loss of valuable process information due to simplification is minimized. We describe the algorithm and the representation, we argue for its general applicability and potential power and we underline the challenges that will need to be overcome.