| Summary: | The upcoming high population density rise in metropolitan areas is anticipated to further deteriorate the traffic conditions. To tackle this problem, advanced ICT applications have been employed, able to monitor and manage traffic in real time. In practice, to efficiently correspond to dynamic traffic conditions those applications require to be frequently reconfigured – an operation that usually involves expert-teams manually adjusting the traffic-regulating strategies regularly. However, these manual procedures are not adequately aligned with the traffic situation since complicated stochastic dynamics, model unavailability and data inner-transmission constraints usually emerge. In order to overcome such cumbersome and expensive adjustment procedures modern decentralized adaptive optimization is widely accepted and recognized as an efficient automated solution for tuning the control strategy on-the-fly. Motivated by the above, L4GCAO, a decentralized, model independent, flexible optimization technique has been designed for optimizing cycle management at a local level to improve network performance at the global level, by automatically adjusting the cycle-regulating parameters in an intersection-centric manner, through cooperating self-learning agents.This paper studies L4GCAO's first application on a realistic traffic-network simulation scheme that examines the online fine-tuning process of the cycle-regulating parameters. Moreover, in order to evaluate the decentralized L4GCAO performance, two levels of performance benchmarking have been considered: a comparison with CAO - its well-established centralized counterpart; an already well-designed fixed-time management plan. In all cases, L4GCAO exhibits an almost equivalent performance improvement compared to CAO, both with respect to a properly fixed-time traffic management plan, while utilizes less parameters in a non-centralized manner.
|
|---|
