Demo: A Probabilistic Model-Based Deep Reinforcement Learning Strategy to Maximize Safety Under Cyberattacks on a Connected Intersection-Bridging Stochasticity and Real-World Driving Data

With the increasing risks posed by malicious cyber-attacks on connected vehicles and transportation infrastructures, we propose a probabilistic model-based deep reinforcement learning strategy to maximize safety under cyberattacks by theoretically bridging the stochasticity and the real-world driving data. The probability of collision is calculated using a probabilistic model framework of collision under cyberattack and was calibrated based on the vehicle trajectory data obtained from driving simulation experiments by applying the trajectory reconstruction model and Monte-Carlo simulation. While applied in the deep reinforcement learning algorithm, the probability of collision is set as a reward to optimize vehicles and traffic signal control. The model is calibrated using real-world driving data and numerical simulation, and the optimization method is validated through a case study. It is demonstrated that the proposed framework is promising to maximize safety under cyberattacks without increasing the algorithm's interference time, which means the reinforcement learning algorithm has the advantage to achieve real-time control as compared to other existing optimization strategies such as dynamic programming.