Multi-user collaboration platform based on edge computing and directed acyclic graph and method thereof

Abstract

The application discloses a multi-user cooperation platform based on edge computing and a directed acyclic graph, and the cooperation method is based on an edge computing platform, wherein the edge computing platform comprises a base station, an external state acquisition module, a deep reinforcement learning decision module and a coordination system; the external environment information acquisition module is used for acquiring the channel gain between the current base station and each user, so that the coordination system knows the current Internet condition; the deep reinforcement learning decision module is used for outputting a multi-user cooperation strategy, an unloading strategy and a local computing frequency strategy; the deep reinforcement learning decision module comprises a multi-element neural network unit, an intelligent agent, a convergence training unit and an experience pool; and the application solves the functions of dynamic change of transmission data volume of multi-user cooperation, dynamic change of user sub-task unloading decision and dynamic change of the local computing frequency of each user, and optimizes application performance under the joint consideration of delay, cooperation gain and energy consumption.

Description

Technical fields:

[0001] This invention relates to the fields of multi-access edge computing and multi-user collaboration in the Internet, and particularly to a multi-user collaboration platform and method based on edge computing and directed acyclic graphs. Background technology:

[0002] Currently, when individual users complete tasks alone, there are often problems with poor application performance. For example, in the field of autonomous driving, existing single-vehicle perception often struggles to guarantee vehicle safety due to issues such as obstacle obstruction and the inherent limitations of the perception algorithm's accuracy.

[0003] Furthermore, in the field of drone target tracking, there is also the problem of single-drone tracking easily losing track of targets. Considering the above issues, an increasing number of scenarios involving multiple mobile devices running the same application utilize the idea of ​​multi-user collaboration. That is, users can collaboratively process the application by sharing intermediate results of subtasks to improve application performance. This is because the same program often has consistent data requirements, and sharing intermediate results can improve their application performance. For example, in the collaborative perception of connected autonomous vehicles, different vehicles share their perception information, such as feature data extracted from camera detection results, to improve their target detection accuracy and perception range. Figure 1 As shown, the step of generating perception results depends on the feature extraction data. After receiving the feature data from vehicles one and three, vehicle two can expand the perception range and improve the accuracy of perception, which is very beneficial for autonomous driving in complex road conditions.

[0004] The aforementioned applications, such as target perception, involve large computational demands and require high real-time performance. They are typical computationally intensive and latency-sensitive applications. Mobile devices typically face challenges in executing these applications due to excessive computational demands making real-time performance difficult to guarantee, and low energy reserves hindering high energy consumption. Multi-access edge computing (MLC) has become an effective solution to these problems. The core of MLC's solution lies in computation offloading. By deploying relatively resource-rich edge computing platforms near base stations close to mobile users, users can wirelessly transfer their computational tasks to nearby base stations and receive the results after execution. Compared to traditional cloud computing, which places high-power servers far from users, MLC reduces network communication overhead by providing close-range services, significantly reducing application latency and energy consumption. This ensures efficient execution of latency-sensitive and computationally intensive applications on mobile devices with limited computing power and energy reserves.

[0005] Typically, computationally intensive applications consist of a series of interdependent subtasks, where the dependencies between subtasks can be modeled using a directed acyclic graph (DAG). According to the DAG, subtasks should be executed in a prescribed order, meaning a subtask can be started when it receives intermediate results from a preceding subtask. Combining these applications modeled as DAGs with computation offloading in multi-access edge computing allows for finer-grained offloading, further enhancing application performance through edge offloading. Numerous DAG task offloading strategies already exist. By incorporating the execution order of subtasks into the offloading strategy, we can make offloading decisions for each subtask, allowing mobile applications to execute completely in parallel across edge servers and mobile devices. This minimizes application execution latency and mobile device power consumption, improving overall application performance.

[0006] Most current research on edge offloading mechanisms based on directed acyclic graphs (DAGs) fails to consider multi-user cooperation, meaning each user's application executes independently. Each user acts as a decision-making agent, making independent offloading strategies for its subtasks based on its external environment (network conditions, computing resources of each node). Because each user only considers its own decisions and ignores those of other users, optimal performance is often difficult to achieve in multi-user scenarios. Furthermore, this approach does not consider the collaborative benefits of multi-user cooperation, which makes it difficult for the system to achieve optimal application performance.

[0007] In other works based on directed acyclic graphs (DAGs) that consider multi-user collaboration, the cooperative relationships between users are assumed to be fixed. Fixed dependencies exist between the subtasks of multiple users; that is, without considering the dynamic changes in computing resources and network conditions, users merge data of fixed types and sizes through fixed patterns. This approach is extremely prone to causing significant latency when network conditions fluctuate, ignoring the dynamic impact of the external environment on the collaborative effect. Furthermore, the aforementioned existing works also neglect the adjustability of local computing frequency, which is highly beneficial for optimizing task execution energy consumption.

[0008] The above analysis reveals two key points. First, the transmission rate during the offloading process from mobile devices to edge servers is influenced by the current network conditions. This transmission rate directly impacts latency and energy consumption, making it crucial to design an offloading scheme that minimizes latency and energy consumption. Second, due to factors such as network bandwidth and computing resources, we need to dynamically consider whether multi-user cooperation is necessary and how much data needs to be transmitted between users, rather than making a static decision about cooperation or non-cooperation. Therefore, our directed acyclic graph (DAG) is dynamically changing, meaning that whether multiple users cooperate affects the offloading decision and the local computing frequency decision, creating a high degree of computational complexity. Failure to quickly find an approximate optimal solution will significantly impact application performance. Summary of the Invention:

[0009] This invention provides a multi-user collaboration platform based on edge computing and directed acyclic graphs that adapts to dynamic network changes, and its application method. This invention uses deep reinforcement learning technology to solve the offloading strategy, user cooperation strategy, and local CPU frequency strategy for each subtask in computationally intensive applications, thereby improving the overall performance of each application in the system from three aspects: latency, energy consumption, and cooperation gain.

[0010] To address the problems existing in the prior art, the present invention adopts the following technical solution:

[0011] A multi-user collaboration platform based on edge computing and directed acyclic graphs (DAGs) is proposed. The collaboration method is based on an edge computing platform, which includes a base station, an external state acquisition module, a deep reinforcement learning decision-making module, and a coordination system.

[0012] The external environment information acquisition module is used to acquire the channel gain between the current base station and each user, so that the coordination system knows the current Internet situation.

[0013] The deep reinforcement learning decision module is used to output multi-user collaboration strategies, offloading strategies, and local computation frequency strategies; the deep reinforcement learning decision module includes multiple neural network units, agents, convergence training units, and experience pools.

[0014] The multi-element neural network units obtain the final decision by solving the uninstallation decision for a single user, the cooperative decision for multiple users, and the local computation frequency decision using the BSAC algorithm; that is:

[0015]

[0016] Where: α: weight of cooperative gain; G(t): cooperative gain in time slot t; dn,M(t): completion time of user n's application in time slot t; β: weight of delay; e n(t): Energy consumption of user n in time slot t; γ: Weight of energy consumption;

[0017] The intelligent agent generates a shared strategy through parallel computation based on the network state space provided by the external environment information acquisition module and the final decision provided by the multi-dimensional neural network unit.

[0018] The experience pool storage external environment information acquisition module provides historical decision information on the stored state between the current base station and each user;

[0019] The convergence training unit continuously updates the agent based on historical decision information and application performance provided by the experience pool.

[0020] To address the problems of the prior art, the present invention also adopts the following technical solution:

[0021] A multi-user collaboration platform based on edge computing and directed acyclic graphs includes the following steps:

[0022] The external environment information acquisition module obtains the channel gain between each user and the base station using the following formula.

[0023]

[0024] Among them: A d Antenna gain f c : Carrier frequency PL: Path loss index; Calculate the Euclidean distance between the user and the base station; It is based on the coordinates of the base station. and the coordinates of user n Calculate the distance between user n and the base station

[0025] The deep reinforcement learning decision module outputs the multi-user collaboration strategy, offloading strategy, and local computation frequency strategy in the following process: Step 2.1: A multi-neural network is constructed using the BSAC algorithm through the reinforcement learning network, outputting the following final decision:

[0026]

[0027] Where: α: weight of cooperative gain; G(t): cooperative gain in time slot t; dn,M(t): completion time of user n's application in time slot t; β: weight of delay; e n (t): Energy consumption of user n in time slot t; γ: Weight of energy consumption;

[0028] Step 2.2: The external environment information acquisition module updates the agent according to the following state space S;

[0029]

[0030] in: Let n be the channel gain for user n in time slot t.

[0031] Step 2.3: The agent calculates simultaneously using the information obtained from the state space and multiple neural networks, integrates the results of the multiple neural networks, and outputs the current action 'a' to the external environment information acquisition module.

[0032] Step 2.4: While continuously making decisions, the external environment information acquisition module generates a series of information and stores them in combination form as T = [s′, a, r, s]. This information is first input into the experience pool. During the training of the multivariate neural network, each combination is extracted from the experience pool. The training network will be performed at intervals based on the information to train and update the agent network; wherein:

[0033] s′: State space information of the previous time step; a: Action selected this time; r: Reward value of the previous selected action. Step 2.5: The external environment information acquisition module updates the system target of the coordination system using the following formula:

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] Furthermore, the cooperative gain G(t) of the time slot t is established by the following formula:

[0040]

[0041] Wherein: the decision for multi-user cooperation in time slot t of the system is when The amount of data transmitted at that time is O m1,m2 .

[0042] Furthermore, the energy consumption e of user n in time slot t n (t) is established by the following formula:

[0043]

[0044] In this invention, frequency decision is calculated locally. This controls local computing power consumption; local computing power consumption is calculated as the product of the CPU revolutions required for the task and the energy consumption per CPU revolution; the energy consumption per CPU revolution... The time required for local to edge transmission is The transmission frequency is P n .

[0045] Beneficial effects:

[0046] (1) The present invention is based on a multi-user collaboration strategy of directed acyclic graph: dynamically determines whether different cooperation relationships between multiple users need to exist and how much data needs to be transmitted, so as to achieve optimal application performance.

[0047] (2) The optimal unloading strategy of the task graph in edge scenarios: Under the dynamic user task graph, an approximately optimal unloading scheme is obtained for each task to optimize application performance.

[0048] (3) The present invention determines a dynamic local computing frequency for each user: dynamically adjusts the local computing frequency of each user to optimize application performance.

[0049] (4) This invention proposes a method for joint multi-user cooperative decision-making and task offloading decision-making in multi-user multi-access edge computing scenarios, thereby achieving the function of optimizing application performance.

[0050] (5) This invention proposes that the amount of data transmitted in multi-user cooperation can be dynamically changed, and the local computing frequency of each user can also be dynamically variable. It optimizes application performance by considering latency, cooperation gain, and energy consumption.

[0051] (6) This invention proposes a BSAC algorithm adapted to large decision spaces. Based on the original SAC algorithm, the Actor part is extended to multiple neural networks for parallel decision-making, which improves the performance of the system in large decision spaces. Attached image description:

[0052] Figure 1 This invention relates to a structural schematic diagram of multi-user collaborative relationships in vehicle perception;

[0053] Figure 2 This is a schematic diagram of the multi-user collaboration strategy module based on edge computing and directed acyclic graph in this invention;

[0054] Figure 3 This is a flowchart of the external environment information acquisition module in this invention;

[0055] Figure 4 This is a schematic diagram of the deep reinforcement learning network structure in this invention. Detailed Implementation

[0056] This invention proposes, in conjunction with the following appendix Figure 2 ~Attached Figure 4 The implementation process of this patent application will be described in further detail.

[0057] This invention provides a multi-user collaboration method based on edge computing and directed acyclic graphs. The collaboration method is based on an edge computing platform, which includes a base station, an external state acquisition module, a deep reinforcement learning decision module, and a coordination system, i.e., as shown below. Figure 2 As shown. Wherein:

[0058] The external environment information acquisition module is used to acquire the channel gain between the current base station and each user, so that the coordination system knows the current Internet situation.

[0059] The deep reinforcement learning decision module is used to output multi-user collaboration strategies, offloading strategies, and local computation frequency strategies; the deep reinforcement learning decision module includes multi-dimensional neural network units, agents, convergence training units, and experience pools, wherein:

[0060] The multi-element neural network units obtain the final decision by solving the uninstallation decision for a single user, the cooperative decision for multiple users, and the local computation frequency decision using the BSAC algorithm; that is:

[0061]

[0062] Where: α: weight of cooperative gain; G(t): cooperative gain in time slot t; dn,M(t): completion time of user n's application in time slot t; β: weight of delay; e n (t): Energy consumption of user n in time slot t; γ: Weight of energy consumption.

[0063] The intelligent agent generates a shared strategy through parallel computation based on the network state space provided by the external environment information acquisition module and the final decision provided by the multi-dimensional neural network unit.

[0064] The experience pool storage external environment information acquisition module provides historical decision information on the stored state between the current base station and each user;

[0065] The convergence training unit continuously updates the agent based on historical decision information and application performance provided by the experience pool.

[0066] To address the problems of the prior art, the present invention also adopts the following technical solution:

[0067] A multi-user collaboration method based on edge computing and directed acyclic graphs includes the following steps:

[0068] The external environment information acquisition module obtains the channel gain between each user and the base station using the following formula.

[0069]

[0070] Among them: A d Antenna gain f c : Carrier frequency PL: Path loss index; Calculate the Euclidean distance between the user and the base station; It is based on the coordinates of the base station. and the coordinates of user n Calculate the distance between user n and the base station

[0071] The deep reinforcement learning decision module outputs multi-user collaboration strategy, offloading strategy, and local computation frequency strategy in the following process: Step 2.1: Construct a multivariate neural network using the BSAC algorithm through a reinforcement learning network; In constructing the multivariate neural network, this invention trains multiple individual neural networks for the actor part according to different action attributes based on the convergence training unit, and inputs the results into a critic network for training, ultimately continuously updating the agent. The agent obtains the final decision by solving the offloading decision for a single user, the multi-user collaboration decision, and the local computation frequency decision. This method can effectively solve the problem of difficulty in convergence caused by the excessively large action space in deep reinforcement learning. The final decision is as follows:

[0072]

[0073] Where: α: weight of cooperative gain; G(t): cooperative gain in time slot t; dn,M(t): completion time of user n's application in time slot t; β: weight of delay; e n (t): Energy consumption of user n in time slot t; γ: Weight of energy consumption; Simultaneously:

[0074] The cooperation gain G(t) of time slot t: Since multi-user cooperation has marginal utility, meaning that the performance improvement is minimal when the amount of transmitted data is too large, we set the cooperation gain to log10 of the transmitted data amount. The decision for multi-user cooperation in time slot t in the system is... when The amount of data transmitted at that time is O m1,m2 .

[0075]

[0076] The completion time dn,M(t) of application n in time slot t: In the task dependency graph, the completion of the last subtask is considered as the completion of the entire application. The completion time of the last subtask (numbered M) is taken as the completion delay of application n.

[0077] User n's energy consumption e in time slot t n (t): Consists of local computing energy consumption and transmission energy consumption to the edge, and is expressed by the following formula:

[0078]

[0079] In this invention, frequency decision is calculated locally. This is used to control local computing power consumption. Local computing power consumption is calculated as the product of the CPU revolutions required for the task and the energy consumption per CPU revolution. The time required for local to edge transmission is The transmission frequency is P n

[0080] Step 2.2: The external environment information acquisition module updates the agent according to the following state space S;

[0081]

[0082] in: Let n be the channel gain for user n in time slot t.

[0083] Step 2.3: The agent calculates simultaneously using the information obtained from the state space and multiple neural networks, integrates the results of the multiple neural networks, and outputs the current action 'a' to the external environment information acquisition module.

[0084] Step 2.4: While continuously making decisions, the external environment information acquisition module generates a series of information and stores them in combination form as T = [s′, a, r, s]. This information is first input into the experience pool. During the training of the multivariate neural network, each combination is extracted from the experience pool. The training network will be performed at intervals based on the information to train and update the agent network; wherein:

[0085] s′: State space information of the previous time step; a: Action selected this time; r: Reward value of the previous selected action. Step 2.5: The external environment information acquisition module updates the system target of the coordination system using the following formula:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A collaborative method based on edge computing and directed acyclic graphs, characterized in that, Includes the following steps: The channel gain between each user and the base station is obtained using the following formula. ; =A d ( ) PL Among them: A d Antenna gain f c : Carrier frequency PL: Path loss index; This represents the Euclidean distance between the user and the base station; It is based on the coordinates of the base station ( .x, .y), and the coordinates of user n ( .x, .y), calculate the distance between user n and the base station. =(( .x- .x) 2 -( .y- .y) 2 ) 0.5 ; The process of outputting multi-user collaboration strategy, offloading strategy, and local computing frequency strategy: Step 2.1: Construct a multi-neural network using the BSAC algorithm through a reinforcement learning network, and output the following final decision: Where: α: weight of the cooperative gain; G(t): cooperative gain in time slot t; : The completion time of user n's application in time slot t; β: The weight of the delay; e n (t): Energy consumption of user n in time slot t; γ: Weight of energy consumption; Step 2.2: Based on the following state space Update the intelligent agent; S=[ , ] in: Let n be the channel gain for user n in time slot t; Step 2.3: The agent calculates based on the information obtained from the state space and multiple neural networks simultaneously, integrates the results of multiple neural networks, and outputs the current action to the α external environment information acquisition module; Step 2.4: While making decisions, the external environment information acquisition module generates a series of information and stores them in a combined form. First, these combinations are input into an experience pool. During training of the multivariate neural network, each combination is extracted from the experience pool. The network will be trained and updated periodically based on this information. Wherein: : State space information from the previous time step; The action chosen this time; The reward value of the previous selection action; Step 2.5: Update the coordination system target using the following formula: 。 2. The collaborative method based on edge computing and directed acyclic graphs according to claim 1, characterized in that, The cooperative gain G(t) of time slot t is established by the following formula: G(t)=log10( O m1,m2 ) Wherein: the decision for multi-user cooperation in time slot t of the system is ,j∈{1,2,...,J};when The amount of data transmitted when =1 is O m1,m2 .

3. The collaborative method based on edge computing and directed acyclic graphs according to claim 2, characterized in that, The energy consumption e of user n in time slot t n (t) is established by the following formula: Among them, frequency decision is made through local computation. The range ∈ [0.1,1] is used to control local computing power consumption; the local computing power consumption is calculated as the product of the number of CPU revolutions required for the task and the energy consumption per CPU revolution; the energy consumption per CPU revolution... The time taken for local transmission to the edge is The transmission frequency is P n .

4. A multi-user collaboration platform based on edge computing and directed acyclic graphs, characterized in that: The multi-user collaboration platform implements the method described in any one of claims 1-3, and the multi-user collaboration platform includes a base station, an external environment information acquisition module, a deep reinforcement learning decision module, and a coordination system, wherein: The external environment information acquisition module is used to acquire the channel gain between the current base station and each user, so that the coordination system knows the current Internet situation. The deep reinforcement learning decision module is used to output multi-user collaboration strategies, offloading strategies, and local computation frequency strategies; the deep reinforcement learning decision module includes multi-functional neural network units, agents, convergence training units, and experience pools; wherein: The multi-element neural network units obtain the final decision by solving the uninstallation decision for a single user, the cooperative decision for multiple users, and the local computation frequency decision using the BSAC algorithm; that is: Where: α: weight of the cooperative gain; G(t): cooperative gain in time slot t; : The completion time of user n's application in time slot t; β: The weight of latency; en(t): The energy consumption of user n in time slot t; γ: The weight of energy consumption; The intelligent agent generates a shared strategy through parallel computation based on the network state space provided by the external environment information acquisition module and the final decision provided by the multi-dimensional neural network unit. The experience pool storage external environment information acquisition module provides historical decision information on the stored state between the current base station and each user; The convergence training unit continuously updates the agent based on historical decision information and application performance provided by the experience pool.

Images