KR20230008321A - Method and system for optimal partial offloading in wireless edge computing network - Google Patents

Abstract

The present invention relates to a method for optimal partial offloading in a wireless edge computing-based network comprising: a second base station located within the coverage of a first base station; and a server connected to the first base station. The method comprises the steps of: calculating offloading data for offloading a task created from at least one user terminal connected to the second base station; calculating the waiting time and energy consumption required to offload the offloading data; and processing the offloading data based on the calculated waiting time and energy consumption. Therefore, the method can schedule task offloading and resource allocation problems between wireless backhaul and wireless fronthaul in a mobile edge computing-based network.

Description

Optimal partial offloading method and system for offloading in wireless edge computing based network

The present invention relates to an optimal partial offloading method and a system for offloading in a wireless edge computing-based network, and more particularly, to a wireless edge computing-based network for minimizing overhead calculated in terms of latency and energy consumption. to an optimal partial offloading method and a system for offloading.

The explosion of mobile data is expected to impose critical requirements such as high reliability and low latency to achieve optimal performance and introduce more challenges to various aspects of wireless networks in the future.

The significant increase in intelligent mobile devices (e.g., smart watches, smartphones, virtual reality glasses, etc.) has resulted in the emergence and growing popularity of various computationally intensive applications such as interactive online gaming, natural language processing, augmented reality and facial recognition.

However, mobile devices have limited computational resources such as battery energy, memory size and processing speed, which limits their ability to achieve acceptable quality of experience (QoE) and quality of service (QoS).

Therefore, a wireless network (5G or higher) is required to support both timely wireless access and computational offloading provisioning for user terminals, and small cell communication is widely used as one of the technologies of 5G networks to meet these wireless access requirements. It is known.

Related prior art is Republic of Korea Patent Publication No. 10-2020-0017589 (title of invention: cloud server for offloading tasks of mobile nodes in a wireless communication system and its operation method, publication date: February 19, 2020) ) is there.

Accordingly, an embodiment of the present invention provides an optimal partial offloading method and method for offloading in a wireless edge computing-based network for scheduling task offloading and resource allocation problems between wireless backhaul and wireless fronthaul in a mobile edge computing-based network. The purpose is to provide a system.

The problem to be solved by the present invention is not limited to the above-mentioned problem (s), and another problem (s) not mentioned above will be clearly understood by those skilled in the art from the description below.

Partial off in a wireless edge computing-based network including at least one second base station located within the coverage of the first base station and a server connected to the first base station according to an embodiment of the present invention for achieving the above object. A loading method comprising: calculating offloading data for offloading a task generated from at least one user terminal connected to the second base station; calculating waiting time and energy consumption required for offloading the offloading data; and processing the offloading data based on the calculated waiting time and energy consumption.

In one embodiment of the present invention, the first base station and the second base station interwork through a wireless backhaul, the second base station and the at least one user terminal interwork through a wireless fronthaul, and the wireless backhaul and the Wireless fronthaul can share frequency bandwidth.

In one embodiment of the present invention, the calculating step is based on a variable related to an offloading rate set based on a channel state between the first base station and the second base station and a channel state between the second base station and the user terminal Offloading data for offloading a task may be calculated.

In one embodiment of the present invention, the calculating may include, when the variable related to the offloading speed has a value between 0 and 1, the first data offloaded as part of the task and the remainder except for part of the task. Second data processed by the user terminal may be calculated.

In one embodiment of the present invention, the calculating step calculates the waiting time based on the following Equation 1 showing the waiting time required to process the first data and the second data, and the first data and The amount of energy consumption may be calculated based on Equation 2 representing the amount of energy required to process the second data.

[Equation 1]

[Equation 2]

Here, y ^t _ij is the total waiting time required to offload the offloading data, y ^t _ij,l is the waiting time required to process the second data, and y ^t _ij,r is the first data , ε ^t _ij is the total energy consumption required to offload the offloading data, e ^t _ij,l is the energy consumption required to process the second data, e ^t _{ij ,r} means energy consumption required to process the first data.

An optimal partial offloading method in a wireless edge computing-based network according to an embodiment of the present invention further includes optimizing offloading by applying the calculated latency and energy consumption to a pre-learned learning algorithm, wherein the As the learning algorithm, reinforcement learning may be applied to a Markov Decision Process composed of a state space, an action space, and a utility function of an offloading environment.

In one embodiment of the present invention, the state space is a channel state between a first base station and a second base station, a channel state between the second base station and the user terminal, and a state between cells covered by the first base station and the second base station. interference, a set of buffer states in which offloaded tasks in the server are stored, the action space includes a set of variables related to offloading speed and a variable related to offloading bandwidth allocation, and the utility function has the following mathematical expression It can be expressed based on Equation 3.

[Equation 3]

Here, X ^t is the state space, a ^t is the action space, ε ^t _ij is the total energy consumption required to offload the offloading data, and y ^t _ij is the offloading data The total waiting time required, λ is a weight variable related to task drop, Г ^t is a weight variable related to the waiting time and energy consumption of the user terminal, r ^t _jm is the first base station and the second base station Refers to the data transfer rate of the wireless backhaul.

In an embodiment of the present invention, the reinforcement learning may use deep deterministic policy gradient (DDPG) based on actors and critics.

A system for optimal partial offloading in a wireless edge computing-based network according to an embodiment of the present invention includes a first base station, a second base station located within the coverage of the first base station, and at least one connected to the second base station. It may include a user terminal and a server for processing tasks offloaded from the user terminal.

According to an embodiment of the present invention, task offloading and resource allocation problems between wireless backhaul and wireless fronthaul may be scheduled in a mobile edge computing-based network.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

1 is a conceptual diagram illustrating a system for partial offloading in a wireless edge computing-based network according to an embodiment according to an embodiment of the present invention.
2 is a flowchart illustrating a partial offloading method in a wireless edge computing-based network according to an embodiment of the present invention.
3 is a diagram illustrating a framework of DDPG learning according to an embodiment of the present invention.
4 is a diagram illustrating execution codes for processing partial offloading in which DDPG learning is reflected in an embodiment of the present invention.
5A is a graph showing overhead due to waiting time compared to comparative examples according to an embodiment of the present invention.
5B is a graph showing overhead due to energy consumption compared to comparative examples according to an embodiment of the present invention.
6 is a graph showing average task drop rates of user terminals compared to comparative examples according to an embodiment of the present invention.
7 is a graph showing total system overhead versus the number of user terminals compared to comparative examples according to an embodiment of the present invention.
8 is a graph showing total system overhead versus task size compared to comparative examples according to an embodiment of the present invention.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. Like reference numerals have been used for like elements throughout the description of each figure.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no other element exists in the middle.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in the present application, they should not be interpreted in an ideal or excessively formal meaning. don't

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a conceptual diagram illustrating a system for partial offloading in a wireless edge computing-based network according to an embodiment according to an embodiment of the present invention.

Prior to description, the network constituting the system of the present invention is a 5th generation (5G) wireless edge computing (MEC) network, which provides a cloud computing function and an IT service environment at the edge of the network.

In such a 5G wireless edge computing network, it is necessary to optimize task allocation to quickly process tasks required by multiple user terminals in real time, and for this, task offloading techniques that minimize standby time and energy consumption are becoming important.

Therefore, the present invention proposes a partially performed offloading technique for optimized task allocation.

Referring to FIG. 1 , a system for partial offloading in a wireless edge computing-based network according to an embodiment of the present invention includes a first base station, a second base station, a user terminal, and a server.

A first base station may be an electronic device that provides access to a network for wireless terminals in 5G cellular communication. The first base station may manage a coverage cell of a predetermined size and may overlap with at least one small cell that is a coverage cell of a smaller size. For example, the first base station may be a central base station gNodeB (gNB).

The second base station may manage a small cell as a base station located within the coverage of the first base station. For example, the second base station may be a small cell base station (sbs).

The user terminal may be connected to the second base station. At this time, one or more user terminals may be connected to the second base station, and the user terminals may perform wireless communication through a coverage band of a small cell managed by the connected second base station. For example, a user terminal may be a UE.

In one embodiment, the first base station and the second base station may interwork through a wireless backhaul link, and the second base station and at least one user terminal may interwork through a wireless fronthaul link.

In this case, the wireless backhaul link and the wireless fronthaul link may share a frequency bandwidth. In this embodiment, a bandwidth sharing coefficient k ^t _jm (where 0 <= k ^t _jm <= 1) may be applied. In detail, k ^t _jm is the total number allocated to the wireless backhaul link between the first base station and the second base station Bandwidth, and (1-k ^t _jm ) may mean the total bandwidth allocated to the wireless fronthaul link between the second base station and the user terminal.

The server is a central server connected to the first base station and may be capable of communicating with the first base station, a second base station capable of interworking with the first base station, and a user terminal capable of interworking with the first base station or the second base station. For example, the server may be a mobile edge computing server (MEC server).

The server may receive and process a task offloaded from the user terminal through the second base station. To this end, a buffer for storing a task offloaded from a user terminal may be provided in the server.

Hereinafter, with reference to FIG. 2, a partial offloading method performed based on the system having the above structure will be described.

2 is a flowchart illustrating a partial offloading method in a wireless edge computing-based network according to an embodiment of the present invention.

For reference, the partial offloading method described later has been described as being performed by a user terminal, but is not limited thereto and may be performed by a separate device capable of communicating with the user terminal. Also, for convenience of explanation, the offloading target of the task is limited to the server.

Referring to FIG. 2 , in step S100 , offloading data for offloading a task generated in at least one user terminal connected to the second base station may be calculated.

인 태스크가 생성될 수 있다. In the user terminal, the size per 1 bit is

An in-task can be created.

Thereafter, in the step S100, offloading data for offloading a task based on a variable related to an offloading speed set based on a channel state between the first base station and the second base station and a channel state between the second base station and the user terminal. can be calculated. That is, offloading data for offloading a task generated in a user terminal to a server may be calculated according to variables related to a preset offloading speed in consideration of channel conditions of the wireless backhaul link and the wireless fronthaul link.

In this case, only a part of the task generated in the user terminal may be offloaded, and the rest may be processed in the user terminal. In other words, tasks may be partially offloaded according to variables related to offloading speed according to each channel state.

및

로 나타낼 수 있으며, 두 데이터의 크기 합은 생성된 태스크의 크기와 동일할 수 있다.To this end, in the step S100, when the offloading speed variable has a value between 0 and 1, the first data offloaded as part of the task and the third data processed by the user terminal as the remainder except for part of the task. 2 data can be calculated. Accordingly, the first data may be offloaded while the second data may be processed in the user terminal. For reference, the first data and the second data are respectively

and

, and the sum of the sizes of the two data may be equal to the size of the created task.

Of course, the offloading data further includes data for offloading all generated tasks (third data) and data for processing all generated tasks in the user terminal (fourth data) in addition to the data for partial offloading. can do.

In this regard, in the step S100, when the offloading speed variable has a value of 1, third data for offloading all generated tasks may be calculated, and the offloading speed variable is 0. If it has a value of , fourth data for processing all of the generated tasks in the user terminal may be calculated. Accordingly, while the third data is offloaded to the server, the fourth data may be processed in the user terminal. Task sizes of the third data and the fourth data are the same, and may be distinguished according to offloading.

Next, in step S200, waiting time and energy consumption required for offloading offloading data may be calculated.

The waiting time is the total waiting time and may include a waiting time when the task is offloaded and a waiting time when the task is processed in the user terminal. That is, the waiting time may include a waiting time required to process the first data and a waiting time required to process the first data.

In this regard, in the step S200, the total waiting time may be calculated based on Equation 1 below.

[Equation 1]

Here, Y ^t _ij is the total waiting time required to offload the offloading data, y ^t _ij,l is the waiting time required to process the second data, and y ^t _ij,r is the waiting time required to process the first data. Indicates the waiting time required for

In one embodiment, the waiting time required to process the first data is four types of waiting time, that is, transmission time (y ^t _ij ) of data offloaded from the user terminal to the second base station, from the second base station to the first base station The transmission time of the offloaded data (y ^t _jm,i ), the queue delay (y ^t _ij,q ) occurring in the server, and the time to process the offloaded data in the server (y ^t _ij,r ). can

When calculating the latency, it is preferable that the data transfer rate of the wireless fronthaul link satisfies a condition equal to or smaller than the data transfer rate of the wireless backhaul link.

The waiting time y ^t _ij,r required to process the first data can be expressed by the following equations.

[Equation 1a]

[Equation 1b]

[Equation 1c]

[Equation 1d]

[Equation 1e]

As an embodiment, y ^t _ij,l, which is a waiting time required to process the second data, may be expressed as in the following equation.

[Equation 1b]

The energy consumption amount is a total energy consumption amount and may include an energy consumption amount when the task is offloaded and an energy consumption amount when the task is processed in the user terminal. That is, the amount of energy consumption may include the amount of energy required to process the first data and the amount of energy required to process the second data.

In this regard, in step S200, the total energy consumption may be calculated based on Equation 2 below.

[Equation 2]

Here, ε ^t _ij is the total energy consumption required to offload the offloading data, e ^t _ij,l is the energy consumption required to process the second data, and e ^t _ij,r is the energy consumption required to process the first data. refers to the amount of energy required for

In an embodiment, e ^t _ij,r, which is the amount of energy consumed to process the first data, may be expressed as in the following equation.

[Equation 2a]

In an embodiment, e ^t _ij,l, which is the amount of energy consumed to process the second data, may be expressed by the following equation.

[Equation 2b]

For reference, the parameters applied to the above equations can be summarized as follows.

Next, the offloading data may be processed based on the waiting time and energy consumption calculated in step S300.

In other words, the offloading method of the task created in the user terminal is determined according to the variable related to the offloading speed, and if partial offloading is possible, the standby calculated according to offloading and non-offloading (processing in the user terminal) Based on time and energy consumption, some of the tasks can be offloaded while the rest can be handled by the user terminal.

Meanwhile, after step S200, a step of optimizing the offloading bandwidth by applying the calculated waiting time and energy consumption to a pre-learned learning algorithm may be performed.

That is, the offloading bandwidth can be optimized by applying a learning algorithm described later in order to reduce overhead generated on the MEC network by minimizing the latency and energy consumption calculated during partial offloading.

Hereinafter, a learning algorithm applied to partial offloading will be described with reference to FIGS. 3 and 4 .

3 is a diagram showing a framework of DDPG learning in an embodiment of the present invention, and FIG. 4 is a diagram showing execution codes for processing partial offloading in which DDPG learning is reflected in an embodiment of the present invention. to be.

In the learning algorithm applied to the partial offloading of the present invention, reinforcement learning may be applied to a Markov Decision Process (MDP) composed of a state space, an action space, and a utility function of an offloading environment.

In reinforcement learning, it selects the most advantageous action in a certain state, uses a policy to determine the optimal action, and gives a reward according to the action. However, the Markov Decision Process can be used to find these actions, rules, or rewards. The Markov decision process is based on action-oriented value evaluation, and the main purpose of the Markov decision process is to find the best decision-making rule and the maximum reward, that is, the decision-making rule and reward with the greatest sum of values according to actions.

The Markov decision process can be composed of utility functions and rules as concepts of state space, action space, and reward, and applying this to the partial offloading method of the present invention is as follows.

In one embodiment, the state space is a channel state between the first base station and the second base station, a channel state between the second base station and the user terminal, interference between cells covered by the first base station and the second base station, and offloaded in the server. It can contain a set of buffer states where tasks are stored.

In one embodiment, the action space may include a set of variables related to offloading speed and variables related to offloading bandwidth allocation.

In one embodiment, the utility function may be expressed based on Equation 3 below.

[Equation 3]

Here, X ^t is the state space, a ^t is the action space, ε ^t _ij is the total energy consumption required to offload the offloading data, and y ^t _ij is the offloading data The total waiting time required, λ is a weight variable related to task drop, Г ^t is a weight variable related to the waiting time and energy consumption of the user terminal, r ^t _jm is the first base station and the second base station Refers to the data transfer rate of the wireless backhaul.

As an embodiment, the rule may be expressed based on Equation 4 below based on the state space, action space, and utility function.

[Equation 4]

Reinforcement learning may be applied to the Markov decision process applied as described above. In this case, reinforcement learning may use deep deterministic policy gradient (DDPG) based on actors and critics.

DDPG is a generally known reinforcement learning technique, and its principle will be omitted, but it will be applied to the partial offloading method of the present invention and described as follows.

As shown in Figure 3, the primary network (Actor PN and Critic PN), the target network (Actor TN and Critic TN), and the replay buffer correspond to the three main components of reinforcement learning for offloading in the present invention. . We use function approximation using a deep neural network (DNN) for both the primary and target networks to compute the action value function of the critic.

로 표시되는 Q-함수를 근사화하기 위해 매개 변수 벡터 θ를 사용하며, 이는 리플레이 버퍼에 저장된 전환 경험(transition experiences)을 사용하여 업데이트 된다. 경험 튜플(experience tuple) e^t=(X^t, a^t, U^t, X^t+1)을 저장하기 위해 제한된 크기의 리플레이 버퍼를 사용하고, 경험 풀(experience pool)은 R로 표시된다. 이를 통해, 학습 프로세스 중 샘플 상관 관계를 분리하고 에이전트(agent)가 다양한 경험에서 효율적으로 학습 할 수 있다. 경험 리플레이 기술에 따르면, 각 시간 슬롯 C의 경험 풀 R에서 R0 경험의 미니 배치(mini-batch)를 무작위로 샘플링하여 프라이머리 및 타켓 네트워크의 네트워크 매개 변수를 업데이트 할 수 있다.

We use the parameter vector θ to approximate the Q-function denoted by , which is updated using the transition experiences stored in the replay buffer. We use a finite size replay buffer to store the experience tuple e ^t = (X ^t , a ^t , U ^t , X ^t+1 ), and the experience pool is denoted by R. Through this, sample correlation can be separated during the learning process and the agent can learn efficiently from various experiences. According to the experience replay technique, a mini-batch of R0 experiences from experience pool R at each time slot C is randomly sampled to update the network parameters of the primary and target networks.

을 줄임으로써 업데이트 될 수 있다.The critic network of deep reinforcement learning evaluates the policy specified by the actor based on the Q-value function. The primary critical network, Critic PN, is trained and updated by reducing the loss. As in FIG. 3, the loss function is composed of a target value of the target network utility and a critical value of the Critic PN. Therefore, the parameters of Critic PN are lost in mini-batch experiences sampled using the gradient descent algorithm.

can be updated by reducing

에서 타겟값 함수 y^t는 경험 유틸리티와 디스카운트된 미래 유틸리티의 합이며, τ는 디스카운트 인자이다.

는 Critic TN에서 얻은 타겟 Q-값이다. Critic PN 변수 θ는 θ에 대한 손실 L(θ)에 대해 미니 배치(mini-batch) 경사 하강법 알고리즘을 수행하여 업데이트 된다.math formula

In , the target value function y ^t is the sum of the experience utility and the discounted future utility, and τ is the discount factor.

is the target Q-value obtained from Critic TN. The critical PN variable θ is updated by performing a mini-batch gradient descent algorithm on the loss L(θ) for θ.

로 유도되는 Actor PN이 훈련된다.Actor PN takes as input the state for transition X and generates an action determined by the neural network parameter ω. Based on the current state, the critical value for the action taken in the Critic PN is displayed. Rules based on the criticism of the current rule using the gradient ascent algorithm.

Actor PN derived from is trained.

와 같이 나타낼 수 있다. 여기서, ηω는 액터 네트워크의 학습율을 의미함.Then, rule parameters are updated in the direction of improving performance. In particular, Actor PN uses the sampled rule gradient to update the parameter ω to maximize J, where ω is

can be expressed as Here, ηω is the learning rate of the actor network.

와

는 매개 변수를 직접 복사하는 대신 소프트 업데이트 후에 업데이트 된다. 타겟 크리틱(critic)/액터(actor) 네트워크는

,

와 같이 업데이트 된다.As shown in Figure 3, the target critical and actor networks can be defined as time-delayed copies of the primary critical and actor networks, respectively, and are used to calculate target values. . Parameters of Actor TN and Critic TN of target network

Wow

is updated after the soft update instead of directly copying the parameters. Target critic/actor network

,

updated as

Execution code for processing partial offloading based on reinforcement learning described above is shown in FIG. 4 .

Hereinafter, evaluation results for the partial offloading method will be described through comparison of task processing between the present invention and comparative examples. Comparative examples include a case in which all tasks are processed by a user terminal (UE execution), a case in which all tasks are offloaded to a server (MEC execution), and a case in which tasks are randomly offloaded (Random offloading).

5A is a graph showing overhead due to standby time according to an embodiment of the present invention compared to comparative examples, and FIG. 5B shows overhead due to energy consumption according to an embodiment of the present invention compared to comparative examples. This is a graph for comparison.

As shown in FIGS. 5A and 5B , it can be seen that the tendency of overhead generated in both the present invention and comparative examples is constant regardless of the number of experiments.

Meanwhile, the overhead value due to latency is large in the order of UE execution, MEC execution, random offloading, and DDPG-PTORA, and the overhead value due to energy consumption is large in the order of UE execution and MEC execution. It can be confirmed that it is almost similar, and since the user terminal has limited computing power, it can be said that the overhead generated is the highest compared to other task processing methods.

6 is a graph showing average task drop rates of user terminals compared to comparative examples according to an embodiment of the present invention.

As shown in FIG. 6, it can be seen that the partial offloading method of the present invention, DDPG-PTORA, has the lowest task drop rate compared to the other MEC execution and random offloading methods. Specifically, the total task drops of DDPG-PTORA, Random offloading, and MEC execution were 22.7964MB, 91.8372MB, and 229.5930MB, indicating that DDPG-PTORA is 75% lower than Random offloading and 90% lower than MEC execution.

7 is a graph showing total system overhead versus the number of user terminals compared to comparative examples according to an embodiment of the present invention.

As shown in FIG. 7 , it can be seen that as the number of user terminals increases, the total system overhead value increases in both the present invention and comparative examples.

On the other hand, it can be seen that the total system overhead value is larger in the order of UE execution, Random offloading, MEC execution, and DDPG-PTORA. However, when the number of user terminals is 90 or more, the total system overhead value of MEC execution increases more than the total system overhead of random offloading. This is because, if the number of user terminals continuously increases, the queue delay of the server and the bandwidth allocated to each user terminal decrease, thereby increasing the offloading time.

8 is a graph showing total system overhead versus task size compared to comparative examples according to an embodiment of the present invention.

The experiment of FIG. 8 was performed assuming that the number of user terminals is 50 and the number of second base stations is 5.

As shown in FIG. 8 , it can be seen that the total system overhead value of both the present invention and the comparative examples increases as the size of the task increases.

On the other hand, it can be seen that the total system overhead value is larger in the order of UE execution, Random offloading, MEC execution, and DDPG-PTORA.

Accordingly, according to an embodiment of the present invention, task offloading and resource allocation between wireless backhaul and wireless fronthaul can be scheduled in a mobile edge computing-based network.

The above description is merely an example of the technical idea of the present invention, and various modifications and variations can be made to those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments implemented in the present invention are not intended to limit the technical idea of the present invention, but to explain, and the scope of the technical idea of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed according to the following claims, and all technical ideas within the equivalent range should be construed as being included in the scope of the present invention.

Claims (9)

A partial offloading method in a wireless edge computing-based network comprising at least one second base station located within coverage of a first base station and a server connected to the first base station,
calculating offloading data for offloading a task generated from at least one user terminal connected to the second base station;
calculating standby time and energy consumption required to offload the offloading data; and
processing the offloading data based on the calculated waiting time and energy consumption;
Partial offloading method in a wireless edge computing-based network comprising a.
According to claim 1,
The first base station and the second base station interwork through a wireless backhaul, and the second base station and the at least one user terminal interwork through a wireless fronthaul;
Partial offloading method in a wireless edge computing-based network, characterized in that the wireless backhaul and the wireless fronthaul share a frequency bandwidth.
According to claim 1,
The calculation step is
Calculating offloading data for offloading a task based on a variable related to an offloading rate set based on a channel state between the first base station and the second base station and a channel state between the second base station and the user terminal. Partial offloading method in a wireless edge computing-based network characterized by.
According to claim 3,
The calculation step is
When the variable related to the offloading speed has a value between 0 and 1, calculating first data offloaded as part of a task and second data processed by the user terminal as the rest except for a part of the task A partial offloading method in a wireless edge computing-based network with
According to claim 1,
The calculation step is
Calculating the waiting time based on Equation 1 below showing the waiting time required to process the first data and the second data, and showing the energy consumption required to process the first data and the second data Partial offloading method in a wireless edge computing-based network, characterized in that for calculating the energy consumption based on Equation 2.
[Equation 1]

[Equation 2]

Here, y ^t _ij is the total waiting time required to offload the offloading data, y ^t _ij,l is the waiting time required to process the second data, and y ^t _ij,r is the first data , ε ^t _ij is the total energy consumption required to offload the offloading data, e ^t _ij,l is the energy consumption required to process the second data, e ^t _{ij ,r} means energy consumption required to process the first data.
According to claim 1,
Further comprising optimizing offloading by applying the calculated waiting time and energy consumption to a pre-learned learning algorithm,
The learning algorithm is
A partial offloading method in a wireless edge computing-based network, characterized in that reinforcement learning is applied to a Markov Decision Process composed of a state space, an action space, and a utility function of an offloading environment.
According to claim 6,
The state space is a channel state between a first base station and a second base station, a channel state between the second base station and the user terminal, interference between cells covered by the first base station and the second base station, and offloaded information in the server. Contains a set of buffer states in which tasks are stored;
The action space includes a set of variables related to offloading speed and variables related to offloading bandwidth allocation;
The utility function is a partial offloading method in a wireless edge computing-based network, characterized in that it can be expressed based on Equation 3 below.
[Equation 3]

Here, X ^t is the state space, a ^t is the action space, ε ^t _ij is the total energy consumption required to offload the offloading data, and y ^t _ij is the offloading data The total waiting time required, λ is a weight variable related to task drop, Г ^t is a weight variable related to the waiting time and energy consumption of the user terminal, r ^t _jm is the first base station and the second base station Refers to the data transfer rate of the wireless backhaul.
According to claim 7,
The reinforcement learning method of partial offloading in a wireless edge computing-based network, characterized in that using an actor and critic-based Deep Deterministic Policy Gradient (DDPG).
a first base station;
a second base station located within the coverage of the first base station;
at least one user terminal connected to the second base station; and
A server for processing tasks offloaded from the user terminal
A system for partial offloading in a wireless edge computing-based network comprising a.

Images