Simulation Fidelity, Abstraction and Resolution in Real-Time Multi-objective Optimisation of Air Traffic Complexity

Abstract

Understanding the role resolution, abstraction and fidelity play when solving problems is critical to the quality of decisions produced by automated systems. Resolution is the lens to see what is relevant and what is not in a system. Abstraction offers a mechanism to simplify systems by eliminating those factors that are not relevant to the phenomena of interest. Fidelity is a decision on the level of details in the data we need to have on those factors that are relevant. In particular, in real-time time-constrained environments, it is important to understand the relationship between resolution, abstraction and fidelity on the one hand, and the speed and accuracy to obtain a decision on the other hand. In this thesis, we will explore the effect of the level of resolution, abstraction and fidelity of simulators on decisions in the context of air traffic control. We design and use four simulators with different levels of abstraction and fidelity and compare their operation and output. We model reality with a very high resolution simulator that works at a higher level of fidelity than those used for comparison. This allows us to have a ground-truth to compare against. We then evaluate the effectiveness of the four simulators on optimising air traffic controllers task load in real-time. Each simulator is used to perform look-ahead operations within a multi-objective optimization algorithm to identify an aircraft-specific action to either reduce or increase complexity. Given that an air-traffic scenario has a minimum energy required to perform the task, the optimization finds opportunities to load-balance the workload over the time horizon of the scenario. This load balancing causes upward and downward shifts of complexity. This phenomenon is analysed in details in the thesis. Despite that a simulator may produce a large deviations from reality, if these deviations are systematic, we can predict it with a static model like an artificial neural network and use the prediction to correct for the simulator’s deviation. We conduct a series of analysis using artificial neural networks and linear regression to study the nature of the deviations. In summary, this thesis demonstrates that decisions on resolution, fidelity and abstraction have a great impact on performance. This impact can be studied and quantified. If used appropriately, it offers an evidence-based rational for the modeller to justify decisions made on resolution, abstraction and fidelity.