Offline Datasets for "DataLight: Data-Centric Offline Reinforcement Learning for Traffic Signal Control"

This record contains the offline datasets used in the ECML PKDD 2026 paper “DataLight: Data-Centric Offline Reinforcement Learning for Traffic Signal Control”. The datasets were constructed for offline reinforcement learning in traffic signal control and include trajectories generated under different traffic signal control policies, such as expert, moderate, and random policies. Each dataset records offline interaction samples used for training and evaluating DataLight and baseline offline RL methods, including state representations, actions, rewards, next states, and related experimental metadata. These datasets support the reproducibility of the experiments reported in the paper, including offline policy learning, data-quality analysis, robustness evaluation under different offline data sources, and comparisons with representative offline and online reinforcement learning baselines for traffic signal control.