CBLab Supporting the Training of Large-scale Traffic Control Policies with Scalable Traffic Simulation_2

2025-04-27 0 0 4.48MB 13 页 10玖币

侵权投诉

CBLab: Supporting the Training of Large-scale Traic Control

Policies with Scalable Traic Simulation

Chumeng Liang

caradryan2022@gmail.com

Shanghai Jiao Tong University

China

Zherui Huang

huangzherui@sjtu.edu.cn

Shanghai Jiao Tong University

China

Yicheng Liu

liuyicheng1515@sjtu.edu.cn

Shanghai Jiao Tong University

China

Zhanyu Liu

zhyliu00@sjtu.edu.cn

Shanghai Jiao Tong University

China

Guanjie Zheng∗

gjzheng@sjtu.edu.cn

Shanghai Jiao Tong University

China

Hanyuan Shi

shihanyuan@sjtu.edu.cn

Independent Researchers

China

Kan Wu

kanwu@zhejianglab.com

Research Center for Intelligent

Transportation, Zhejiang Lab

China

Yuhao Du

apiadu17a6@gmail.com

Independent Researchers

China

Fuliang Li

tjfulianglee@gmail.com

Baidu

China

Zhenhui Li

jessielzh@gmail.com

Yunqi Academy of Engineering

China

ABSTRACT

Trac simulation provides interactive data for the optimization

of trac control policies. However, existing trac simulators are

limited by their lack of scalability and shortage in input data, which

prevents them from generating interactive data from trac simula-

tion in the scenarios of real large-scale city road networks.

In this paper, we present City Brain Lab, a toolkit for scalable

trac simulation. CBLab consists of three components: CBEngine,

CBData, and CBScenario. CBEngine is a highly ecient simulator

supporting large-scale trac simulation. CBData includes a trac

dataset with road network data of 100 cities all around the world.

We also develop a pipeline to conduct a one-click transformation

from raw road networks to input data of our trac simulation.

Combining CBEngine and CBData allows researchers to run scal-

able trac simulations in the road network of real large-scale cities.

Based on that, CBScenario implements an interactive environment

and a benchmark for two scenarios of trac control policies respec-

tively, with which trac control policies adaptable for large-scale

urban trac can be trained and tuned. To the best of our knowledge,

CBLab is the rst infrastructure supporting trac control policy

optimization in large-scale urban scenarios. CBLab has supported

∗corresponding author

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than the

author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or

republish, to post on servers or to redistribute to lists, requires prior specic permission

and/or a fee. Request permissions from permissions@acm.org.

KDD ’23, August 6–10, 2023, Long Beach, CA, USA

ACM ISBN 979-8-4007-0103-0/23/08. . . $15.00

https://doi.org/10.1145/3580305.3599789

the City Brain Challenge @ KDD CUP 2021. The project is available

on GitHub: https://github.com/CityBrainLab/CityBrainLab.git.

CCS CONCEPTS

•Information systems →Spatial-temporal systems.

KEYWORDS

Trac Simulation, Trac Control Policy, Trac Signal Control,

Large-scale Data

ACM Reference Format:

Chumeng Liang, Zherui Huang, Yicheng Liu, Zhanyu Liu, Guanjie Zheng,

Hanyuan Shi, Kan Wu, Yuhao Du, Fuliang Li, and Zhenhui Li. 2023. CBLab:

Supporting the Training of Large-scale Trac Control Policies with Scalable

Trac Simulation. In Proceedings of the 29th ACM SIGKDD Conference on

Knowledge Discovery and Data Mining (KDD ’23), August 6–10, 2023, Long

Beach, CA, USA. ACM, New York, NY, USA, 13 pages. https://doi.org/10.

1145/3580305.3599789

1 INTRODUCTION

Well-crafted trac control policies, such as trac signal control and

congestion pricing, are expected to improve the eciency of urban

transportation. In recent years, many studies have been conducted

to optimize the trac control policies according to real-time trac

data [

–

]. These policies depend on

data generated by interaction with the trac environment where

they explore to make good decisions under dierent consequences.

However, real-world urban trac cannot provide enough inter-

active data to train these policies, because the exploration of the

policy may have a toxic impact on the urban trac e.g. provoke

severe congestion. Trac simulators are therefore born as alterna-

tives to provide trac environments for trac control policies to

arXiv:2210.00896v2 [physics.soc-ph] 5 Jun 2023

KDD ’23, August 6–10, 2023, Long Beach, CA, USA Liang et al.

Raw Traffic Data Traffic Control

Policies

CBData CBEngine CBScenario

Input Data Observation

Action

Store Train

1 2

Deploy

CBLab

Simulator

Data Tool Environment

Figure 1: An overview of CBLab.

interact with. These simulators [

] simulate the microscopic

evolution of the urban trac. For each time step, they describe

the trac state, obtain a trac action from the decision made by

trac control policies and make it happen in the simulation. Trac

control policies can then learn from how the trac evolves under

certain actions and improve decision-making.

While existing trac simulators help hatch various trac con-

trol policies successfully, they still come with drawbacks. Current

simulators, as they were designed primitively, support simulation in

road networks smaller than one hundred intersections and cannot

scale to city-level trac, which involves thousands of intersections.

Due to limits in eciency and scalability, these simulators are either

not able to conduct a city-level simulation in a feasible time or set

to prevent masses of vehicles from coming in the trac.

Another concern lies in the shortage of input data for large-scale

trac simulation. Although the map data of the main cities in the

world is completed, there is an absence of infrastructure for access to

the map data and a pipeline to transform it into simulation inputs.

Therefore, inputs for trac simulation only come from manual

work and are limited to a small set of road networks [

]

whose scales are often dozens of intersections (e.g. 4x3 or 4x4) -

much smaller than real urban road networks.

To overcome two aforementioned drawbacks, we propose City

Brain Lab, a novel toolkit for scalable trac simulation. CBLab con-

sists of three components: a microscopic trac simulator CBEngine,

a data tool CBData, and a trac control policy environment CBSce-

nario (See Figure 1). Beneting from well-designed parallelization,

CBEngine is of high eciency and scalability and is capable of

running the trac simulation on the scale of 10,000 intersections

and 100,000 vehicles with a real-simulation time ratio of 1:4 with or-

dinary computing hardware. CBData includes an accessible dataset

that contains raw road networks of 100 main cities all around the

world. A pipeline is prepared to automatically transform the raw

data into input data for trac simulation. Combining CBEngine

and CBData, users can easily start up trac simulation on real

city-level road networks. Based on the scalable trac simulation,

we implement CBScenario as an environment for two trac control

policies: trac signal control and congestion pricing. Users can de-

sign, develop, and train trac control policies in the framework of

CBScenario. To the best of our knowledge, we are the rst to provide

infrastructure for large-scale trac control policy optimization.

Our contribution can be summarized as follows.

•

We develop a scalable trac simulator CBEngine that sup-

ports city-level microscopic trac simulation for the rst

time.

•

We develop a data tool CBData to provide input data for

large-scale trac simulation.

•

We implement an interactive environment CBScenario for

training trac control policies under a large-scale setting.

The original version of CBLab has successfully supported the

City Brain Challenge @ KDD CUP 2021 with 1,156 participating

teams. See Appendix C for details.

2 CBENGINE: CITY-SCALE TRAFFIC

SIMULATION ENGINE

In this section, we introduce CBEngine, the trac simulator. We

demonstrate its trac modeling and conduct extensive experiments

to show the eciency, scalability, and plausibility of CBEngine.

2.1 Overview of Simulation Modeling

Trac simulators take the road network and the trac ow (vehi-

cles in the trac) as inputs and aim to simulate their interaction.

Road networks describe the topology of roads and intersections.

Trac ows describe the origins, destinations, and routes of vehi-

cles. As shown in Figure 2, the road network interacts with vehicles

through trac signal lights, which control the passing of vehicles

at intersections. When the simulation starts, vehicles in the trac

ow set out from their origins, travel down the routes, and nally

arrive at their destinations. Roads, intersections, trac signal lights,

and vehicles can be considered as trac elements.

We can formulate the simulation in the form of states and actions.

The simulator holds states for trac elements at the time step

𝑡

For the vehicle, the state is its location and speed. For the trac

signal light, the state is the current signal. Actions of elements then

iterate the current state to that in the next time step. For example,

the acceleration of a vehicle rises up the vehicle’s speed at the

next time step. Finally, the simulation moves on to the next time

step, where a new state-action iteration will begin. Let

𝑠𝑡

𝑎𝑡

, and

𝑠𝑡+1

denote the current state, action, and the next state, the trac

modeling in the simulation can be concluded by Eq 1.

𝑠𝑡+𝑎𝑡→𝑠𝑡+1(1)

2.2 Road Network

In CBEngine, Road Network is the topological network where ve-

hicles drive. It consists of two components: roads and intersections.

Road models the road segment in the real-world road network.

A road may include multiple lanes. Each lane holds one or more

vehicles. Intersection is the nexus of dierent roads. Through lane

links in the intersection, lanes of dierent roads connect to each

CBLab: Supporting the Training of Large-scale Traic Control Policies with Scalable Traic Simulation KDD ’23, August 6–10, 2023, Long Beach, CA, USA

Vehicle

Routing Driving

CBEngine Model

Traffic Signal

1 2 34

567 8      

1 2 34

567 8

action : change the signal phase from 1 to 4

step

tt+t

action : change the route action : change the acceleration

step step

t+ t t+ t

Figure 2: The Trac Model of CBEngine

other. Trac signal light is another key element in the intersec-

tion, which assigns a true-or-false signal for each lane link. Vehicles

can only move from one road to another through available lane

links with a true signal.

2.3 Driving Model

Behaviors of vehicles are controlled by the driving model in the

trac simulation. Driving models determine how vehicles move on

the road. Towards simulating the behaviors of vehicles, researchers

have proposed various models [8, 24, 25].

The default driving model of CBEngine is a modied version

of the driving model used in Cityow [

], originating from the

driving model proposed by Stefan Krauß [

]. The key idea is that:

the vehicle will drive as fast as possible subject to safety regular-

ization. Specically, the maximum speed of the vehicle is subject

to several static or dynamic speed constraints. Vehicles will accel-

erate or decelerate to the speed with the maximum acceleration

or deceleration. In the implementation, the maximum acceleration

(deceleration) is a hyperparameter and is open to users to tune with

an API. The considered speed constraints are listed and discussed

respectively as follows:

•Road speed limit

•Collision-free following & leading speed

•Cutting-in collision-free speed

•Trac-signal safe speed

Road speed limit. Each road has its own speed limit. This is a

static speed constraint.

Collision-free following & leading speed. To avoid collisions, ve-

hicles need to adapt their speed to the speed of their following

and leading vehicles. We use the collision-free following speed to

model the max speed constrained by the leading vehicle. Following

the driving model in Cityow, we compute these two constraints

with Eq 2. It takes

𝑣

current speed of the vehicle,

𝑣𝐿

current speed

of the leading vehicle,

𝑑

maximum deceleration of the vehicle,

𝑑𝐿

maximum deceleration of the leading vehicle,

𝐷

the current dis-

tance between two vehicles,

𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙

the length of each time step

as parameters to compute the collision-free following speed 𝑠𝑐 𝑓 𝑠 .

𝑎=

2·𝑑, 𝑏 =

𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙

𝑐=

𝑣·𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙

2−𝑣2

𝐿

2·𝑑𝐿−𝐷

𝑠𝑐 𝑓 𝑠 =−𝑏+√𝑏2−4𝑎𝑐

2·𝑎

(2)

To compute the collision-free leading speed

𝑠𝑐𝑙𝑠

, we use a mirror

equation of Eq 2. We use the

𝑑𝐹

the maximum deceleration of

the following vehicle and

𝑣𝐹

the speed of the following vehicle

to replace

𝑑𝐿

and

𝑣𝐿

in the Eq 2. The collision-free following and

leading speed constraints can be summarized as Eq 3.

𝑠𝑐𝑙𝑠 ≤𝑣≤𝑠𝑐 𝑓 𝑠 (3)

Cutting-in collision-free speed. CBEngine supports self-adaptive

lane changing within the road (Cutting-in). This asks for cutting-in

collision-free following and leading speed, avoiding collisions with

the leading and following vehicles in the target lane. We compute

this constraint using Eq 2 with

𝑣𝐿

𝑑𝐿

, and

𝐷

given by the leading

vehicle in the target lane and

𝑣𝐹

𝑑𝐹

, and

𝐷

given by the following

vehicle in the target lane.

An exception happens when the vehicle needs to conduct an

emergent cutting-in. This takes place when the vehicle is very close

to the intersection but still in the wrong lane (e.g. The vehicle is in

the go-straight lane but needs to turn left). On this occasion, the

vehicle tries to stop to wait until it is able to change the lane. There-

fore, the maximum speed constraint equals zero and the vehicle

will decelerate to zero with its maximum deceleration.

Trac-signal safe speed. Vehicles heading for an intersection are

subject to two constraints determined by the trac signal. First,

the current speed can be decelerated to zero within the remained

passing time of the trac signal. Second, the driving distance cannot

exceed the distance to the intersection, under the decelerating

process dened in the rst constraint.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CBLab:SupportingtheTrainingofLarge-scaleTrafficControlPolicieswithScalableTrafficSimulationChumengLiangcaradryan2022@gmail.comShanghaiJiaoTongUniversityChinaZheruiHuanghuangzherui@sjtu.edu.cnShanghaiJiaoTongUniversityChinaYichengLiuliuyicheng1515@sjtu.edu.cnShanghaiJiaoTongUniversityChinaZhanyuLiuzh...

展开>> 收起<<

CBLab Supporting the Training of Large-scale Traffic Control Policies with Scalable Traffic Simulation_2.pdf

共13页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

CBLab Supporting the Training of Large-scale Traffic Control Policies with Scalable Traffic Simulation_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: