• A novel MARL-based end-to-end parking allocation approach with intermodal transport. • GRA is used to model personalised parking profiles for individuals. • A comparative analysis evaluates GRA’s against TOPSIS in the MARL paradigm. Rapid economic growth and technological advancement have fostered increased car ownership around the world. Despite the critical role of vehicles in modern life, parking-related challenges persist, leading to negative externalities such as delays, fuel consumption, and environmental impacts. Although Smart Parking Systems (SPSs) have been developed to address these parking issues, they typically only provide available parking spaces with direct walking access from a car park to a destination, thereby restricting the range of parking options available to drivers. By expanding the parking allocation framework to consider the entire journey from origin to destination rather than solely to a car park, a wider range of available parking options can be explored, which may yield more optimal solutions for reducing negative externalities. In addition, SPSs usually assume uniform parking preferences among drivers, which may not reflect the diverse preferences observed in real-world scenarios. To accommodate varying individual preferences, a personalised parking solution is preferred to optimise parking allocation with a particular focus on alleviating negative externalities. Therefore, this paper develops a personalised end-to-end parking allocation algorithm using Multi-Agent Reinforcement Learning (MARL) to broaden the search for available parking spaces and provide intermodal travel solutions to help drivers reach their destinations from car parks. Real-world data from Nottingham, UK, are used to calibrate the simulation model which is employed to evaluate the learning performance of MARL algorithms, including Deep Q-Network (DQN) and Advantage Actor-Critic (A2C). Additionally, as two commonly-used methods for multi-attribute decision making problems, Grey Relational Analysis (GRA) and Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) are compared in this paper for their effectiveness in modelling personalised parking profiles. The results demonstrate the superiority of the A2C-GRA algorithm, with a significant average total reward improvement of 19% over benchmark models at 95% confidence interval. On average, the travel time and distance optimised by the A2C-GRA algorithm are 39 min and 6.6 km, respectively, representing reductions of 5.22% and 3.62% compared to the benchmark models.
Qiu et al. (Fri,) studied this question.