Federated Deep Reinforcement Learning Based Task Offloading with Power Control in Vehicular Edge Computing.

Area of Science:

• Computer Science
• Electrical Engineering
• Network Engineering

Background:

• Vehicular edge computing (VEC) supports low-latency applications by offloading tasks to edge servers.
• High vehicle density and mobility in VEC environments cause channel interference, increasing power consumption and latency.
• Existing multi-agent DRL methods suffer from limited local information, while single-agent DRL requires excessive data exchange.

Purpose of the Study:

• To propose a novel task offloading method with power control for VEC systems.
• To maximize VEC system throughput under vehicle power constraints.
• To address the limitations of existing DRL approaches in dynamic vehicular networks.

Main Methods:

• A federated deep reinforcement learning (FL) method is proposed, combining centralized and distributed approaches within the deep deterministic policy gradient (DDPG) framework.
• The method considers dynamic channel interference and conditions for optimized task offloading and power control.
• Deep reinforcement learning (DRL) is leveraged for handling complex environments and high-dimensional inputs.

Main Results:

• The proposed FL-based DDPG method effectively manages task offloading and power control in dynamic VEC environments.
• Demonstrated significant improvements in system throughput compared to conventional methods.
• Showcased a reduction in queueing delay for vehicles in dynamic vehicular networks.

Conclusions:

• The federated deep reinforcement learning approach offers a superior solution for task offloading in VEC systems.
• The method effectively balances performance and resource constraints in dynamic vehicular networks.
• This research contributes to enhancing the quality of service (QoS) in VEC applications.