Real-time Delay Prediction and Quality of Service Assurance for Traffic Management in Service Systems

Ensuring quality of service (QoS) is one of the important challenges facing service providers. A key metric of the QoS is the experienced delay by the customers, which can depend on various factors such as the demand patterns, service capacity and the scheduling discipline. In this thesis, we study delay distribution prediction in service networks and propose novel QoS assurance methods which target the mentioned factors. In particular, the proposed methods require no knowledge of the system model or parameters, which is an important feature for real-world applications. We first consider the delay distribution prediction problem in tandem and acyclic queueing networks. Our analytical results suggest that the Gaussian mixture model can be a good candidate for estimating the end-to-end delay distribution in these systems. Motivated by this result, we use mixture density networks to propose a delay distribution predictor based on queue length information, which requires no knowledge of the system model or parameters. As the next step, we take a reinforcement learning (RL) approach and propose an admission controller with the goal of providing probabilistic upper-bounds on the end-to-end delay of the accepted jobs, while minimizing the probability of unnecessary rejections. Since admission control might not be a viable solution in some applications, we propose a deep RL-based service-rate controller as an alternative, which is capable of providing probabilistic upper-bounds on the end-to-end delay of the system by dynamic adjustment of the service rates. In the last part of this thesis, we study urban traffic management, as another important example of service systems. Specifically, we introduce two notions of fairness in this context, which are concerned with the delay of the vehicles and throughput of the traffic flows at an intersection. This is particularly important asneglecting fairness can lead to situations where some vehicles experience unacceptable long delays, or where the throughput of a particular traffic flow is highly impacted by the fluctuations of another conflicting flow. Finally, we propose two traffic signal control methods for implementing these fairness notions, which also consider the efficiency of the system at the same time.

Department dc:contributor.department

Electrical and Computer Engineering

Year dc:date.issued

2021

Author dc:creator

• Raeis, Majid

Advisor dc:contributor.advisor

• Leon-Garcia, Alberto

• Quality of service
• queueing
• Reinforcement learning
• service systems

Handle dc:identifier.uri

http://hdl.handle.net/1807/108721

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/108721