Champion! SEUers won in 2020 “Big Data World Cup”
Editor’s note
Do you still remember the team TLab guided by Prof. Liu Zhiyuan from School of Transportation won two awards in KDD CUP also known as the “Big Data World Cup”? This year, this team participated in KDD CUP again, stood out among 1,000 participating teams from all over the world and finally won the champion.
The International Conference on Knowledge Discovery and Data Mining (ACM SIGKDD, KDD for short) recently announced the results of its annual event KDD CUP 2020 on its official website. The team TLab guided by Prof. Liu Zhiyuan from School of Transportation of Southeast University (members including Dr. Liu Yang, graduate student Lyu Cheng, etc.) stood out among more than 1,000 participating teams from all over the world and finally won the champion upon fierce competition in the Reinforcement Learning Competition Track: Vehicle Repositioning Task. The LAMDA Group from Nanjing University and the NTTdocomo team from Japan ranked second and third respectively.
01
About the event
The International Conference on Knowledge Discovery and Data Mining (ACM SIGKDD, KDD for short) is sponsored by the Special Interest Group on Knowledge Discovery and Data Mining (SIGKDD) of U.S. Association for Computing Machinery (ACM), and is recommended by the Chinese Computer Federation (CCF) as a Class-A international academic conference. It has been held for 26 sessions up to now. Among them, KDD CUP, which is also known as the “Big Data World Cup”, is an annual event organized by ACM SIGKDD and currently the international top event at the highest level with the most influences and on the largest scale in data mining.
02
Track setting
2020 KDD CUP sets 3 competition tracks, of which the intensive learning competition track titled “Learning to Dispatch and Reposition on a Mobility-on-Demand Platform” is undertaken by Didi.
The participants had to solve the problem of order matching and vehicle scheduling on the online car-hailing platform, in which the order matching task needs to be matched every two seconds, and the vehicle scheduling task needs to be dispatched within over 8,000 hexagonal grids in Chengdu City. Related algorithm shall match the potential travel needs of passengers with suitable drivers so as to use vacant vehicles more efficiently, increase the vehicle turnover, improve the user’s experience and the driver’s income, and optimize the system operating efficiency.
The existing multi-agent reinforcement learning method regards each vehicle as an agent. During simulation, the vehicle goes to the destination according to the scheduling algorithm or performs a biased random driving without considering the driver’s preference. This may bring the following problems. When multiple vehicles share the same state (concurrent situation), these vehicles will make the same decision, in this case, multiple agents are “redundant” and function as the “duplicate” of one vehicle.
Considering the said shortcomings, the TLab team guided by Prof. Liu Zhiyuan optimized the existing method in a systematic and comprehensive manner, and further designed a more logical single-agent deep reinforcement learning method.
For example, this method considers a single agent as an intelligent “dispatch center”. When the vehicle has sent a dispatch request, the “dispatch center” will perform a global capacity dispatch. In combination with the previous experience in large-scale spatio-temporal forecasting, the team further screened the research area and customized an N×N grid to combine the global information (the temporal and spatial distribution of orders, vehicles and value throughout the city) and the local information (the vehicle’s current location, etc.) as a state. A pruned global action space was adopted to prevent the vehicle from falling into a local optimum.
Prof. Liu Zhiyuan said: “The data provided by Didi roughly included 210,000 orders per day, 30,000 taxi drivers and dynamic data subject to changes in 24 hours. A total of over one-month data was provided to us for model training. Afterwards we submitted the model to the organizer who would evaluate our algorithm according to its simulation environment. Finally, upon multiple rounds of rigorous evaluation in the preliminary and semi-finals, they believe our method is superior to others teams.”
Upon fierce competition for several months, SEU TLab team took a significant lead in the Dev Reposition Score and Total Reposition Score, and finally won the champion with excellent results among more than 1,000 participating teams.
03
Team members
Tutor: Prof. Liu Zhiyuan
Prof. Liu Zhiyuan, a doctoral supervisor, graduated from National University of Singapore with a Ph.D. and once taught at Monash University in Australia, is currently serving as Deputy Dean of School of Transportation, Southeast University, a member of the Academic Committee and the Visiting Professor of Monash University, Australia. His main research includes transportation network planning and management, transportation big data analysis and modeling, public transportation, and intelligent transportation systems, etc. He has published more than 100 SCI papers and served as Associate Editor of ASCE Journal of Transportation Engineering and IET Intelligent Transport Systems, the well-known SCI journals in the field of transportation research.
Team member: Liu Yang
Dr. Yang Liu, enrolled in 2017, became a member of Prof. Liu Zhiyuan’s research team from the School of Transportation of Southeast University. His research direction concerns machine learning algorithms and their applications in traffic engineering. He has published many papers in international journals, including IEEE Transactions on Cybernetics, Transportation Research Part C, IEEE Transactions on Intelligent Transportation Systems, etc., and won more than ten champions and runners-up in several competitions regarding artificial intelligence algorithm, including three Champions in Alibaba Tianchi Algorithm Competition (2016, 2018 and 2019), Champion in IEEE CS Ucar Artificial Intelligence Competition (2018), Champion in IJCAI Artificial Intelligence Confrontation Challenge (2019), Recommended Runner-up in KDD CUP Multi-Mode Transportation (2019), etc.
Team member: Lyu Cheng
Lyu Cheng, a graduate student, graduated with a bachelor’s degree in traffic engineering from SEU in 2018 and joined in Prof. Liu Zhiyuan’s team to study for the master’s degree. His research direction concerns traffic big data analysis and modeling. He has published 5 papers in such SCI journals as IEEE Transactions on Intelligent Transportation System, Computer-Aided Civil and Infrastructure Engineering, etc., and has also won a number of awards in competitions regarding artificial intelligence, including the Gold Award in 2019 National University New Energy Vehicle Big Data Innovation and Entrepreneurship, and the First Prize in 2019 National University Big Data Application Innovation Competition, the Third Prize in ICME 2019 Grand Challenge on Short Video Understanding Challenge, etc.
Considering the business pain points in Internet enterprises and the transportation industry, the transportation engineering discipline of the School of Transportation, Southeast University has been working closely with leading companies in various fields such as Huawei, Didi, China Mobile, and China Telecom, etc. to develop technological achievements in artificial intelligence and big data, apply them to the data analysis platform, and thus provide technical support for the construction of smart cities.
Behind the champion,
it is the hard work and perseverance of the TLab team.
Let’s cheers for them!
Written by Liu Yang, Guo Qin
Photos by Cong Cong, Zhao Mengzhe, Wang Chenxiao
Revised by Shu Yuan
Proofread by Chen Si, Eric Song, Melody Zhang
Edited by Sun Shukai