Decentralized Computation Offloading Strategy via Multi-Agent Deep Reinforcement Learning for Multi-Access Edge...

Area of Science:

• Computer Science
• Artificial Intelligence
• Telecommunications

Background:

• Multi-access edge computing (MEC) is crucial for offloading intensive computations from edge devices.
• Decentralized decision-making is needed to manage resource-intensive applications efficiently.
• Simultaneous access attempts in MEC create challenges for optimal offloading.

Purpose of the Study:

• To propose a decentralized offloading decision strategy using multi-agent deep reinforcement learning (MADRL).
• To minimize overall task completion latency for edge devices in MEC environments.
• To enable edge devices to learn offloading policies based on local observations.

Main Methods:

• A decentralized computation offloading strategy based on multi-agent deep reinforcement learning (MADRL).
• Utilizing a deep Q network (DQN) for a discrete action space deep reinforcement learning (DRL) approach.
• Implementing a grant-free access mechanism for decentralized offloading initialization.
• Jointly optimizing user association and offloading decisions to mitigate collisions.

Main Results:

• The proposed MADRL strategy effectively reduces overall task completion latency.
• Faster convergence of learning performance is achieved compared to conventional schemes.
• The decentralized approach demonstrates efficiency and scalability in multi-user MEC environments.

Conclusions:

• The proposed MADRL-based decentralized offloading strategy is efficient for MEC systems.
• The approach successfully minimizes task completion latency and improves learning convergence.
• This method offers a scalable solution for managing computation load in multi-user edge environments.