This repository contains the LaTeX source for my master's thesis.
Autonomous systems often operate under multiple objectives that cannot be naturally reduced to a single scalar cost. This is especially true in autonomous driving, where safety, rule compliance, lane keeping, and progress towards a destination must be prioritized differently depending on the situation. This thesis studies that problem through games of ordered preference, a game-theoretic framework in which each agent ranks its objectives lexicographically.
The thesis focuses on three aspects required for the practical use of this framework. First, it develops lexicographic IBR over time, an approximate receding-horizon solution method that exploits temporal continuity to reduce the computational burden of solving games of ordered preference while preserving the qualitative influence of preference order on behavior. Results in simulated traffic scenarios show substantial reductions in compilation and solution times relative to the baseline formulation, especially as the number of preference levels increases. Second, it formulates the inverse problem of estimating the preference order of other agents from observations. Whereas prior inverse-game methods can recover latent goals or objective functions, this thesis targets the ordering of objectives itself, which is the defining information in games of ordered preference, and evaluates initial offline and online estimators. Third, it introduces Duckietown as a low-cost robotic platform for experimental validation under sensing noise, actuation errors, and other real-world imperfections. The experiments show that the framework can already run in a near-real-time physical loop, although with limited timing slack and a visible gap between planned acceleration-domain trajectories and the motion executed through the robot's native interface.
Taken together, the thesis shows that games of ordered preference provide a promising framework for modelling interactive decision making with prioritized objectives, and that receding-horizon approximations can make their use considerably more tractable. At the same time, it shows that practical preference-order estimation depends on both observation richness and candidate quality, and identifies broader sim-to-real validation, faster replanning, and tighter actuation matching as essential directions for future work.
include/: custom LaTeX class and template utilities.chapters/: chapter-by-chapter thesis content.images/: local images.misc/: bibliography and extra LaTeX snippets.thesis.tex: main manuscript entry point.Makefile: PDF build commands.presentation/: The thesis presentation.
Generate the PDF locally, or download it from GitHub Actions artifacts or the releases page.
Simply install latex and other packages. If I'm not missing anything:
sudo apt install texlive-full biber libpgf6 texlive-lang-spanish inkscapeEasy!
makeThis generates thesis.pdf in the repository root.
You can run make redo to generate a pdf even if one has already been created.
Even easier!
make presentationThis generates presentation.pdf in the repository root.