Safe Reinforcement Learning for Intersection Management in RITI Communities Under Rare Extreme Events

Abstract

The rapid advancement of artificial intelligence (AI) is reshaping the transportation sector, with applications spanning autonomous vehicles, driver injury prevention, and traffic management. Efficient traffic management, particularly through adaptive intersection control, holds significant potential for reducing congestion. This study explores the application of reinforcement learning (RL) to adaptive traffic signal control in rural, isolated, tribal, and indigenous (RITI) communities, which face unique challenges such as rare extreme weather events. Standard RL approaches struggle in these contexts due to limited exposure to these rare events. In our study, we first evaluate several mainstream RL algorithms and identified two most promising approaches. Then, we propose to use offline RL algorithms, which can train on existing datasets before interacting with the real environments. This provides a robust solution because (1) it is costly to deploy the algorithm and let the traffic network operate under suboptimal policies before the algorithm learns the optimal policy, and (2) it mimics the scenario where some events are not seen in the training dataset. We compare the performance of offline RL algorithms using different offline datasets, generated by policies of different levels of expertise, in realistic test cases. Results indicate that offline RL approaches perform better when trained on datasets from expert policies, stressing the importance of the quality of the offline datasets. These findings highlight the potential of RL-based adaptive traffic control for improving transportation efficiency, especially when tailored to the specific conditions of RITI communities.