Active Learning Metamodels for ATM Simulation Modeling

Transportation systems are particularly prone to
exhibiting overwhelming complexity on account of the numerous
involved variables and their interrelationships, unknown stochastic phenomena, and ultimately human behavior. Simulation
approaches are commonly used tools to describe and study such
intricate real-world systems. Despite their obvious advantages,
simulation models can still end up being quite complex themselves. The field of Air Traffic Management (ATM) modeling
is no stranger to such concerns, as it traditionally involves
laborious and systematic analyses built upon computationally
heavy simulation models. This rather frequent shortcoming can
be addressed by employing simulation metamodels combined
with active learning strategies to approximate the input-output
mappings inherently defined by the simulation models in an
efficient way.
In this work, we propose an exploration framework that
integrates active learning and simulation metamodeling in a
single unified approach to address recurrent computational
bottlenecks typically associated with intense performance impact
assessments within the field of ATM. Our methodology is designed to systematically explore the simulation input space in
an efficient and self-guided manner, ultimately providing ATM
practitioners with meaningful insights concerning the simulation
models under study. Using a fully developed state-of-the-art ATM
simulator and employing a Gaussian Process as a metamodel,
we show that active learning is indeed capable of enhancing
both the modeling and performances of simulation metamodeling
by strategically avoiding redundant computer experiments and
predicting simulation outputs values.