DTR-DDQN: A Reinforcement Learning Framework for Delay-Tolerant Joint Path Selection and Task Offloading

Dynamic route planning is vital for autonomous driving in IoV systems. However, fluctuating RSU computing power, wireless channel gains, and traffic conditions often cause mismatches between the fastest task execution path and the shortest travel time path. To tackle this, we jointly optimize vehicle routing and task offloading to minimize total transit time and reduce risks from traffic and processing delays. Given the exponential action space and dynamic environments, this becomes a complex combinatorial challenge. We propose a dual-input double deep Q-network (DTR-DDQN) to address this, using a DNN to estimate Q-values and avoid action space explosion. By separately capturing traffic and wireless features, DTR-DDQN generates efficient, integrated strategies. Extensive simulations confirm the method’s accuracy and effectiveness.