Deep Reinforcement Learning for Computation Offloading and Resource Allocation in Unmanned-Aerial-Vehicle Assisted...

Area of Science:

• Computer Science
• Electrical Engineering
• Artificial Intelligence

Background:

• Mobile edge computing (MEC) extends cloud capabilities to network edges, benefiting resource-limited devices.
• Existing MEC offloading methods struggle with dense user and sparse infrastructure scenarios.
• Unmanned Aerial Vehicles (UAVs) offer a flexible solution for MEC in challenging environments.

Purpose of the Study:

• To address the challenge of offloading decisions and resource allocation in multi-user, multi-server UAV-assisted MEC systems.
• To optimize the weighted total cost of delay, energy consumption, and discarded tasks for end-users.
• To develop an efficient computation offloading policy for UAV-assisted MEC.

Main Methods:

• Formulated the joint optimization problem as a Markov decision process.
• Applied the soft actor-critic (SAC) deep reinforcement learning algorithm.
• Developed a SAC-based dynamic computing offloading (SACDCO) algorithm.

Main Results:

• The SACDCO algorithm effectively reduces delay, energy consumption, and discarded tasks.
• Achieved approximately 50% reduction in system delay compared to fixed local-UAV schemes.
• Achieved approximately 200% reduction in energy consumption compared to fixed local-UAV schemes.

Conclusions:

• The proposed SACDCO algorithm provides an effective solution for computation offloading in UAV-assisted MEC.
• Demonstrates significant improvements in Quality of Service (QoS) for end-users.
• Highlights the potential of UAV-assisted MEC for future mobile computing.