Towards autonomic urban traffic control with collaborative multi-policy reinforcement learning
| Autor: | Ivana Dusparic, Vinny Cahill, Julien Monteil |
|---|---|
| Rok vydání: | 2016 |
| Předmět: |
050210 logistics & transportation
Engineering business.industry 05 social sciences Real-time computing 02 engineering and technology computer.software_genre VisSim Set (abstract data type) Intelligent agent Urban traffic control 0502 economics and business Traffic optimization Scalability 0202 electrical engineering electronic engineering information engineering Reinforcement learning 020201 artificial intelligence & image processing business Adaptation (computer science) computer Simulation computer.programming_language |
| Zdroj: | ITSC |
| Popis: | Various multi-agent decentralized approaches based on reinforcement learning (RL) have been proposed to increase scalability and real-time adaptiveness of urban traffic control (UTC) systems. In such approaches, traffic light control parameters are not pre-defined, but intelligent agents controlling the junctions learn the suitable traffic signal settings. In order to consider applications of RL in commercial UTC products, they need to enable fine-grained optimization of multiple traffic objectives and be validated in realistic UTC simulations. This paper presents REALT, a UTC system based on Distributed W-Learning (DWL), a multi-policy multi-agent RL-based optimization technique, which enables it to address multiple traffic optimization goals simultaneously. We introduce an extension of DWL with fine-grained traffic observation to enable adaptation of phase duration rather than just phase selection. We also evaluate the impact of action set selection in RL applications to UTC, by introducing a phase generation module which enables REALT to use different phase sets. We simulate REALT performance in VISSIM, on a detailed model of the road network representingWestern Road, the main traffic artery in Cork, Ireland. We use precise historic traffic counts as input and present results of the comparison of REALT performance to that of SCOOT signals deployed in Cork. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|