Distributed Multi-Agent Reinforcement Learning for Collaborative Path Planning and Scheduling in Blockchain-Based Cognitive Internet of Vehicles

The collaborative path planning and scheduling can overcome the limitations of single vehicle intelligence to obtain a globally optimal decision strategy in cognitive internet of vehicles (CIoVs). The collaboration of vehicles necessitates the exchange of environmental and decision information, generating massive collaborative computing tasks with strict latency requirements. Leveraging mobile edge computing (MEC) technology, computing tasks can be processed near the vehicles to reduce latency. However, traffic congestion and computational load imbalance seriously affect traffic efficiency and computational latency. In hybrid driving scenarios, it is challenging to fulfill the diverse service requirements of vehicles with different intelligence levels. Moreover, non-collaborative tend to result in traffic congestion due to vehicle aggregation effects, while centralized solutions lack flexibility and have high computational complexity. To address these concerns, a distributed multi-agent reinforcement learning (DMARL) algorithm is proposed for collaborative path planning and scheduling in a blockchain-based collaboration framework. In this framework, we model the communication, traffic situation and task processing of the system and formulate a joint optimization problem to minimize both travel time and computation latency. Last, we convert the scheduling problem for different types of vehicles into Markov decision processes (MDPs) and propose Q-learning-based DMARL algorithm to achieve proactive load balancing of both road infrastructures and MEC nodes (MECNs). Simulation results demonstrate that the proposed approach outperforms the comparison schemes in terms of load balance indexes of roads and MECNs, travel time, and computation latency.