Research
Pioneering data-driven algorithms to make actionable actions in the real world.
Open Source / Software
Our group values practical and reproducible research. Below are open-source libraries, benchmarks, and tools built by the DaRL group and our collaborators, spanning reinforcement learning for transportation, large-model agents, and spatio-temporal data mining. For the full list of repositories, see our GitHub organization
CityFlow
A multi-agent reinforcement-learning environment for large-scale city traffic scenarios, orders of magnitude faster than SUMO for training RL-based signal control.
Honor of Kings Arena
A competitive multi-agent reinforcement-learning environment built on Tencent’s Honor of Kings MOBA game, designed to benchmark generalization across heroes, lineups, and opponents.
Instructional Agents
A multi-agent LLM system that reduces teaching-faculty workload by automating instructional design โ lecture plans, slides, quizzes, and course material generation.
IntelliLight
A reinforcement-learning approach for intelligent traffic-light control โ one of the earliest deep-RL methods for adaptive signal control on real-world road networks.
RL_Signals
A curated hub of papers, datasets, simulators, and tutorials covering reinforcement learning for traffic signal control โ the go-to reading list for newcomers to the area.
LibSignal
A unified open library for traffic signal control with reproducible RL baselines across simulators (CityFlow, SUMO), standardized benchmarks, and cross-simulator evaluation.
PyDimension
Dimensionless learning โ data-driven discovery of dimensionless numbers and scaling laws from scarce experimental measurements, combining physics priors with sparse regression.
PromptGAT
Prompt-to-Transfer: closing the sim-to-real gap for traffic signal control by conditioning a grounded action transformer on natural-language prompts.
CoMAL
Collaborative Multi-Agent Large Language Models for mixed-autonomy traffic โ coordinating CAVs and human-driven vehicles via LLM-based negotiation and planning.
CityFlowER
An efficient and realistic traffic simulator with embedded machine-learning vehicle-behavior models, bridging the gap between rule-based speed and data-driven realism.
Open-TI
Open Traffic Intelligence โ an augmented-language-model agent that turns natural-language instructions into traffic analysis, simulator control, and signal-policy actions end-to-end.
Sponsors
Learning to Simulate
Papers: ECML-PKDD'24a, PADS'23, ERA'23, KDD'22, AAAI'21, ICDE'21, ECML-PKDD'20, AAAI'20 Workshop
Realistic simulators are a step closer towards policymaking for the real world. We investigate how to build realistic simulators from real-world data.
Simulator/Environment Building/Datasets
Project websites: CityFlowER, Honor of Kings (็่ ่ฃ่), LibSignal, CityFlow, Epidemic, Product Allocator
Simulators are the foundation of reinforcement learning. We built a bunch of simulators for various applications, including MOBA Games, transportation, epidemics, and product allocation.
Trustworthy Deep Learning
Papers: EACL'25b, EACL'25c, SIGKDD Explorations'25, COLM'25, KDD'25a, KDD'25c, ICML'25a, ICML'25b, AAAI'24a, AAAI'24c, ICDM'23, CDC'23a, CIKM'23, KDD'23, IJCAI'23, ERA'23, AAAI'23, IAAI'22, IJCAI'21a, IJCAI'21b, USENIX Security'21 (Adversarial Policies), NeurIPS'20 Workshop
The project investigates different aspects of trustworthy deep learning, including robust modeling for deep learning models with physics, reinforcement learning with offline data, and adversarial policy training.
Deep Reinforcement Learning
Papers: Survey (Arxiv), Survey(KDD Explorations), AAAI'24a, CDC'23a, CDC'23b, CASE'23, IJCAI'23, AAAI'23, AAAI'20, KDD'19, CIKM'19a, CIKM'19b, KDD'18
The project systematically investigates "smart" traffic light control systems using deep reinforcement learning and evaluate its effectiveness on both synthetic and real-world traffic data.
