Solving a game of ordered preference consists of lexicographically minimizing for all the objectives in the desired order.
Games of ordered preference are a framework that can model agent interactions with multiple (and strictly ordered) objectives.
For example, autonomous vehicles want to stay on the lane while maintaining the desired speed, avoiding collisions, etc.
Excited to present our team's work this week in IEEE's International Conference on Intelligent Transport Systems (ITSC) !
We efficiently approximate optimal joint trajectories in games of ordered preference.
Approximate solutions to games of ordered preference: arxiv.org/abs/2507.11021 .
A 🧵:
We build on Dong Ho Lee's work on games of ordered preference: arxiv.org/pdf/2410.21447 , which had exponential growth as players had more objectives.
Our contribution is a method of efficiently approximating joint solutions to games of ordered preference using Lexicographic IBR over time.
Video
Lexicographic Iterative Best Response over time solves games of ordered preference in a receding horizon setting and uses the previously computed (but not executed) actions to warm-start the solver.
Combined with iterative best response, we end up solving a multiple smaller (and easier) problems.
Test results show us that computed solutions approximate the baseline with linear growth!
Thanks to @paudelasheras.bsky.social , Dong Ho Lee, Lasse Peters, David Fridovich-Keil, and @exomorphic.bsky.social for this fantastic collaboration! 🥳
arxiv.org
The problem is that in complex environments (like intersections) agents can't comply with all these objectives, and they have to give up on something. Games of ordered preference let the agents decide which objective has a higher preference among the others.
