AAAI Publications, Twenty-Third International FLAIRS Conference

Font Size: 
Evaluating Multi-Agent Traffic Controllers
Crystal Redman, Adele Howe

Last modified: 2010-05-06


Many multi-agent traffic control designs have been proposed; each makes different assumptions about the environment, which makes it difficult to compare their performance. In this paper, we present a testbed for flexible and consistent evaluation ofmulti-agent approaches to urban traffic control. The testbed, an extension of the MASON simulator, varies parameters such as traffic load, existence of arterial street and grid size. We instrument the testbed to collect a set of metrics defined in the literature: delay, normal travel time, number of stops, percent stopped, and wait time ratio. We implement three distinct, well known multi-agent traffic controllers and use the testbed to assess the impact of different scenarios by comparing agent performance. The different metrics highlight clear trade-offs between time and flow metrics, but the more challenging scenarios dilute the distinctions. The testbed supports evaluation, comparative analyses and hybridization of the approaches. We use our analyses to suggest modifications to the agents and show that the agent designs can be improved.


evaluation; traffic control agents; trade-off analysis

Full Text: PDF