Automated Planning for Urban Traffic Control with LLM-Generated Configurations

Abstract

Automated planning approaches have proven effective in performing traffic signal optimisation, and their deployability has been demonstrated by their ability to incorporate constraints and features of the real-world infrastructure on which they will operate. A major constraint is the need to know in advance, for each junction of the controlled urban region, the set of configurations (i.e., the length of all stages) that can be considered for the optimisation process. Configurations therefore play a pivotal role as they effectively allow control of the traffic flows; their quality is of crucial importance.

In the literature, configurations have been generated synthetically or by leveraging historical data. In this paper, we explore the use of off-the-shelf Large Language Models (LLMs) to generate good-quality traffic signal configurations to address a range of traffic signal optimisation problems. LLMs hold the promise of generating unusual yet effective configurations with minimal human effort.