Modelling urban street configurations for Autonomous Vehicle flows
File(s)
Author(s)
Ye, Qiming
Type
Thesis
Abstract
Despite the difficulties that have been encountered in the development and deployment of Autonomous Vehicles (AVs) at the current stage, they are still anticipated to become an essential mode of transportation at some point in the coming decades. A growing body of literature has shown that AVs can outperform Human-Driven Vehicles (HDVs) when it comes to fewer driving errors, the ability to predict driving behaviours, and the ability to comply with the increasingly complex traffic regulations in the future. The applications of critical technologies, such as lane-keeping and car-following, could potentially make platoon-based driving viable. That would result in a significant reduction in AV headway, reaction time, and average travel costs, while decreasing the number of road lanes that are required for traffic on the road. Due to the self-cruising and self-parking ability of AVs, novel solutions are being discussed to resolve the paradox of on-street parking problems as a result of the deployment of AV transport systems. A growing body of literature indicates that Shared AV (SAV) mobility will be essential for future Public Transport (PT) services, which are expected to use curbside spaces for parking intensively, frequently visit curbside for Pick-Up and Drop-Off (PUDO) passengers, and for fast-charging operations.
In light of the disruptions brought about by introducing AVs to open roads, more parties from academia, industry, and governments have become increasingly interested in visioning future AV-adaptive streets, which will not only be able to offer AV-based transportation but also encourage the use of various modes of Active Mobility (AM) as well. In order to successfully integrate AV transportation systems into the existing road network, intelligent management over the use of road space - for example, the assignment of road Right-Of-Way (ROWs) and the control of curbside spaces - would be a crucial topic of establishing efficient and AM-aware AV mobility solutions able to interact with the rest of the road traffic.
The present studies contributed to the design of AV-adaptive Complete Streets (CS), however, demonstrating a clear lack of operational and demand-responsive modelling frameworks to ensure that the allocation of road space during the day was coordinated as per the driving, parking, and PUDO demands of AV transport and other forms of transport. To fill these gaps, this thesis proposes three concrete modelling frameworks, aiming to incorporate control over road infrastructure with AV fleet operations.
In concrete terms, the first model designs a Reinforcement Learning (RL)-based model that evolves the configurations of road ROWs and assigns them to respective road users according to pedestrian flow and AV flow in real-time. The objective minimises travel costs and frees up more road space for AM. The second model optimises the layout of curbside parking lanes for SAVs stops in a network by solving a non-linear 0/1 Knapsack Problem (0/1-KP). The third modelling framework trains a RL model to evolve the assignment strategy for the real-time allocation of curbside parking space to diverse types of parking activities, i.e. on-street, PUDO operations and curbside delivery, given a real-world complex network.
To effectively solve these problems, specific metaheuristics and RL algorithms incorporating microscopic traffic simulation have been applied. Various model training strategies were examined, and these models were applied to solve the optima (near optima) under divergent scenarios, to improve the quality of the solution and higher algorithmic performances in respective modelling settings.
According to the results, the proposed method improved the allocation of ROWs in real-time and liberated more room for the use of AM road users with an equivalent to one driving lane width. For the second research problem, the proposed method reduced travel delays by 38.61 s/veh (24.37%) for the SAV fleet compared with respective benchmark conditions. Additionally, the third model balances curbside occupancy with demand-responsive (acceptance rates) to have improved the comprehensive curbside performance by 37.72% from the benchmark scenario. Modelling outcomes also revealed that for multi-agent systems, centralised and distributive training strategies were distinctively beneficial for tackling different problems. Particularly, for the ROWs assignment problem, the distributive strategy was superior, whereas the centralised one achieved better outcomes in the third research problem.
This thesis provides viable Artificial Intelligence (AI)-based tools to plan urban space under considerable uncertainties in the advent of autonomous transportation. The proposed modelling frameworks present novel approaches to the design and management of AV-adaptive streets, which have not been explored in the past. They have the potential to provide feasible and scalable technologies that dynamically assign ROWs to operational AVs and fine-grained curbside parking spaces to accommodate their parking and PUDO demands. Meanwhile, these models allow AVs to safely interact with other road users, especially pedestrians and ensure the efficiency of the transport system under diverse traffic scenarios.
In light of the disruptions brought about by introducing AVs to open roads, more parties from academia, industry, and governments have become increasingly interested in visioning future AV-adaptive streets, which will not only be able to offer AV-based transportation but also encourage the use of various modes of Active Mobility (AM) as well. In order to successfully integrate AV transportation systems into the existing road network, intelligent management over the use of road space - for example, the assignment of road Right-Of-Way (ROWs) and the control of curbside spaces - would be a crucial topic of establishing efficient and AM-aware AV mobility solutions able to interact with the rest of the road traffic.
The present studies contributed to the design of AV-adaptive Complete Streets (CS), however, demonstrating a clear lack of operational and demand-responsive modelling frameworks to ensure that the allocation of road space during the day was coordinated as per the driving, parking, and PUDO demands of AV transport and other forms of transport. To fill these gaps, this thesis proposes three concrete modelling frameworks, aiming to incorporate control over road infrastructure with AV fleet operations.
In concrete terms, the first model designs a Reinforcement Learning (RL)-based model that evolves the configurations of road ROWs and assigns them to respective road users according to pedestrian flow and AV flow in real-time. The objective minimises travel costs and frees up more road space for AM. The second model optimises the layout of curbside parking lanes for SAVs stops in a network by solving a non-linear 0/1 Knapsack Problem (0/1-KP). The third modelling framework trains a RL model to evolve the assignment strategy for the real-time allocation of curbside parking space to diverse types of parking activities, i.e. on-street, PUDO operations and curbside delivery, given a real-world complex network.
To effectively solve these problems, specific metaheuristics and RL algorithms incorporating microscopic traffic simulation have been applied. Various model training strategies were examined, and these models were applied to solve the optima (near optima) under divergent scenarios, to improve the quality of the solution and higher algorithmic performances in respective modelling settings.
According to the results, the proposed method improved the allocation of ROWs in real-time and liberated more room for the use of AM road users with an equivalent to one driving lane width. For the second research problem, the proposed method reduced travel delays by 38.61 s/veh (24.37%) for the SAV fleet compared with respective benchmark conditions. Additionally, the third model balances curbside occupancy with demand-responsive (acceptance rates) to have improved the comprehensive curbside performance by 37.72% from the benchmark scenario. Modelling outcomes also revealed that for multi-agent systems, centralised and distributive training strategies were distinctively beneficial for tackling different problems. Particularly, for the ROWs assignment problem, the distributive strategy was superior, whereas the centralised one achieved better outcomes in the third research problem.
This thesis provides viable Artificial Intelligence (AI)-based tools to plan urban space under considerable uncertainties in the advent of autonomous transportation. The proposed modelling frameworks present novel approaches to the design and management of AV-adaptive streets, which have not been explored in the past. They have the potential to provide feasible and scalable technologies that dynamically assign ROWs to operational AVs and fine-grained curbside parking spaces to accommodate their parking and PUDO demands. Meanwhile, these models allow AVs to safely interact with other road users, especially pedestrians and ensure the efficiency of the transport system under diverse traffic scenarios.
Version
Open Access
Date Issued
2023-03
Date Awarded
2023-10
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Angeloudis, Panagiotis
Stettler, Marc
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)