Dynamic Truck Parking Reservation Using Reinforcement Learning and Mixed-Integer Linear Programming

Abstract: In North Carolina, a severe shortage of truck parking facilities along major interstates such as I-26, I-77, I-85, and I-95 has led to trucks parking illegally along freeway shoulders and interchange ramps, posing safety risks and operational challenges. Despite efforts to increase capacity, significant gaps in truck parking supply remain on critical corridors. Additionally, existing trucker mobile applications face several limitations: they often do not employ learning algorithms to optimize space management over time, lack the ability to reallocate spaces dynamically when cancellations occur, and do not offer robust mechanisms for emergency reallocation. These challenges highlight the need for a more adaptive and intelligent solution. This research proposes a Hybrid RL + MILP model that integrates real-time adaptability with periodic global optimization to address these issues in truck parking management, particularly in North Carolina.

The Reinforcement Learning (RL) component of the model dynamically reallocates parking spaces in real-time, adjusting to system states such as emergency truck arrivals, cancellations, and fluctuating demand. To improve efficiency in real-time decision-making, the RL component draws upon techniques from Integer Gradient Ascent (IGA) and policy-gradient reinforcement learning, facilitating faster convergence to optimal solutions during high-demanding situations. Although the model does not include dynamic pricing, it conceptually aligns with dynamic pricing mechanisms, adjusting space allocation based on fluctuating demand in real time.

In addition to the real-time adaptability provided by RL, the Mixed-Integer Linear Programming (MILP) component performs periodic global optimization, recalculating allocation strategies at regular intervals to optimize facility capacities, balance regional demand, and reserve spaces for emergencies. This dual approach ensures that the system responds promptly to short-term changes while maintaining long-term operational efficiency across multiple facilities. A key innovation of this model is its ability to learn from historical data, enabling the RL agent to anticipate demand surges and proactively adjust parking allocations. Moreover, the integration of multi-facility optimization allows the system to balance demand across multiple locations, such as those in high-traffic corridors, preventing overload at individual facilities.

By combining real-time adaptability with long-term optimization, our hybrid model presents a scalable and adaptive solution for truck parking management in North Carolina. This framework addresses critical gaps in current methodologies, ensuring efficient allocation of parking spaces in response to both routine and emergency situations without compromising overall system performance.

Keywords:
Truck parking management, Reinforcement Learning, Mixed-Integer Linear Programming, Integer Gradient Ascent, policy-gradient learning, hybrid optimization, real-time decision-making, dynamic reallocation, multi-facility integration, emergency parking prioritization, scalable parking systems.