Adaptive multi-agent optimization in intelligent-connected vehicles based on deep reinforcement learning

In order to provide offloading services for computationally demanding and time-sensitive operations on vehicles, multi-access edge computing has emerged as a successful paradigm. However, there are typically spatiotemporal load mismatches across edge servers due to significant vehicle mobility. As a result, task migration is used to transfer excessive tasks from overloaded to underloaded servers in order to maintain dynamic workload balance. Deep reinforcement learning techniques are used in recent research to produce offloading and migration decisions based on system observations. On the other hand, task migration in the incorrect direction may result in further delays as the migration direction is said to be heavily reliant on vehicle movements. The challenges intelligent connected vehicles encounter during service migration in a highly dynamic Internet of vehicles environment, such as multi-user resource competition and dynamic changes in edge node availability, were addressed by a service migration technique based on multi-agent group relative policy optimization (MAGRPO). The formulation of the service migration problem as a long-term multi-user joint optimization problem with resource constraints led to the development of a MAGRPO approach that did not require an explicit critic network. A policy update signal based on the relative ordering of discounted returns within the group was developed to lower training expenses and effectively alleviate training instability caused by severe penalties (such as node overload or failure). Simulation findings show that the proposed method outperforms existing baseline methods in key metrics such as migration success rate, overall service delay, and migration energy usage. In situations where edge node resources are few and their availability fluctuates, it demonstrates greater resilience and scalability.