Several tutorials have been accepted and will be scheduled in the conference program. Here are their details and speakers. The program will be published in some weeks, once the peer review process of technical contributions is over and notifications have been issued to the authors. Stay tuned for updates, and do not forget to register to attend these tutorials!
Tutorial 1:
Data-driven techniques on incident detection for freeways
Motivation and objectives:
Incident detection is a key component in real-time traffic management systems that allows efficient response plan generation and decision making by means of risk alerts at critical affected sections in the network. State-of-the-art incident detection techniques traditionally require: i) good quality data from closely located sensor pairs, ii) a minimum of two reliable measurements from the flow occupancy-speed triad, and iii) supervised adjustment of thresholds that will trigger anomalous traffic states.
Despite such requirements may be reasonably achieved in simulated scenarios, realtime downstream applications rarely work under such ideal conditions and must deal with low reliability data, missing measurements, and scarcity of curated incident labelled datasets, among
other challenges. Non-recurrent situations, such as incidents, appear as outliers or anomalies in the real-time traffic data feeds.
Hence, outlier detection becomes a key task for road traffic incident detection systems that has been considered in the state of the art. This tutorial aims at providing a brief theoretical framework on both traditional and data-driven methods for traffic data analysis, pattern identification, detecting incidents, as well as hands-on practice on specific algorithms and machine learning techniques.
- Topics: AI, Machine Learning and Deep learning for ITS, management of incidents and evacuation
- Format: Half day (4 hours)
Intended audience:
PhD candidates and practitioners in the field of ITS. The workshop will give the participants, through a hands-on exercise, an opportunity to understand and contribute to the challenges related to freeway mobility and improving transport resilience. In the era of Intelligent Transport Systems (ITS), technologies such as advanced traffic management systems and travel information systems can benefit from the information provided by the data. The availability of rich data derived from diverse data sources can better support and improve the capabilities of traffic models to evaluate such systems. In the context of incident detection, rich and diverse data can enhance the capabilities of data-driven solutions to predict accurately and timely network anomalies.
All necessary material (data sets, network characteristics, Python scripts) and guidelines will be prepared by the organizers and shared in advance to the registered participants for the workshop in order to have time to familiarize with the data set, network, etc. During the workshop an introduction will be first provided on anomaly detections and traffic data analysis using machine learning techniques. Following, the process of cleaning the data and classification of traffic patterns will be presented. The exercise can include the processes mentioned above.
Materials and equipment needed for the tutorial:
Python Jupyter notebooks using Google Colaboratory services (https://colab.google).
Tutorial agenda:
14:30:Start
- Introduction to data-driven methods for traffic data analysis (outlier removal, traffic pattern identification) (30 minutes) (Athina Tympakianaki, Antonio Pellicer)
- Hands-on exercises for data cleaning and descriptive analysis using provided data sets and scripts (1hour) (Antonio Pellicer)
- Introduction to statistical and machine learning methods for incident detection (30 minutes) (Monica Dominguez, Mahdi Rahimiasl)
Break 16:30 – 17:00
Continue 17:00 – 18:45
- Hands-on exercises on incident detection (traffic anomalies) using provided data sets and scripts (1,5 hours) (Monica Dominguez, Mahdi Rahimiasl, Antonio Pellicer, Athina Tympakianaki)
- Discussion of results (15 minutes)
Speakers:
Academic partners from the Horizon2020 projects FRONTIER (https://www.frontierproject.eu) and TANGENT (https://tangent-h2020.eu) will also participate in the tutorial.
Athina Tympakianaki, Aimsun, athina.tympakianaki@aimsun.com
Athina has twelve years’ experience as a researcher and traffic analyst in academia and industry, with expertise in calibrating traffic simulation models and particularly in estimating network demand. Athina has a strong background in traffic simulation and optimization. She has in-depth knowledge and experience in traffic management and control, as well as in the analysis of heterogeneous data in transportation using statistical and machine learning methods. Athina joined the Aimsun team in 2019 as Senior Scientific Researcher in Aimsun Research Project team. Athina participates in several European funded research projects as a technical lead and project manager as well as in internal research developments.
Monica Dominguez, Aimsun, monica.dominguez@aimsun.com
Dr. Dominguez has a PhD in Computer Science and more than 10 years of expertise in the area of machine learning and artificial intelligence. She joined the Aimsun Research Projects’ team in 2022 as a scientific researcher and has worked on project management and data science. She participates in European projects leading tasks, carrying out data-driven experiments, data analysis and processing. Her research interests range from data analysis to deep learning with a special focus on graph neural networks applied to traffic tasks.
Antonio Pellicer, Aimsun, antonio.pellicer@aimsun.com
Antonio holds a PhD in civil engineering and has more than 5 years of experience as a researcher, with expertise in the area of transport resilience, drivers’ behaviour and mesoscopic traffic modelling. He has also worked as a transport consultant, developing public transport plans and traffic studies. Antonio joined the Aimsun Research Projects team in 2022 as a scientific researcher and, since then, he has been involved in EU-funded research projects, working on project management and supporting technical tasks. He has participated in a wide variety of projects, covering areas from connected and autonomous vehicles (CAV) and traffic management of incidents to safety analysis and pedestrians’ behaviour modelling.
Mohammadmahdi (Mahdi) Rahimiasl, University of Antwerp and imec, mahdi.rahimiasl@imec.be
Mahdi is a PhD student in Applied Engineering at the IDLab research group of the University of Antwerp and imec. He has been involved in several research projects, primarily focusing on spatiotemporal deep learning models, graph neural networks, and time series in mobility tracks. He is working on innovative solutions for complex and vital problems like traffic forecasting, public transportation delay predictions, and traffic event detection.
Tutorial 2:
Cross-simulator Datasets and Evaluations for Traffic Control Policies
Motivation and objectives:
Traffic control policies are essential for managing traffic flow on roadways and highways. By implementing traffic signal control and dynamic pricing, cities can reduce congestion, increase safety, reduce environmental impact, better use resources, and improve the quality of life for residents. With traffic control policies commonly tested and evaluated using simulations like SUMO or CityFlow since simulations provide a safe and controlled environment. However, existing traffic simulators are limited by their shortage in input data and possible inconsistency between different simulators, which prevents them from generating interactive data from traffic simulation in the scenarios of real road networks. Different simulations are designed with different assumptions or objectives, which can lead to bias and inconsistency in the results. This motivates us to understand the evaluation of traffic control policies using cross simulator data and evaluation tools as proposed by this tutorial.
This tutorial is to provide both lecture and hands on experience for using the toolkits for cross simulator traffic control with enriched datasets. We will review the existing techniques in traffic control literature, introduce the ke y techniques like reinforcement learning, analysis their pros and cons and compare their experimental settings. As most of the literature on traffic control is originally implemented under different simulators or conducted under different datasets, this tu torial provides the usage of two tools, CBLab and LibSignal, to provide a fair comparison between the published literature and potentially act as a reference for future works. This tutorial will also provide the hands on usage on CBLab and LibSignal, to compare different control policies across different simulation environments with different datasets. Our libraries have the following features, which make them a high quality benchmark for cross simulator comparison: (1) Unified: our library builds a systematic pipeline to implement, use and evaluate traffic signal control models in a unified platform. The data configuration, model instantiation, and standardized evaluation procedure are shared across simulators. (2) Comprehensive: we provide over ten models covering two widely used traffic simulators reproduced to form a comprehensive model warehouse and multiple datasets commonly used from different resources.
Both lectures and hands on experiences will be held during the tutorial:
Introduction:
- Data, simulation, and methods in traffic control policies
- Why deep RL in traffic control ?
- SUMO, CityFlow, CBLab and LibSignal
Hands on Simulation with CBLab:
- Introduction to existing datasets and scenarios
- Creating road networks from OpenStreetMap, Custom Templates and existing datasets
- Running experiments on CBLab
Reinforcement learning (RL) and RL for traffic signal control
- Introduction of the principles of RL + high level examples
- Frequently used methods: value-based RL and policy-based RL
- Deep value-based RL (Deep Q Learning), deep policy-based RL (Deep deterministic RL) and other frequently used methods
- Formulating traffic control problem as a reinforcement learning problem
Hands on Implementation for Traffic Signal Control:
- Design of reward, state, action
- Different measures of reward, state (e.g., queue length, waiting time) and action and their pros and cons
- Different models in RL (e.g., Policy based, value based) and their pros and cons
- Step-by-step instructions on setting up the Python environment with LibSignal
Open problems in RL for traffic control
- Simulating environment to real world
- Building a real simulator (e.g., learning to simulate)
- Sim2Real transfer
- Other issues: benchmarks, interpretability, safety issues, etc.
Intended audience:
This tutorial would like to bring together the researchers who are interested in the exciting RL techniques to solve transportation problems. This tutorial is different from previously held tutorials on pure introduction to simulators or reinforcement learning. We plan to have a more focused theme and provide hands-on experiences on the evaluation of traffic control methods with data, simulator and well-implemented toolkits for tasks like traffic signal control and dynamic pricing. We believe this focused theme can enable more in-depth discussions on related topics and techniques.
Materials and equipment needed for the tutorial:
Each presenter will bring their own laptop. The equipment to be needed is a projector, power socket, and laptop adapters. The attendees are not required to bring any equipment.
Tutorial site:
Tutorial website is available here
Speakers:
Hua Wei, Arizona State University, hua.wei@asu.edu
Hua Wei is an assistant professor at the School of Computing and Augmented Intelligence (SCAI) in Arizona State University (ASU). He also affiliates with the Lawrence Berkeley National Laboratory. Before joining ASU, he worked as an Assistant Professor at New Jersey Institute of Technology and a Staff Researcher at Tencent AI Lab. He got his PhD from Pennsylvania State University in 2020 under the supervision of Dr. Zhenhui (Jessie) Li. Before that, he received his master and bachelor degree from Beihang University (BUAA) majoring in Computer Science, working with Prof. Jinpeng Huai and Dr. Tianyu Wo. He has rich experiences in teaching, including BSc and MSc level courses. He has successfully organized workshops on top-tier computer science conferences, including Data-driven Intelligent Workshop Series on SIGKDD/CIKM/ICDM. He has also organized a tutorial “Deep Reinforcement Learning for Traffic Signal Control” on ITSC 2020.
Guanjie Zheng, Shanghai Jiao Tong University, gjzheng@sjtu.edu.cn
Guanjie Zheng is an Assistant Professor at Shanghai Jiao Tong University. He has successfully organized workshops including “The City Brain Workshop” in KDD 2021 and “Prescriptive Analytics for the Physical World” in KDD 2020. He has multiple experiences in giving tutorials in conferences and lectures, including “Learning with Small Data” in KDD 2020 and university seminars for undergraduate and graduate students.