A DRL based cooperative approach for parking space allocation in an automated valet parking system

Automated valet parking (AVP) is one of the most advanced technologies for improving parking efficiency and security. However, in an AVP system, the traditional vehicle-side greedy search strategy for available parking spaces is likely to achieve low global efficiency and poses a high risk of collision. Therefore, in this study, a system-side deep reinforcement learning (DRL)-based cooperative approach is proposed to solve the parking space allocation problem in a large AVP environment. First, the problem of parking space allocation is formulated as a Markov decision process (MDP). Then, a reward shaping method oriented to the global objective is designed. Next, because the current reinforcement learning methods are difficult to apply to parking space allocation involving large numbers of discrete actions, a cost-based method of parking allocation action embedding is proposed to embed the discrete parking actions in a continuous space, which the actor can generalize. After action embedding, the deep deterministic policy gradient (DDPG) is employed as the training algorithm. The experimental results show that the proposed DRL -based cooperative approach can converge in the parking space allocation problem involving a large AVP system and achieve greater improvement of global AVP efficiency than can the other parking methods.