A robust deep reinforcement learning approach to driverless taxi dispatching under uncertain demand

With the progressive technological advancement of autonomous vehicles, taxi service providers are expected to offer driverless taxi systems that alleviate traffic congestion and pollution. However, it is challenging to maintain the efficiency and reliability of a taxi service system due to the complexity of the traffic network and fluctuating traffic demand. In this paper, we present a robust variant of the twin delayed deep deterministic policy gradient algorithm (TD3), namely, adaptive TD3 integrated with robust optimization (ATD3-RO), to implement a fleet of autonomous vehicles for a taxi service under uncertain passenger demand. Our proposed method incorporates an adaptive module for integer-valued action generation, which also enhances the model's resilience to a larger action space. Considering the uncertain demand of passengers, we design a perturbation sampling-based method to generate adversarial examples for robust training. Additionally, we propose a robust optimization-based strategy to generate a lower bound and guide the convergence of the critic network during the model training process. In our case study, we validate the efficacy of ATD3-RO by constructing a reinforcement learning simulator of the driverless taxi transportation system using real taxi data. The simulation results demonstrate that ATD3-RO outperforms the general TD3 algorithm and other state-of-the-art reinforcement-learning-based approaches while improving learning efficiency. We assess the algorithm's robustness against sudden changes in requests, e.g., a surge in demand at some traffic nodes caused by an emergent event. The results suggest that ATD3-RO performs adaptive actions that are aligned with the variations in passenger demand. Finally, we prove that our model can provide a reliable dispatching strategy even at various ratios between driverless taxis and passenger demand.