Learning to Share: A Study of Multi-agent Learning in Transportation Systems

  • Edmond Awad

Student thesis: Master's Thesis

Abstract

Intelligent transportation technologies such as car navigation systems can lead to signi?cant energy savings. One task of a navigation system is to help a driver take an ef?cient path to its destination. If we consider the Personal Rapid Transit (PRT) system in Masdar City, we can see that implementing an effective navigation system can play a pivotal role in improving its performance. In this study, we consider a non-centralized navigation system that seeks to route self-interested vehicles ef?ciently taking into consideration the congestion that might happen when most of the vehicles use one link over the others. This case is known in the literature as a congestion game. In a congestion game, each player chooses a resource (a link) out of many resources to use, and the cost of each resource depends on the number of agents who select it. This research aims to ?nd an algorithm that, when used in every vehicle, can secure each vehicle a fair cost, while maximizing the social welfare of the group. In so doing, we evaluate the performance of several traditional algorithms in a small, but intriguing network. Our results demonstrate the dif?culty of even this simple problem. We then propose and evaluate a family of algorithms on a simpler (and related) problem to better understand the kinds of learning biases algorithms should incorporate in this domain.
Date of Award2011
Original languageAmerican English
SupervisorJacob Crandall (Supervisor)

Keywords

  • Transportation -Planning

Cite this

'