Exploring the Feasibility and Sensitivity of Deep Reinforcement Learning Controlled Traffic Signals in Bidirectional Two-Lane Road Work Zones

Reconstruction of bidirectional two-lane roads typically necessitates closing one lane, with traffic signal control (TSC) implemented to prevent direct conflicts between opposing traffic flows. First, the effectiveness of 16 modified discrete and continuous deep reinforcement learning (DRL) algorithms from 4 categories (DQN, AC, PO, DDPG) was systematically investigated in virtual and real-world bidirectional two-lane road work zone scenarios by this study to explore inter-correlations between scenarios and DRL algorithms. Next, multiple sensitivity analyses, including the work zone length, average speed, and directional distribution factor, are conducted to explore the core mechanisms underlying the performance discrepancies of different algorithms. Results indicates that the discrete D3QN-PER-2s (Dueling Double Deep Q network that uses both prioritized experiences replay and double-step temporal difference methods) outperforms continuous DRL and baseline methods in both virtual and real-world scenarios. Overall, D3QN-PER-2s decreases about 68% of the average waiting time and 15 % of the CO₂ emission in three real-world scenarios, indicating the strong transferability of DRL methods. Additionally, with longer work zone distance, the better performances are observed in discrete methods. Last, the fundamental mechanisms underlying the performance divergences are explored. Furthermore, several scenario-specific traffic management approaches are suggested for implementation. These findings offer valuable guidance and insights into the development and implementation of DRL TSC in bidirectional two-lane road work zones.