A Federated Learning and Deep Reinforcement Learning-Based Method with Two Types of Agents for Computation Offload

Area of Science:

• Computer Science
• Artificial Intelligence
• Mobile Computing

Background:

• Mobile edge computing (MEC) faces challenges with latency-sensitive applications due to network variability and task uncertainty.
• Existing computation offloading strategies suffer from decision delays and user data privacy concerns.

Purpose of the Study:

• To develop an efficient computation offloading strategy for MEC environments that minimizes system latency.
• To address decision-making difficulties caused by task uncertainty and dynamic network conditions.
• To enhance user data privacy and reduce communication overhead in MEC systems.

Main Methods:

• Proposed FDRT, a federated learning and deep reinforcement learning (DRL) method utilizing a multi-agent collaborative offloading strategy (DRT).
• Implemented two DRL agents (DDQN and D3QN) for joint decision-making on local computation vs. MEC offloading.
• Integrated federated learning with a novel parameter aggregation method to optimize model training and ensure data privacy.

Main Results:

• The DRT strategy reduced average task execution delay by up to 50% compared to baseline and state-of-the-art methods.
• FDRT accelerated the convergence rate of multi-agent training.
• Training time for DRT was reduced by 61.7% through federated learning.

Conclusions:

• FDRT offers an effective solution for optimizing computation offloading in MEC, balancing latency reduction, privacy, and efficiency.
• The proposed federated learning approach enhances DRL model training in collaborative MEC environments.
• This work provides a significant advancement in intelligent resource management for mobile edge computing.