Posted on Leave a comment

Energy Harvesting for Dynamic Computation Offloading using GRL


J. Chen and Z. Wu, “Dynamic Computation Offloading with Energy Harvesting Devices: A Graph-Based Deep Reinforcement Learning Approach,” in IEEE Communications Letters, vol. 25, no. 9, pp. 2968-2972, Sept. 2021,

doi: 10.1109/LCOMM.2021.3094842.


The author presents a unique technique called GCN-DDPG for joint partial offloading and resource allocation (JPORA) in mobile edge computing (MEC) systems with energy harvesting (EH). The dynamic nature of the MEC environment, with its changing numbers of mobile devices (MDs), compute jobs, and accumulated amounts of energy, is the difficult to be handled. Because traditional DDPG is not very good at extracting latent representations from Euclidean data, it does not generalize well in such dynamic changing conditions. In order to overcome this restriction, author propose GCN-DDPG and it Improving JPORA choices for MDS including uplink transmission power, local computing capacity, and offloading ratio is the goal of this paper.

The issues author spots in other paper methods for JPORA in MEC for Energy Harvesting:

  • Some paper suggests convex relaxation and heuristic search as optimization-methodologies though they are in successful some situation and they take a long time to spot workable solutions, so this may not be appropriate for the dynamic nature of MEC systems in large scale situations.
  • Other researchers propose that algorithms utilize Deep Q-Networks (DQN), which have trouble with dynamic JPORA issues’ continuous action spaces and determining the right degree of discretization is difficult since changing it too much might result in higher-dimensional complexity or the loss of behavioral information.
  • Only latent representations may be extracted from Euclidean data by DDPG agents using neural networks such as FCN, CNN, and LSTM. But these networks can’t be able to adequately replicate the graph-like features that task offload distribution and allocation of resources display in MEC systems.
  • Neural network exists are not able to recognize graph-based relations among network components, which is required to properly handle JPORA issues.

Therefore, These issues point to the need for a more flexible and successful method of JPORA in MEC with EH devices. So, the author proposes a GCN-DDPG agent.

problem statement:

The objective of this study is to reduce, over time, the average weighted cost of job computation time and energy consumption on MDs Therefore the author proposes a problem formulation as:


lim(T → ∞) (1/T) * Σ(t=1 to T) (1/M) * Σ(m=1 to M) (ω1 * τt^m + ω2 * Et^m)

  •  Represents the time horizon, indicating the duration over which the average tradeoff is considered.
  • 𝑀: Represents the number of devices.
  • 𝜔1 and 𝜔2: Tradeoff weights, determining the relative importance of computation time (𝜏𝑡𝑚) and energy consumption (𝐸𝑡𝑚).
  • 𝜏𝑡𝑚: Computation time of device 𝑚 at time 𝑡.
  • 𝐸𝑡𝑚: Energy consumption of device 𝑚 at time 𝑡.



The fig of proposed model illustrated by GCN-DDPG frame from the study by J. Chen (2021)

Methodology described in this work offers a unique solution to Energy Harvesting devices, mobile edge computing (MEC) systems face the dynamic JPORA problem in GCN-DDPG GCN collects the graphical structure whereas the DDPG is used in decision making. The initialization of the MEC system with EH MDs includes setting up parameters such as the quantity of MDs, calculation jobs, max battery capacity, and obtained energy levels. Task data, uplink channel gains, battery energy levels, and harvested energy levels for each MD are represented as the major state of the MEC system. The GCN-DDPG agent determines continuous actions, such as offloading ratios, local computing capacity, and uplink transmission powers, for each MD on the basis of observed condition. The effectiveness and durability of the proposed technique in resolving the JPORA issue in MEC with EH devices have been validated   thorough performance assessments and analyses of the experimental data.

Results obtained during validation:


                                                 Results obtained in the study by J. Chen (2021)


  • Used GCN-DDPG to address dynamic decision-making in MEC with EH MDs.
  • Furthermore, it Reduced the average weighted cost of energy use and job completion time.
  • GCN makes it easier to study the MEC network topology and make effective decisions.
  • Based on experimental data, additionally GCN-DDPG works better than the most advanced techniques.
  • Future scope: Investigate how to allocate computing resources at edge and servers for additional optimization.


Sakthivel R

I am a First-year M.Sc., AIML student enrolled at SASTRA University

I possess a high level of proficiency in a variety of programming languages and frameworks including Python. I have experience with cloud and database technologies, including SQL, Excel, Pandas, Scikit, TensorFlow, Git, and Power BI

Leave a Reply

Your email address will not be published. Required fields are marked *