Computation Offloading Strategy Based on Multi-Agent Reinforcement Learning in Vehicular Edge Computing Networks.

Area of Science:

• Intelligent Transportation Systems
• Vehicular Edge Computing
• Multi-Agent Reinforcement Learning

Background:

• Vehicular applications require intensive computation and low latency.
• Existing offloading strategies fail in dynamic vehicular networks due to mobility and communication risks.
• Need for advanced solutions to handle complex vehicular network environments.

Purpose of the Study:

• To design a novel computation offloading algorithm for vehicular edge networks.
• To address challenges of dynamic environments, high mobility, and concurrent tasks.
• To minimize latency and energy consumption in intelligent transportation systems.

Main Methods:

• Developed a computation offloading algorithm using multi-agent reinforcement learning (MARL).
• Incorporated queue load, communication links, task attributes, and computing resources.
• Formalized the problem as a multi-agent Markov decision process (MAMDP).
• Designed an improved Multi-Agent Proximal Policy Optimization (MAPPO)-based MATPPO-T algorithm.

Main Results:

• The proposed MATPPO-T algorithm achieved joint optimization of task offloading, resource allocation, and result migration.
• Reduced total system cost by approximately 22% compared to benchmark algorithms (MAPPO, PPO).
• Demonstrated the lowest offloading overhead and fastest convergence speed.

Conclusions:

• The developed MARL-based algorithm is robust and scalable for dynamic vehicular edge networks.
• Effectively manages complex vehicular environments with high-speed mobility.
• Offers significant improvements in latency and energy efficiency for intelligent transportation systems.