Programming Traffic: Modelling and Control of Microscopic Vehicle Manoeuvres

Abstract

This dissertation explores microscopic decision by human drivers and builds control frameworks that optimise microscopic actions of autonomous vehicles. Existing studies often treat mandatory lane-changes as absolute, neglecting the possible option of not performing them. As such, we explore the incentive behind the manoeuvre by analysing and modelling the cost for not performing lane-changes for exiting freeways. The developed mathematical model is deterministic and can estimate the exit-missing cost with simple network-level variables. Both simulation and empirical results indicate that exit missing costs are not substantial.

Driving is a two-dimensional task, yet studies have often neglected the interdependence between the two dimensions. This study integrates longitudinal car-following with lateral lane-keeping and discretionary lane-changing behaviours to create a realistic model that replicates human-driven trajectories. Recognising the hierarchical nature of human decision-making, we incorporate a game-theoretic reasoning structure and validate the model using a naturalistic trajectory dataset. Model calibration is conducted heterogeneously through a Bayesian framework. The model is able to replicate macroscopic traffic patterns with high accuracies and can generate smooth human-like trajectories.

Building on the two-dimensional model, we design an optimal controller for autonomous vehicles operating in a lane-free freeway environment. Vehicles are assumed to be communication-free, with each making decisions independently. A game-theoretic formulation ensures that vehicles account for the potential actions of others. The full framework is evaluated at both the microscopic and macroscopic scales. Experiments demonstrate how reasoning depth and vehicle-size heterogeneity influence overall traffic efficiency. The proposed lane-free controller is also compared against a similarly defined lane-based controller.