Towards Adaptive Abstraction in Agent Based Simulation

Humans often switch between different levels of abstraction when reasoning about salient properties of systems with complex dynamics. In this paper, we study and compare multiple modelling and simulation techniques for switching between abstractions. This improves insight and explainability as well as simulation performance, while still producing identical answers to questions about properties. Traffic flow modelled using an Agent Based Simulation formalism is used to demonstrate the introduced concepts. The technique requires explicit models (1) of the dynamics of both individual cars and of emergent "jams", (2) of the conditions -often involving complex temporal patterns- under which switching between the levels of abstraction becomes possible/necessary and (3) of the state initialization after a switch. While aggregation is natural when going from detailed to abstract, the opposite direction requires additional state variables in the abstract model.