A multi-person AR system data offloading and resource allocation method under a 6G network framework

Abstract

The application provides a multi-person AR system data offloading and resource allocation method under a 6G network framework, relates to the technical field of 6G mobile communication, selects an AR user as a Host user in each time slot, and the rest are Resolver users; the interaction weight between the Host user and the Resolver user is calculated by using a graph attention network; a data offloading model based on Stackelberg game is established, a multi-party utility function is constructed, a double-layer optimization problem is converted into a single-layer optimization problem, and a sequential quadratic programming algorithm is adopted to solve, so that the optimal data offloading decision is obtained; the resource allocation process is constructed as a Markov decision process, a reward function containing average delay and delay variance is defined, a double-delay deep deterministic policy gradient algorithm is adopted to output a resource allocation strategy of a continuous action space; through the combination of game theory and reinforcement learning, the environmental data redundancy and server rendering load are effectively reduced, the interaction consistency among multi-users is optimized while ensuring low delay, and the collaborative experience of the multi-person AR system is improved.

Description

Technical Field

[0001] This invention relates to the field of 6G mobile communication technology, and in particular to a method for data offloading and resource allocation in a multi-user AR system under a 6G network framework. Background Technology

[0002] With the development of mobile augmented reality (AR) technology, multi-user AR systems are gradually becoming an important application in smart cities, industrial collaboration, remote education, and immersive entertainment. In these scenarios, multiple users need to simultaneously perceive, operate, and share the same virtual object or environment, and the core experience relies on the consistency of multiple users' understanding of the virtual object's state. In multi-user AR systems, different users' terminal devices need to maintain as consistent a state update as possible for the same virtual object, including position, posture, and interaction results. If there are significant time delay differences in state updates between different users, it will cause the virtual object to appear asynchronous or even misaligned from different perspectives, thus undermining the immersiveness and reliability of collaborative interaction. Therefore, consistency has become one of the key issues restricting the quality of multi-user AR system experiences.

[0003] However, in practical applications, the environmental information collected by different users from their respective perspectives often has a high degree of overlap. If each user uploads complete raw environmental data to the server, it will not only cause a large amount of redundant data transmission but also significantly increase the consumption of communication bandwidth and the server's computing load. On the other hand, the edge server based on the 5G network architecture performs rendering calculations and distributes results to multiple users separately, lacking a reasonable decision-making mechanism, leading to duplicate calculations and wasted resources. Therefore, for multi-user AR applications, it is urgent to explore efficient data offloading and resource allocation methods under the 6G network framework. Summary of the Invention

[0004] This invention proposes a data offloading and resource allocation method for a multi-user AR system under a 6G network framework, in order to solve the problems of data redundancy and server re-rendering caused by overlapping perspectives of multiple users when users make data offloading decisions in existing multi-user AR systems.

[0005] This invention provides a method for data offloading and resource allocation in a multi-user AR system under a 6G network framework, the method comprising the following steps:

[0006] Step S1: In each time slot, select one AR user from multiple AR users as the Host user, and the remaining users as Resolver users;

[0007] Step S2: Calculate the interaction weights between the Host user and each Resolver user using a graph attention network; the interaction weights are used to characterize the intensity of interaction between users.

[0008] Step S3: Establish a data offloading model based on Stackelberg game, construct the environmental data upload process and the rendering data acquisition process as a sequential game with a leader-follower structure, and establish a multi-party utility function that includes the Host user, Resolver user and MEC server in the upload and rendering stages.

[0009] Step S4: Under the condition of satisfying the participation constraints, the two-layer game optimization problem constructed in step S3 is transformed into a single-layer optimization problem with only the Host user decision variables, and the sequential quadratic programming algorithm is used to solve it to obtain the optimal data unloading decision. The data unloading decision includes the optimal amount of data uploaded by the Host user and the rendering price, as well as the optimal amount of data uploaded by each Resolver user and the rendering data acquisition method.

[0010] Step S5: Construct the resource allocation process as a Markov decision process, and define the state space, action space and reward function, wherein the reward function includes the multi-user end-to-end average latency and latency variance, and the latency variance is used to characterize the interaction consistency among multiple users.

[0011] Step S6: Train the algorithm using a dual-delay deep deterministic strategy gradient algorithm to output a resource allocation strategy for the continuous action space. The resource allocation strategy includes the uplink bandwidth ratio, downlink bandwidth ratio, and edge computing resource ratio allocated to each user. Then, allocate communication and computing resources to each user according to the resource allocation strategy.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] 1. Reduce data redundancy and communication overhead: By introducing Host-Resolver role division and dynamically modeling the interaction intensity between users using graph attention networks, combined with the Stackelberg game-based upload and rendering path selection mechanism, Resolver users are guided to selectively upload supplementary data and obtain rendering results from Host users, significantly reducing the amount of redundant environment data uploaded and the amount of downlink rendering data from the MEC server, thereby reducing the system communication load.

[0014] 2. Addressing the problem of selfish user decision-making: By using Stackelberg game theory to model the leader-follower relationship between Host and Resolver users, the two-level optimization is transformed into a single-level optimization and solved using the SLSQP algorithm. This allows Host users to formulate upload and pricing strategies based on the expected optimal response from Resolver users, thereby improving the overall system utility and avoiding resource waste caused by maximizing individual interests.

[0015] 3. Balancing real-time performance and interactive consistency for multiple users: The resource allocation problem is modeled as a Markov decision process. The average end-to-end latency and latency variance are introduced into the reward function. The TD3 algorithm is used to achieve adaptive and continuous allocation of uplink bandwidth, computing resources and downlink bandwidth. While ensuring low latency, the latency difference between users is effectively controlled, which improves the immersiveness and consistency of multi-user collaboration.

[0016] 4. The algorithm is stable in convergence and adaptable to dynamic environments: The TD3 algorithm overcomes the Q-value overestimation problem of traditional DDPG through a dual Critic network and a delayed update mechanism, and achieves stable training convergence under dynamic channel and task load environments; combined with the basic share mechanism, it ensures the minimum resources for each user in extreme scenarios, thereby enhancing the robustness of the system.

[0017] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0018] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0019] Figure 1 This invention provides a multi-user AR data unloading architecture.

[0020] Figure 2 This is a resource allocation block diagram based on TD3 in this invention;

[0021] Figure 3 This is a graph showing the change of Host user upload utility over time frames under different strategies of this invention;

[0022] Figure 4 This is a graph showing the variation of MEC downlink rendering volume over time frames under different strategies of this invention;

[0023] Figure 5 This invention illustrates the variation of environmental data unloading volume over time frames under different strategies.

[0024] Figure 6 This is a graph showing the reward convergence curves of the TD3 algorithm and the DDPG algorithm of this invention.

[0025] Figure 7 This is a reward convergence graph for different learning rates in this invention;

[0026] Figure 8 This is a graph showing the average latency variation of each algorithm under different user scales in this invention;

[0027] Figure 9This is a graph showing the variation of the standard deviation of latency for each algorithm under different user scales in this invention;

[0028] Figure 10 This is a graph showing the average latency variation of each algorithm under different bandwidths in this invention;

[0029] Figure 11 This is a graph showing the variation of the standard deviation of latency for each algorithm under different bandwidths in this invention. Detailed Implementation

[0030] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0031] In multi-user AR systems, users often prioritize maximizing their own utility when making data offloading decisions, rather than actively pursuing the optimal overall system benefit. This selfish decision-making characteristic can easily affect overall service quality and collaborative efficiency. Therefore, to address the data redundancy and server-side repetitive rendering issues caused by overlapping user perspectives, this invention designs a data offloading model based on a combination of Graph Attention Network (GAT) and Stackelberg game theory. By utilizing the graph attention mechanism to dynamically characterize the interaction intensity between users, interaction weights that change over time are formed. Furthermore, under limited resource competition, users tend to prioritize strategies that maximize their own gains. The environmental data upload process and the rendering data acquisition process are modeled as a sequential game problem with a leader-follower structure, establishing a multi-party utility function that includes the upload and rendering phases. Under the premise of satisfying participation constraints, the two-layer game problem is transformed into a single-layer optimization problem, and a solution method based on Sequential Least Squares Programming (SLSQP) is designed to achieve optimal data offloading within each time slot.

[0032] To address the issue of high latency and significant latency variations among users due to resource constraints in multi-user AR scenarios, this invention constructs a latency model encompassing uplink transmission, edge computing, and downlink transmission, using average latency and latency variance as comprehensive indicators to measure real-time performance and consistency. The joint resource allocation problem is transformed into a Markov decision process, and a continuous action space resource allocation strategy based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is designed to achieve adaptive proportional allocation of communication and computing resources.

[0033] This article considers a... A multi-user collaborative system consisting of augmented reality users, denoted as [user set denoted as ]. Within each time slot, the system selects the user with the most effective interactions with other users as the Host user, denoted as [username]. The remaining users act as resolvers, denoted as... , When a Host user uploads environment-aware data to the server, a Resolver user can also collect environment data and upload it to the server to support the construction of a shared virtual environment. At the same time, each Resolver user determines how to obtain rendering data based on its interaction with the Host user and the pricing mechanism.

[0034] like Figure 1 As shown, in the considered multi-user augmented reality scenario, the overall system operation flow is as follows:

[0035] ① Each AR user uses their own terminal device to perceive the surrounding real environment and collect local environmental data, including images and point clouds, to provide basic information for the construction of a shared virtual environment. Due to the overlap in the observation perspectives of different users, the data they collect often has a certain degree of spatial redundancy.

[0036] ② Within each time slot, the system selects one user from all users as the Host user, and the remaining users as Resolvers. The Host user is responsible for uploading the main environmental awareness data, which is used to construct the core content of the shared virtual space. The Resolver users, based on their spatial location and interaction relationship with the Host user, selectively upload some environmental data to supplement areas not covered or insufficiently covered by the Host user, thereby reducing redundant data transmission to a certain extent.

[0037] ③ After receiving environment data from the Host user and each Resolver user, the MEC server processes all uploaded data uniformly and completes the construction of the shared virtual environment. Subsequently, the server performs rendering calculations based on the uploaded environment data, generating rendering results relevant to the perspective of different users. During this process, the server does not require the Host user to locally aggregate the Resolver user's data; instead, it directly processes the data uploaded by each user centrally.

[0038] ④ After the virtual environment is built, the Host user requests the corresponding rendering data from the MEC server according to its own perspective requirements to meet its local display and interaction needs. Simultaneously, the Host user also undertakes a certain data forwarding function, providing rendering data support to some Resolver users in subsequent stages.

[0039] ⑤ Finally, based on the interaction relationships with each Resolver user, the Host user forwards the rendering data to some Resolver users who have interactive needs. Resolver users can also obtain the required rendering data from the Host user or the MEC server according to the system pricing mechanism and interaction constraints, thereby completing collaborative display and interaction in multi-user augmented reality scenarios.

[0040] I. Dynamic modeling of interaction relationships;

[0041] (1) Spatiotemporal graph modeling and node feature definition;

[0042] Consider by A multi-user system consisting of AR users, in discrete time frames The system is represented as a set of time sequence diagrams. Each user corresponds to a node in the diagram, and the node is in time frame. Its state is represented by its spatiotemporal trajectory information.

[0043] Let the first Individual users in time frames The spatial location is:

[0044]

[0045] To reflect the user's movement trend, a velocity vector obtained from the position difference between adjacent time frames is introduced:

[0046] Therefore, node In time frame The eigenvectors are defined as follows:

[0047]

[0048] (2) Definition of adjacency matrix and interaction conditions;

[0049] In multi-user AR scenarios, not every two users have a potential interaction relationship. To reduce computational complexity and irrelevant noise, a dynamic prior adjacency matrix is ​​introduced to filter potentially interactive user pairs.

[0050] set up Represents time frame The adjacency matrix, where Defined as:

[0051]

[0052] The interaction conditions should specifically consider the following four points:

[0053] 1. Euclidean distance between users:

[0054]

[0055] 2. Similarity in velocity direction:

[0056]

[0057] 3. Difference in speed magnitude:

[0058]

[0059] 4. Relative motion trend:

[0060]

[0061] If the above preset threshold constraints are met, then the user is considered... and In time frame There are interactive relationships; at the same time, in order to preserve the node's own information, set... .

[0062] (3) Interaction modeling based on graph attention networks;

[0063] After obtaining the dynamic adjacency matrix, a graph attention network is introduced to model the interaction strength between users. This method allows for the assignment of continuous weight values ​​to different user pairs under adjacency constraints.

[0064] 1. Characteristic linear transformation;

[0065] For the Each attention head first performs a linear mapping of node features:

[0066]

[0067] 2. Attention score calculation;

[0068] Based on the additive attention mechanism, nodes For nodes The attention score is defined as:

[0069]

[0070] in, and These are learnable attention parameters.

[0071] 3. Normalize attention weights;

[0072] Use Softmax for each node Neighbors (including the node itself) normalization:

[0073]

[0074] This weight represents the weight in the first... Under one attention head, the time frame is Time user For users The relative degree of influence.

[0075] (4) Interactive value output format

[0076] Through multi-head attention mechanisms, interaction weight values ​​in the following form can be obtained:

[0077]

[0078] in, This indicates the number of attention heads. This value is used to describe the intensity of subsequent dynamic user interactions.

[0079] II. Utility Modeling in the Upload Phase;

[0080] Suppose that within a certain time slot, the amount of environmental data uploaded by the Host user is... The amount of environment data uploaded by each Resolver user is This paper assumes that the Host user and each Resolver user directly upload environment data to the server, and the server processes all uploaded data uniformly, without requiring the Host user to perform local aggregation of the Resolver user's data.

[0081] The total amount of environmental data received by the server is:

[0082]

[0083] In multi-user AR scenarios, video frames and environmental point clouds collected by different users often exhibit significant overlap. As the shared virtual environment gradually becomes more complete, the marginal benefit of continuing to upload raw environmental data gradually decreases. Therefore, the overall benefit of the server for environmental data is modeled using a concave function with diminishing marginal returns:

[0084]

[0085] in, This represents the weighting coefficient of environmental integrity on the benefits of scene reconstruction.

[0086] The server pays all uploaders based on the unit of environmental data, with the unit payment price denoted as [missing information]. Therefore, the server's utility function during the upload phase is:

[0087]

[0088] superscript This indicates the transmission phase.

[0089] The host user's utility during the upload phase is as follows:

[0090]

[0091] in This represents the cost per unit of data uploaded by the Host user.

[0092] The utility of each Resolver user during the upload phase is:

[0093]

[0094] in The unit cost of uploading data for Resolver users.

[0095] III. Utility Modeling in the Rendering Stage;

[0096] After the virtual environment is built, each user needs to obtain the rendering results related to their own perspective. To characterize the spatiotemporal interaction between the Host user and each Resolver user, this paper uses a graph attention network to obtain the interaction weights between the Host user and the Resolver users. Interaction weights reflect the probability that the Host user and the Resolver user are in the same interaction area, as well as the degree of scene similarity.

[0097] Suppose that within a certain time slot, the amount of rendering data obtained by the i-th Resolver user from the Host user is... The amount of rendering data obtained from the server is Then the total amount of data rendered is:

[0098]

[0099] This reflects the constraints that interaction relationships impose on the selection of rendering data paths. When At this time, if the Resolver user and the Host user are too far apart or their motion characteristics are completely unrelated, they do not meet the conditions for cooperation. .when At that time, the system determines that there is a valid spatial interaction between the Resolver user and the Host user. .

[0100] (1) Rendering utility for Resolver users;

[0101] To demonstrate that more rendering data leads to a better user experience, and to avoid introducing the assumption of diminishing marginal returns during the rendering stage, this paper models the rendering benefit of the i-th Resolver user as a linear function:

[0102]

[0103] The rendering phase utility for Resolver users is:

[0104]

[0105] (2) The rendering effect of the Host user;

[0106] Host users also need to request rendering data from their own perspective from the server, denoted as... At the same time, the Host user also needs to forward rendering data to the Resolvers users with whom it has an interaction relationship.

[0107] Therefore, the total utility of the Host user during the rendering phase should be:

[0108]

[0109] in: The amount of rendering data requested by the Host user from the server; Host users must pay for the server price for their own rendering needs; The host user bears the cost of all data processing and forwarding.

[0110] (3) The rendering effect of the server;

[0111] The server distributes rendering data to all users, including the Host user and all Resolvers users:

[0112]

[0113] IV. Overall Utility Modeling

[0114] Total utility for Host users:

[0115]

[0116] Total utility for Resolver users:

[0117]

[0118] Overall utility of MEC server:

[0119]

[0120] V. Data offloading optimization issues;

[0121] In the previous section, utility models for the environmental data upload and rendering stages were established, and based on these models, the overall utility functions of the three participants in the game theory were presented. Intuitively, at each time frame, the Host user first announces the upload decision and the rendering price. Subsequently, each Resolver user, after observing these decisions, selects its own upload data volume and the rendering data obtained from the Host user or the server, thus forming a sequential game with a leader-follower structure.

[0122] This section further models the data offloading decision problem in multi-user AR scenarios within the Stackelberg game framework. The data offloading problem can be formulated as a two-level optimization problem under the Stackelberg game:

[0123]

[0124] The optimization objective for the Host user is to maximize their total utility within the Stackelberg game framework. Constraint C1 represents the amount of data uploaded and the rendering data acquisition strategy for each Resolver user, determined by the optimal response to the utility maximization problem, given the Host user's decision. C2 represents the composition of the total rendering data volume for Resolver users during the rendering phase. C3 represents the feasibility constraint for rendering data acquisition based on interaction relationships; when there is no effective interaction relationship between the Host user and the Resolver user, the Resolver user cannot acquire rendering data from the Host user. C4 represents the non-negativity constraint of relevant decision variables, ensuring that the relevant decision variables of both the Host user and the Resolver user are non-negative during the upload and rendering phases. C5 represents the participation constraint of Resolver users, ensuring the rationality of their participation in system collaboration. C6 represents the participation constraint of the server, requiring that the server's total utility in the data upload and rendering service process is not less than zero to ensure the sustainability of system operation.

[0125] VI. Data unloading optimization issues transformation;

[0126] From the aforementioned optimization problem P1, we know that the decision variables for the Host user are: The decision variables of each resolver user Given constraint C1, the essence of problem P1 is to maximize the utility of the optimal response under the given host user decision conditions. Therefore, problem P1 is a typical two-level Stackelberg game optimization problem, and solving it directly is quite complex. To reduce the complexity of solving the problem, it is necessary to transform this two-level problem into a single-level optimization problem with only the host user decision variables in mind.

[0127] (1) Solve for the optimal response of the Resolver user;

[0128] Resolver user utility function about It is separable, wherein:

[0129] Upload stage:

[0130]

[0131] Rendering stage:

[0132]

[0133] In the given and Under these conditions, the optimal rendering decision for the Resolver user can be parsed as follows:

[0134] when and hour:

[0135]

[0136] when or hour:

[0137]

[0138] (2) Substitute the optimal response into the Host user utility function

[0139] The result obtained in step one Substitute the total utility function of the Host user:

[0140]

[0141] This eliminates the decision variables for the Resolver user, making the Host user's utility function depend only on... .

[0142] (3) Transformation of two-level problems into single-level problems

[0143] After the above processing, the original problem P1 can be transformed into the following single-layer optimization problem:

[0144]

[0145] VII. Design of an optimization algorithm based on SLSQP;

[0146] The previous section transformed the original two-level Stackelberg game problem into a single-level nonlinear optimization problem with only the host user's decision variables. Because... For variables that are continuously differentiable and whose constraints are simple linear intervals or non-negative utility constraints, this section proposes to use the SLSQP algorithm for solving the problem.

[0147] In each time slot First, determine the Host user ID. The remaining users are designated as Resolver users, denoted as... Based on the user's spatial location and interaction characteristics, calculate the interaction weights between the Host user and each Resolver user. .

[0148] Interaction weight This is used to determine whether a valid interaction exists between the Host user and the Resolver user, and directly affects the data acquisition method for AR users during the rendering phase. Therefore:

[0149]

[0150] In other words, when there is no valid interaction relationship, the Resolver user cannot obtain rendering data from the Host user.

[0151] Rendering pricing for a given host user and interaction weights Then, each Resolver user selects the rendering data acquisition path based on the principle of maximizing their own utility.

[0152] The optimal response from the Resolver user is as follows:

[0153]

[0154] Therefore, the rendering decisions of each Resolver user can be represented as a function of the Host user's decision variables, and no longer need to be considered as independent optimization variables in the solution.

[0155] Substituting the optimal response obtained by the Resolver user into the total utility function of the Host user yields an objective function that depends only on the Host user's decision variables.

[0156] In the numerical solution process, the maximization problem is equivalently transformed into a minimization problem:

[0157]

[0158] In the single-level optimization problem for Host users, the following constraints must be satisfied simultaneously:

[0159] Variable value range constraints:

[0160]

[0161] Participation constraints:

[0162]

[0163]

[0164] Given the objective function and constraint set, the SLSQP algorithm is used to iteratively update the host user's decision variables. In each iteration, the algorithm constructs a local quadratic programming subproblem and updates the variables while satisfying the constraints. .

[0165] Considering the piecewise structure introduced by the optimal responses of the Resolver users in the objective function, to improve the stability of the solution, we can start from multiple sets of initial feasible points and select the solution that maximizes the total utility of the Host user as the final solution. The strategy is as follows:

[0166] 1. Provide multiple sets of representative initial values ​​for subsequent iterative solutions, and set several sets of initial points. To cover it: area area The area.

[0167] 2. Run SLSQP once for each set of initial points to obtain candidate solutions;

[0168] 3. Select feasible solutions that satisfy all constraints from the candidate solutions, and choose the one with the largest U as the final solution for this time slot.

[0169] Once the algorithm meets the convergence condition, it outputs the optimal decision for the Host user. Subsequently, according to Recalculate the optimal response for each resolver user. The consistency of all constraints is checked to ensure that the obtained solution is feasible in both physical and game theory terms. The specific implementation and update process is shown in Table 1.

[0170] Table 1. Stackelberg game solving algorithm based on SLSQP

[0171]

[0172] This invention focuses on a real-time processing workflow. Each AR device collects environmental data within its current time slot and uploads the raw data to the MEC server. The MEC completes rendering calculations and result encoding within the same time slot, and then distributes the rendered complete image to each terminal. The entire process constitutes an end-to-end real-time processing link. To highlight the resource allocation issue, this chapter abstracts the computational tasks in the virtual object rendering process into rendering computational tasks on the MEC side.

[0173] Within a single time slot, the task processing flow for each user can be summarized into three stages:

[0174] (1) Environmental data upload stage;

[0175] user AR devices in time slots The system collects current environmental and interactive information through sensors such as cameras and IMUs, performs simple encoding, and forms raw data related to rendering. Let the data volume be... The data is uploaded to the MEC server via the uplink. In the method proposed in Chapter 3, whether a user needs to upload data to the MEC and the scale of the uploaded data are jointly determined by the interaction between users and path selection decisions.

[0176] The data is uploaded to the MEC server via the uplink.

[0177] (2) MEC-side rendering and calculation stage;

[0178] MEC servers in time slots After receiving uplink data from all users, the system generates virtual object images or augmented reality overlay content for each user in the current scene. All rendering computation tasks are executed on the MEC (Multi-access Edge Computing) platform, meaning computation on local devices or other terminals is not considered.

[0179] (3) Rendering result delivery stage;

[0180] MEC encodes and compresses the rendered image to obtain the rendered result data that needs to be sent to the user. This is done within a time slot. Internal MEC is distributed to users The amount of data rendered after encoding is All users' rendering results are directly delivered by MEC. Whether a user obtains rendering results from MEC also depends on the interaction relationship and the rendering path selection result. That is, some users can obtain rendering data from the Host user path, thereby reducing the amount of downlink data from MEC.

[0181] This data is sent to the user in real time via the downlink. AR devices are used to display augmented reality images in this time slot.

[0182] Within each time slot, the end-to-end latency from the AR device acquiring current environmental data to the terminal receiving the rendering result consists of three parts: uplink and downlink transmission latency, and MEC rendering computation latency. The optimization goal of this chapter is to minimize the average end-to-end latency and differences among multiple users by rationally allocating communication and computing resources, while optimizing data offloading.

[0183] 8. Communication model;

[0184] The system has two types of communication links: uplink from the user to the MEC, and downlink from the MEC to the user. Both types of links use Orthogonal Frequency Division Multiple Access (OFDMA) technology. Let the set of available sub-channels in the system be... Each sub-channel bandwidth is .

[0185] (1) Uplink;

[0186] During the uplink phase, subchannel allocation from user equipment to MEC uses binary variables. Indicates: If in a time slot In the middle, the first Sub-channels are allocated to users Uplink transmission with MEC, ,otherwise .set up For users uplink transmit power, For time slots Internal users With MEC in sub-channel The channel gain and noise power are... Then the user In the time slot The uplink transmission rate is:

[0187]

[0188] (2) Downlink;

[0189] In the downlink phase, the subchannel allocation from MEC to user equipment uses binary variables. Indicates: If in a time slot In the middle, the first Each sub-channel is used by MEC to communicate with users. The downlink transmission, then ,otherwise .set up This refers to the downlink transmit power of the MEC. For time slots MEC and users In sub-channel The channel gain on the user The downlink transmission rate is:

[0190]

[0191] To facilitate subsequent resource allocation modeling, this chapter adopts a continuous control method of bandwidth ratio allocation in the DRL scheme, treating the total uplink / downlink bandwidth of the system as a continuously divisible resource, and allocating it among users through a ratio variable. and It can be understood as the equivalent rate under given bandwidth allocation ratio and average channel gain conditions.

[0192] IX. Computational Model;

[0193] In the time slot Inside, users The amount of data uploaded is After rendering and encoding are completed, the MEC (Multi-access Edge Computing) module distributes the data to the user. The rendering result data size is In real-world systems, the intermediate data generated during the rendering phase is typically much larger than the final encoded data sent out. To simplify modeling while reflecting the scale of data processed internally during rendering, this chapter introduces a scaling factor. Define user The amount of data rendered internally on the MEC side is:

[0194]

[0195] The computational load required for rendering can be expressed as:

[0196]

[0197] in, This parameter represents the computational complexity of rendering per unit of data, used to characterize the uncertainties related to scene complexity, lighting models, and materials. Let:

[0198]

[0199] in, and These are the shape parameter and the scale parameter, respectively.

[0200] Assume the total computing power of the MEC server is (Unit: CPU cycles / second), in time slots Assigned to user The rendering computing resource ratio is ,satisfy:

[0201]

[0202] Then the user The actual MEC computing resources obtained are:

[0203]

[0204] Based on this, the user In the time slot The rendering computation latency is:

[0205]

[0206] Based on the communication model, the uplink and downlink transmission delays are as follows:

[0207]

[0208] In summary, users In the time slot The total end-to-end latency from acquiring current environmental data to receiving the rendering result for this time slot can be expressed as:

[0209]

[0210] 10. Delay Model;

[0211] In multi-user AR scenarios, the Quality of Experience (QoE) is not only related to the absolute latency of a single user, but also closely related to the latency differences between different users. If there are significant time differences in the state updates of different users when observing the same virtual object, it will lead to inconsistencies in position and posture perception, resulting in a disjointed interactive experience.

[0212] Interactions between different users are typically not pairwise. To accurately characterize users' consistency needs when sharing virtual objects, this section constructs a consistency index based on an interaction weight matrix. The interaction matrix between users is as follows:

[0213]

[0214] when Time indicates user and There is a need for interaction.

[0215] Therefore, QoE and consistency metrics should be based on interaction subgraphs. ,in:

[0216]

[0217] Further define the set of users participating in the interaction as:

[0218]

[0219] Because this invention employs a real-time processing mechanism, the user In the time slot The interaction delay can be approximated as the end-to-end latency:

[0220]

[0221] In the time slot The average latency of the interactive user set is:

[0222]

[0223] Consistency should primarily reflect the latency differences between user pairs that have an interactive relationship:

[0224]

[0225] when Smaller When the system is smaller, it not only has good real-time performance but also provides greater consistency and immersion for multiple users. These two metrics will be considered comprehensively in subsequent optimization objectives.

[0226] XI. Problem Construction for Resource Allocation Optimization;

[0227] Under the above system model, consider each time slot Internally, based on the current channel status and task load, the uplink bandwidth allocation ratio, MEC computing resource allocation ratio, and downlink bandwidth allocation ratio for each user are jointly determined to optimize multi-user end-to-end latency performance.

[0228] Assume the total uplink bandwidth of the system is Total downlink bandwidth is Total computing power is In the time slot Within, each is represented by an allocation to a user. The ratio of uplink bandwidth, MEC computing resources, and downlink bandwidth satisfies:

[0229]

[0230] And nonnegativity constraints:

[0231]

[0232] The corresponding physical resources are:

[0233]

[0234] Given a bandwidth allocation ratio, the equivalent transmission rate can be expressed as:

[0235]

[0236] in, , The effective rate per unit bandwidth.

[0237] During system operation, while satisfying the latency constraints for each user, the goal is to minimize the average and variance of the end-to-end latency for multiple users. The following long-term average optimization objective can be constructed:

[0238]

[0239] in This is a weighting coefficient used to balance overall real-time performance and consistency.

[0240] To simultaneously ensure both the overall real-time performance of the system and the consistency of multi-user interaction, in each time slot Internally, it is necessary to jointly allocate uplink bandwidth, MEC computing resources, and downlink bandwidth. This optimization problem can be described as:

[0241]

[0242] Where C1 represents the uplink bandwidth resource constraint, indicating that in any time slot... Within a given time slot, the sum of the uplink bandwidth allocated to all users cannot exceed the total available uplink bandwidth of the system; C2 is the MEC computing resource constraint, meaning that within any given time slot, the sum of the MEC computing resource proportions occupied by all users cannot exceed the total computing capacity of the MEC server; C3 is the downlink bandwidth resource constraint, meaning that within any given time slot... Within this time slot, the sum of the downlink bandwidth proportions allocated to all users cannot exceed the total available downlink bandwidth of the system; C4 is a resource allocation feasibility constraint, indicating that the uplink bandwidth proportion, computing resource proportion, and downlink bandwidth proportion obtained by each user in each time slot are all non-negative numbers, ensuring that the resource allocation has physical meaning; C5 is a latency constraint, indicating that the end-to-end latency of each user in any time slot must not exceed the preset threshold D. th This is used to indicate the user's minimum latency requirement.

[0243] 12. Transformation of resource allocation optimization issues;

[0244] From the perspective of system operation mechanism, resource allocation decisions are made step-by-step within discrete time slots. The decision in each time slot not only affects the current end-to-end latency performance but may also indirectly affect the system state in subsequent time slots. Therefore, this section equivalently transforms the above optimization problem into a Markov decision process, represented as a quintuple. The resource allocation process based on TD3 is as follows: Figure 2 As shown.

[0245] (1) State space ;

[0246] During decision-making time slots Internal environmental conditions It should be able to represent the current channel conditions and task load. Combining the above communication and computation models, the state vector can be designed as follows:

[0247]

[0248] in, Representing users respectively In the time slot The uplink and downlink equivalent channel gain; Representing users respectively The amount of data uploaded and the amount of data rendered in the current time slot.

[0249] (2) Action space ;

[0250] This invention only considers resource allocation for each user. In the time slot Three continuous control variables are defined internally: , and .

[0251] Representing users respectively The obtained uplink bandwidth allocation weight, MEC rendering computing resource allocation weight, and downlink bandwidth allocation weight are used by the agent in the time slot. The combined output action is:

[0252]

[0253] To prevent individual users from receiving excessively low resources under extreme circumstances, which could affect system stability, this chapter introduces a basic share mechanism in the resource allocation process. Let the basic share parameter be... The actual allocation of the three types of resources is as follows:

[0254]

[0255]

[0256]

[0257] This refers to the total uplink and downlink bandwidth of the system. This represents the total computing power of the MEC. It is the basic share coefficient.

[0258] (3) Reward function ;

[0259] In the time slot Within this framework, the average end-to-end latency for each user can be calculated based on the resource allocation results:

[0260]

[0261] The time delay variance is:

[0262]

[0263] The instant reward is:

[0264]

[0265] in For adjustable weights, Encourage agents to reduce overall system latency and latency differences in each time slot.

[0266] (4) State transition ;

[0267] The system updates the channel state at the end of each time slot, reads the task load for the next time slot, and enters the next state. .

[0268] The state transition satisfies the Markov property, that is:

[0269]

[0270] Within the reinforcement learning framework, the agent's goal is not only to optimize the immediate reward of a single time slot, but also to maximize the cumulative discounted reward:

[0271]

[0272] in, It consists of average latency, latency variance, and consistency metrics. This is the discount factor. When As the value approaches 1, the strategy focuses more on long-term latency levels and their fluctuations, which helps maintain stable overall performance; when... When the latency is smaller, the focus is more on optimizing the real-time latency of the current time slot.

[0273] Thirteen, Design of Optimization Algorithm Based on TD3;

[0274] The core idea of ​​the TD3 algorithm is to introduce a deep neural network to approximate the policy function and action value function based on deterministic policy gradient theory, thereby learning a near-optimal continuous control policy in a high-dimensional, nonlinear environment. In a continuous action space, the algorithm iteratively approximates the optimal resource allocation policy using an Actor-Critic framework. Its specific solution process is as follows.

[0275] (1) Network initialization;

[0276] Policy Network : Receive current state Output deterministic actions.

[0277] Value Network : Estimated state Take action Corresponding action value This is used to evaluate the merits of the current strategy.

[0278] Target Policy Network and target value network The structure is the same as that of the Actor and Critic networks. It maintains similarity to the main network through soft updates, which is used to stabilize the calculation of target values ​​and reduce training oscillations.

[0279] (2) Interaction between the intelligent agent and the environment;

[0280] In each training round, the agent interacts with the system environment:

[0281] 1. In time slots Inside, based on the current state The policy network outputs continuous actions:

[0282]

[0283] 2. The environment normalizes the allocation of uplink bandwidth, MEC computing resources, and downlink bandwidth based on the action vector;

[0284] 3. Calculate the end-to-end latency for each user terminal based on the communication and computation models;

[0285] 4. Constructing instant rewards based on average latency, latency variance, and consistency metrics. ;

[0286] 5. The environment transitions to the next state. ;

[0287] 6. [The following is a list of items / items] Store in the experience replay pool.

[0288] This process is executed cyclically in each time slot to collect training samples.

[0289] (3) Critic network update;

[0290] Mini-batch samples are randomly sampled from the experience replay pool to update the dual critic network.

[0291] First, construct the target action:

[0292]

[0293] in The clipped Gaussian noise is used for target policy smoothing.

[0294] Then construct the target Q value:

[0295]

[0296] By minimizing the mean squared error loss function, the parameters of the two critic networks are updated separately to approximate the target value. A dual-criterion minimum-value mechanism is employed to reduce the Q-value overestimation problem and improve training stability.

[0297] (4) Delayed updates in the Actor network;

[0298] After the critic network completes several updates, the policy network is updated once more. The goal of the policy update is to maximize the critic network's evaluation of the current policy, i.e.:

[0299]

[0300] The Actor network parameters are updated via policy gradient ascent, enabling its output resource allocation scheme to achieve higher long-term cumulative rewards. The delayed update mechanism avoids premature policy adjustments before the value estimation has fully converged, thereby improving algorithm stability.

[0301] (5) Target network soft update;

[0302] To maintain the stability of the training process, a soft update mechanism is used for the target network parameters:

[0303]

[0304] in, This is the soft update coefficient.

[0305] By repeatedly interacting with the environment in consecutive time slots, the TD3 agent gradually learns how to adaptively adjust the resource allocation ratio under different channel conditions and task loads, thereby optimizing the average and variance of multi-user end-to-end latency and improving the overall real-time performance and multi-user behavior consistency of the system.

[0306] Based on the above introduction to the MDP problem and the TD3 algorithm, this paper proposes a resource allocation algorithm based on TD3, as shown in Table 2.

[0307] Table 2. Multi-user AR resource allocation algorithm based on TD3

[0308]

[0309] XIV. Parameter Settings for the Data Unloading Algorithm Experiment

[0310] In this experiment, we selected a subset of indoor environment sequences from the Nymeria dataset and used them as the data set for the experiment. For these sequences, only the 3D coordinate data acquired by the motion capture suit was extracted. The system environment parameters for the simulation are set as follows: Number of AR devices... The simulation used 991 time frames. The adjacency matrix was dynamically generated based on a 4-meter distance threshold. The velocity similarity threshold was 0.7, the velocity magnitude difference threshold was 0.5 meters / frame, and the relative motion trend threshold was 0.1 meters / frame. Other network parameter settings are shown in Table 3, and Table 4 shows the relevant settings for the game theory model. Furthermore, the simulation was run in a Python 3.9.0 environment, and the neural network model was built using TensorFlow 2.10.0.

[0311] Table 3 Network Model Parameters

[0312]

[0313] Table 4 Game Theory Model Parameters

[0314]

[0315] XV. Analysis of Simulation Results of Data Unloading Algorithm

[0316] To verify the advantages of the proposed Stackelberg game-based data offloading strategy, four typical data offloading and rendering acquisition schemes were selected as baseline comparisons. The simulation environment parameters of each comparison scheme were kept consistent with those of this scheme.

[0317] Comparison with Solution 1: No-Interaction Solution. This solution does not consider the dynamic interaction relationships between users during the decision-making process. Resolver users only upload and render data according to fixed rules to analyze the impact of the interaction modeling mechanism on system performance.

[0318] Comparison Option 2: Fixed-Price. In this option, the price parameters related to the host user and the server remain fixed and are not dynamically adjusted according to the system status. This is used to compare the effect of adaptive pricing mechanisms on utility optimization and rendering path selection.

[0319] Comparison Solution 3: Distributed across all servers. In this solution, all users directly obtain rendering results from the MEC server, which is used to evaluate the performance advantages of the proposed solution in reducing the downlink rendering load of MEC.

[0320] Comparison Scheme 4: Uniform Uploading. In this scheme, all users adopt a uniform upload method, without differentiating upload strategies between Host users and Resolver users, to evaluate the effectiveness of the proposed scheme in reducing the total upload overhead of the system.

[0321] Figure 3This study demonstrates the changes in host user upload utility across different time frames using three methods: Game Equilibrium, No-Interaction, and Fixed-Price. Experimental results show that the upload utility of the Proposed method remains relatively stable throughout the simulation, with an average value of approximately 3.20 and a standard deviation of approximately 0.05. The difference between its peak and trough values ​​does not exceed 0.10, and the curve exhibits small fluctuations without a trend of cumulative amplification over time, indicating good temporal consistency in its upload decisions. In contrast, the upload utility of the No-Interaction and Fixed-Price methods shows significant oscillations over time, reflecting their high sensitivity to changes in the interaction state. The upload decisions fail to fully incorporate the dynamic changes in user interaction relationships, making the strategies highly sensitive to disturbances in the interaction state. Furthermore, the pricing mechanism lacks the feedback capability to adaptively adjust with the system state; the fixed price easily amplifies the adjustment magnitude of the upload decision when the interaction structure changes, resulting in significant peak-trough oscillations in the time series. The above results indicate that the stability advantage of this scheme mainly stems from the constraint effect of dynamic upload boundary control and auxiliary upload mechanism on extreme decisions. It can effectively balance the upload benefits and costs of host users in different interactive environments, thereby improving the stability of upload utility.

[0322] Figure 4 The results compare the MEC downlink rendering volume over time frames during simulations for the Game Equilibrium, No-Interaction, Fixed-Price, and All-Server methods. The figures show fluctuations in MEC downlink rendering volume across all four schemes. This is primarily because user rendering requests differ across time frames, resulting in a dynamic variation in the amount of rendering data required for each frame. Therefore, the differences in overall downlink rendering data volume across different schemes should be emphasized. The Game Equilibrium method exhibits a relatively low overall MEC downlink rendering volume, indicating a lower dependence on server rendering resources. In contrast, the No-Interaction and Fixed-Price methods maintain consistently high downlink rendering volumes, suggesting that in the absence of effective path selection and price adjustment mechanisms, rendering requests are more likely to concentrate on the MEC side. The All-Server scheme, where all users obtain rendering results through the MEC server, results in a significantly higher downlink rendering volume than the other three schemes. The results show that this solution can guide more rendering data to be forwarded through the Host user path while meeting rendering requirements, thereby effectively reducing the amount of data sent from the MEC side and alleviating the server load pressure.

[0323] Figure 5 This paper presents a comparison of the Proposed method and the Uniform Upload method in terms of total upload volume. Experimental results show that the Uniform Upload method maintains a high total upload volume throughout the simulation, remaining consistently high. In contrast, the Proposed method has a significantly lower total upload volume, indicating that it effectively suppresses unnecessary upload overhead while meeting system requirements. This result demonstrates that the Proposed method, by combining Host user decision-making with Resolver user-assisted upload strategies, achieves more refined upload resource allocation, thereby reducing system communication burden while maintaining relatively stable operation.

[0324] 4.3.1 Parameter Settings

[0325] The system environment parameters for the simulation are set as follows: the network scenario includes one MEC server and 10 AR users; the simulation area is 30m × 30m; the total uplink and downlink bandwidth is 15MHz; and other parameters are shown in Table 5. Furthermore, this section uses Python 3.7 to build a multi-user interactive AR system environment and PyTorch 1.4.0 to build the neural network model. TD3 parameter settings are shown in Table 6.

[0326] Table 5 Environmental Parameters

[0327]

[0328] Table 6 TD3 Parameter Settings

[0329]

[0330] 4.3.2 Simulation Result Analysis

[0331] To verify the effectiveness of the proposed Consistency Driven (CD) scheme in a multi-user augmented reality system, two typical latency optimization schemes were selected as baselines for comparison:

[0332] Comparison Scheme 1: Resource allocation strategy based on Max-Latency Minimization (MLM). Strategy optimization is achieved by designing a reward mechanism with the objective function of minimizing the maximum user latency of the system. Simulation environment parameter settings are consistent with the proposed scheme.

[0333] Comparison Scheme 2: A task scheduling strategy based on Max-Latency Minimization (MLM). Strategy optimization is achieved by designing a reward mechanism with the objective function of minimizing the system's average user latency. Simulation environment parameters are set consistently with the proposed scheme.

[0334] Figure 6 The graph illustrates the reward convergence of the TD3 and DDPG algorithms during training. As can be observed, the reward value fluctuates significantly in the early stages of training, which is due to the randomness of resource allocation behavior in the policy network's exploratory phase. With increasing training rounds, the reward value gradually rises and then stabilizes within a certain range, indicating that the agent gradually learns a more stable resource allocation strategy. During the convergence phase, the fluctuation amplitude of the reward curve decreases significantly, suggesting that the dual-Critic network and delayed update mechanism improve training stability to some extent. From a convergence perspective, the TD3 algorithm can achieve relatively stable training under dynamic channel and load environments.

[0335] Figure 7 This paper compares the convergence performance of the TD3 algorithm under different learning rates. Experimental results show that the algorithm exhibits the fastest convergence speed and best performance when the learning rate is set to 0.00005; the orange curve represents the performance at a lower learning rate. The slow convergence is due to the small update step size. While a higher learning rate initially shows a very fast increase, the large step size can lead to excessively large steps in policy space exploration, resulting in multiple significant negative fluctuations, with the final convergence fluctuation slightly larger than that of the blue curve. Therefore, while a low learning rate ensures robust training, it results in an excessively slow gradient descent process, making it difficult to quickly capture features in dynamic environments. Conversely, an excessively high learning rate can easily cause system oscillations or even ineffective learning, making it difficult for the policy to stabilize near the optimal solution. Therefore, in optimizing resource allocation strategies in a multi-user AR simulation environment, the proposed algorithm must select an appropriate learning rate to ensure efficient and steady convergence to the optimal policy within a limited number of iterations.

[0336] Figure 8The average latency changes of the three algorithms under different user numbers are presented. As the number of users increases from 2 to 10, the average latency of all schemes shows a monotonically increasing trend. This is because the increase in the number of users leads to increased competition for system resources and increased complexity of collaborative processing, resulting in an increase in end-to-end latency. The comparison results show that when the user scale is small, the difference in average latency among the three algorithms is relatively small; however, as the number of users further increases, the average latency of the MLM method increases significantly faster, while the ALM method, although maintaining a low average latency under a medium user scale, also shows a significant increase under high load scenarios. Until the number of users increases to 10, the minimum latency requirement for users can no longer be met. Due to the consideration of consistency issues, this scheme makes a slight compromise on the average latency, achieving a high user experience and low latency difference in the interactive experience even with multiple users.

[0337] Figure 9 The paper demonstrates the variation of latency standard deviation for three algorithms under different user scales. Latency standard deviation is an important indicator for measuring the consistency of multi-user interactions. It can be observed that the latency standard deviation of all three algorithms increases with the number of users. This is because increased resource competition may lead to some users receiving relatively fewer resources, thus widening the latency gap between users. The MLM method shows a more significant increase in standard deviation under high user scales. This method focuses on minimizing the maximum latency, and may skew resource allocation for some users during the optimization process, resulting in an uneven overall latency distribution. The ALM method mainly focuses on average latency; although it maintains a certain balance at some scales, it struggles to suppress the widening of user differences when resources are scarce. In contrast, the proposed TD3 method maintains a lower latency standard deviation across all user scales, and its growth trend is relatively gradual. Therefore, from a consistency perspective, the TD3 scheme exhibits more stable performance in multi-user collaborative scenarios.

[0338] Figure 10The changes in average end-to-end latency difference for each algorithm under different system bandwidth conditions are shown. The overall trend indicates that as the available system bandwidth increases, the average latency of all three algorithms decreases, suggesting that the more abundant the bandwidth resources, the less resource contention pressure in the system. In low-bandwidth scenarios, due to limited communication resources, resource competition among users is intense. The MLM method may allocate resources preferentially to individual users during the optimization of maximum latency, resulting in a relatively high overall average latency. The ALM method, with average latency as its optimization objective, can compress the overall latency to some extent under low bandwidth conditions. The TD3 algorithm, with average latency and latency standard deviation as its optimization objectives, although its average latency is close to that of the ALM method under different bandwidth conditions, can adaptively adjust the allocation ratio of communication and computing resources according to real-time channel status and task load. When resources are scarce, it prioritizes the basic needs of users with high loads or poor channel conditions, achieving more reasonable resource utilization and better interaction consistency without significantly sacrificing overall performance.

[0339] Figure 11 The paper presents the changes in the standard deviation of latency for the three algorithms under different bandwidth conditions. It shows that as bandwidth increases, the standard deviation of latency for each scheme generally decreases, indicating that increased communication resources help alleviate latency differences between users caused by resource competition. However, in terms of overall performance, the TD3 scheme maintains a low and relatively stable standard deviation across the entire bandwidth range, with minimal fluctuations. In contrast, the MLM method has a significantly higher standard deviation under low bandwidth conditions, suggesting that it may sacrifice latency uniformity for some users when focusing on maximum latency optimization. While the ALM method shows some improvement under medium to high bandwidth conditions, it is still higher than TD3 overall and exhibits some fluctuations during bandwidth changes. Combined with the modeling of the consistency index in the paper, it can be seen that the TD3 scheme introduces a latency variance term into the reward function, making the resource allocation process not only focus on overall real-time performance but also consider the control of differences between users. Therefore, in an environment of dynamically changing bandwidth, this method can effectively suppress latency fluctuations between different users and maintain relatively stable consistency performance.

[0340] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for data offloading and resource allocation in a multi-user AR system under a 6G network framework, characterized in that, The method includes the following steps: Step S1: In each time slot, select one AR user from multiple AR users as the Host user, and the remaining users as Resolver users; Step S2: Calculate the interaction weights between the Host user and each Resolver user using a graph attention network; the interaction weights are used to characterize the intensity of interaction between users. Step S3: Establish a data offloading model based on Stackelberg game, construct the environmental data upload process and the rendering data acquisition process as a sequential game with a leader-follower structure, and establish a multi-party utility function that includes the Host user, Resolver user and MEC server in the upload and rendering stages. Step S4: Under the condition of satisfying the participation constraints, the two-layer game optimization problem constructed in step S3 is transformed into a single-layer optimization problem with only the Host user decision variables, and the sequential quadratic programming algorithm is used to solve it to obtain the optimal data unloading decision. The data unloading decision includes the optimal amount of data uploaded by the Host user and the rendering price, as well as the optimal amount of data uploaded by each Resolver user and the rendering data acquisition method. Step S5: Construct the resource allocation process as a Markov decision process, and define the state space, action space and reward function, wherein the reward function includes the multi-user end-to-end average latency and latency variance, and the latency variance is used to characterize the interaction consistency among multiple users. Step S6: Train the algorithm using a dual-delay deep deterministic strategy gradient algorithm to output a resource allocation strategy for the continuous action space. The resource allocation strategy includes the uplink bandwidth ratio, downlink bandwidth ratio, and edge computing resource ratio allocated to each user. Then, allocate communication and computing resources to each user according to the resource allocation strategy.

2. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The step S1, which involves selecting one AR user from multiple AR users as the Host user, specifically involves selecting the user with the most effective interaction relationships with other users as the Host user within each time slot, based on the dynamic interaction relationships between users.

3. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The specific method for calculating the interaction weights between the Host user and each Resolver user using a graph attention network in step S2 includes: Node feature vectors are constructed based on the spatiotemporal trajectory information of each user. In discrete time frames The system is represented as a set of time sequence diagrams; each user corresponds to a node in the diagram, and the node is in time frame. Its state is represented by its spatiotemporal trajectory information; Let the first Individual users in time frames The spatial location is: in, This represents the spatial location feature vector of the i-th user at time frame t; and This indicates the user's x-coordinate and y-coordinate position in the coordinate system at time t; Represents the transpose of a vector; To reflect the user's movement trend, a velocity vector obtained from the position difference between adjacent time frames is introduced: Therefore, node In time frame The eigenvectors are defined as follows: A dynamic adjacency matrix is ​​constructed based on the Euclidean distance, similarity of velocity direction, difference in velocity magnitude, and relative motion trend among users; set up Represents time frame The adjacency matrix, where the elements of the matrix are... Defined as: Where h specifically refers to the index of the user selected as the Host user; i generally refers to the index of the i-th Resolver user in the system other than the Host user. The interaction conditions should be considered from the following four aspects: Euclidean distance between users: in, This represents the Euclidean distance between the Host user and the i-th Resolver user at time frame t. This represents the spatial location vector of the Host user in time frame t; This represents the spatial location vector of the i-th Resolver user in time frame t; h specifically refers to the index of the user selected as the Host user; Representing vectors Norm; Similarity in velocity direction: in, This represents the cosine value of the motion direction between the h-th Host user and the i-th Resolver user within time slot t; and This represents the velocity vector of the Host user and the i-th Resolver user in time slot t; Difference in speed magnitude: in, This represents the rate difference between the h-th Host user and the i-th Resolver user in time slot t; Relative motion trend : If the above preset threshold constraints are met, then the user is considered... and In time frame There are interactive relationships; at the same time, in order to preserve the node's own information, set... ; Under the constraints of the dynamic adjacency matrix, the normalized attention weights between the Host user and each Resolver user are calculated through a multi-head attention mechanism and used as the interaction weights. Characteristic linear transformation: For the Each attention head first performs a linear mapping of node features: in, This represents the transformed feature vector of the i-th node under the k-th attention head; Learnable weight matrix; Attention score calculation: Based on the additive attention mechanism, nodes For nodes The attention score is defined as: in, This represents the original attention score of node i to itself under the k-th attention head in time slot t; and These are the self-attention learning vector and the neighbor attention learning vector; It is a non-linear activation function; Normalized attention weights: Use Softmax for each node Normalize the neighbors and the node itself: This weight represents the weight in the first... Under one attention head, the time frame is Time user For users The relative degree of influence; Interactive value output format: Through multi-head attention mechanisms, interaction weight values ​​in the following form can be obtained: in, Indicates the number of attention heads; This value is used to describe the intensity of subsequent dynamic user interactions.

4. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The establishment of the multi-party utility function in step S3 specifically includes: The server's upload phase utility is the total benefit to the environment data minus the payment to all uploaders; the host user's upload phase utility is the payment received from the server minus the upload cost; and the resolver user's upload phase utility is the payment received from the server minus the upload cost. The total amount of environmental data received by the server, X, represents the amount of environmental data uploaded by the Host user. and the amount of environmental data uploaded by the i-th Resolver user sum: The overall benefit of environmental data adopts a concave function with diminishing marginal returns. : ,in, This represents the weighting coefficient of environmental integrity on the benefits of scene reconstruction. The server pays all uploaders based on the unit of environmental data, with the unit payment price denoted as [missing information]. The server's utility function during the upload phase is: Among them, superscript Indicates the transmission stage; The host user's utility during the upload phase is as follows: in, This represents the cost per unit of data uploaded by a Host user. The utility of each Resolver user during the upload phase is: in, The unit cost of uploading data for Resolver users; The rendering stage utility of a server is the revenue from distributing rendering data to all users minus the generation cost. The rendering stage utility of a host user is its own rendering revenue minus the rendering price and forwarding cost paid to the server. The rendering stage utility of a resolver user is its revenue from obtaining rendering data from a host user or the server minus the corresponding cost. Suppose that within a certain time slot, the amount of rendering data obtained by the i-th Resolver user from the Host user is... The amount of rendering data obtained from the server is Then the total amount of data rendered is: This reflects the constraints that interaction relationships impose on the selection of rendering data paths; when At this time, if the Resolver user and the Host user are too far apart or their motion characteristics are completely unrelated, they do not meet the conditions for cooperation. ;when At that time, the system determines that there is a valid spatial interaction between the Resolver user and the Host user. ; (1) Rendering utility for Resolver users; The rendering gains of a Resolver user measure the economic or technical value of the data uploaded by that user in contributing to the system's scene reconstruction and rendering: ; This represents the amount of environmental awareness data uploaded by the i-th user; This represents the marginal revenue that each unit of data brings to the server. The rendering phase utility for Resolver users is: (2) The rendering effect of the Host user; Host users also need to request rendering data from their own perspective from the server, denoted as... At the same time, the Host user also needs to forward rendering data to the Resolvers users with whom they have an interaction relationship; Therefore, the total utility of the Host user during the rendering phase should be: : in, The amount of rendering data requested by the Host user from the server; The unit service price paid by the Resolver user to the Host user; Host users must pay for the server costs for their own rendering needs; The host user bears the cost of all data processing and forwarding. (3) The rendering effect of the server; The server distributes rendering data to all users, including the Host user and all Resolvers users: in, This refers to the unit operating cost of the server.

5. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, Step S4, which transforms the two-layer game optimization problem into a single-layer optimization problem, specifically includes: Given the Host user's decision variables, solve for the optimal response that maximizes the utility of each Resolver user, and obtain the analytical expressions for the optimal amount of data uploaded and the optimal amount of data acquired by the Resolver user with respect to the Host user's decision variables. Find the optimal response for the Resolver user; Resolver user utility function about It is separable, wherein: Upload stage: Rendering stage: in, It is the cost incurred by Resolver users in downloading data; In the given and Under these conditions, the optimal rendering decision for the Resolver user can be parsed as follows: when and hour: in, This represents the optimal collaborative decision. This represents the optimal demand function, where the Resolver user determines the price based on the price provided by the Host user. The optimal amount of data to acquire; The optimal direct connection decision is 0; this means that user i decides not to download data directly from the MEC server. when or hour: in, This indicates that user i directly requests services from the MEC server; Substituting the analytical expression into the total utility function of the Host user, the decision variables of the Resolver user are eliminated, resulting in a single-level optimization problem that depends only on the decision variables of the Host user; Substitute the optimal response into the Host user utility function: The above results Substitute the total utility function of the Host user: This eliminates the decision variables for the Resolver user, making the Host user's utility function depend only on... ; Transform a two-level problem into a one-level problem; After the above processing, the two-layer game optimization problem constructed in step S3 can be transformed into the following single-layer optimization problem: in, It is the upper limit of uplink bandwidth. It is the upper limit of the price.

6. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The method of solving the problem using a sequential quadratic programming algorithm specifically includes: (1) The optimization problem of maximizing the total utility of the Host users is equivalently transformed into a minimization problem; In the numerical solution process, the maximization problem is equivalently transformed into a minimization problem: (2) Construct a local quadratic programming subproblem in each iteration and update the Host user's decision variables under the conditions of satisfying the variable value range constraints and participation constraints; In the single-level optimization problem for Host users, the following constraints must be satisfied simultaneously: Variable value range constraints: Participation constraints: Given the objective function and constraint set, the SLSQP algorithm is used to iteratively update the decision variables of the host user. In each iteration, the algorithm constructs a local quadratic programming subproblem and updates the decision variables while satisfying the constraints. ; (3) Solve the problem from multiple initial feasible points, select the feasible solution with the largest total utility of the Host user as the optimal solution for this time slot, and output the optimal amount of data uploaded by the Host user and the rendering price. (4) Based on the optimal solution, combined with the optimal response parsing expression of each Resolver user, the optimal upload data volume and rendering data acquisition method of each Resolver user are calculated.

7. The method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The participation constraints include: participation constraints for each Resolver user, ensuring that their utility in participating in system collaboration is not less than zero; and participation constraints for the server, ensuring that its total utility in the data upload and rendering service process is not less than zero.

8. A method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The definition of the state space, action space, and reward function in step S5 specifically includes: (1) State space S; During decision-making time slots Internal environmental conditions It should be able to represent the current channel conditions and task load; the state vector can be designed as follows: in, Representing users respectively In the time slot The uplink and downlink equivalent channel gain; Representing users respectively The amount of data uploaded and the amount of data rendered in the current time slot; (2) Action space A; For each user In the time slot Three continuous control variables are defined internally: , and , respectively representing users The obtained uplink bandwidth allocation weight, MEC rendering computing resource allocation weight, and downlink bandwidth allocation weight; the agent in the time slot The combined output action is: To prevent individual users from receiving excessively low resources under extreme circumstances, thus affecting system stability, let the base ratio parameter be... The actual allocation of the three types of resources is as follows: in, This refers to the total uplink and downlink bandwidth of the system. This represents the uplink transmission weight of the i-th user at time t; It is the user's calculated weight; User's downlink transmission weight; This represents the total computing power of the MEC. The basic share coefficient; N is the total number of AR users in the system; (3) Reward function R; In the time slot Within this framework, the average end-to-end latency for each user can be calculated based on the resource allocation results: in, Represents a set of users with interactive relationships; This represents the total number of users in the set; Let represent the individual end-to-end latency of the i-th user at time t; The time delay variance is: The instant reward is as follows: in, The weights for the average end-to-end delay and delay variance, Encourage agents to reduce overall system latency and latency differences in each time slot; (4) State transition P; The system updates the channel state at the end of each time slot, reads the task load for the next time slot, and enters the next state. The state transition satisfies the Markov property, that is: Within the reinforcement learning framework, the agent's goal is not only to optimize the immediate reward of a single time slot, but also to maximize the cumulative discounted reward: in, It consists of average latency, latency variance, and consistency metrics. As a discount factor; when As the value approaches 1, the strategy focuses more on long-term latency levels and their fluctuations, which helps maintain stable overall performance; when... When the latency is smaller, the focus is more on optimizing the real-time latency of the current time slot.

9. A method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, The end-to-end latency includes uplink transmission latency, edge computing rendering latency, and downlink transmission latency; wherein, the rendering latency is calculated based on the amount of rendering data within each user and the proportion of allocated computing resources, and the amount of rendering data within each user and the amount of rendering data sent down are correlated through a proportional coefficient.

10. A method for data offloading and resource allocation in a multi-user AR system under a 6G network framework according to claim 1, characterized in that, Step S6, which describes training using a dual-delay deep deterministic strategy gradient algorithm, specifically includes: (1) Initialize the policy network, the dual-value network, and the corresponding target network; Policy Network : Receive current state Output deterministic actions; Value Network : Estimated state Take action Corresponding action value This is used to evaluate the merits of the current strategy; Target Policy Network and target value network The structure is the same as that of the Actor and Critic networks. It maintains similarity to the main network through soft updates to stabilize the target value calculation and reduce training oscillations. (2) In each time slot, the policy network outputs resource allocation actions according to the current state, the environment calculates the end-to-end latency of each user and feeds back the instant reward, and stores the experience tuple into the experience replay pool. In the time slot Inside, based on the current state The policy network outputs continuous actions: The environment normalizes the allocation of uplink bandwidth, MEC computing resources, and downlink bandwidth based on the action vector; Calculate the end-to-end latency for each user terminal; Construct an instant reward based on average latency, latency variance, and consistency metrics. ; The environment transitions to the next state. ; Will Store in the experience replay pool; (3) Randomly sample a small batch of samples from the experience replay pool, construct the target Q value by taking the minimum value of the dual value network, update the dual value network; delay the update of the policy network, and perform soft update on the target network; Mini-batch samples are randomly sampled from the experience replay pool to update the dual critic network; First, construct the target action: in, It is the smoothed target action; The state in the next moment; Target policy network The clipped Gaussian noise is used for target policy smoothing; Then construct the target Q value: in, It's an instant reward. It is a discount factor. and Two independent networks of target commentators; Actor network delayed updates; After the critic network completes several updates, the policy network is updated once more; the goal of the policy update is to maximize the critic network's evaluation of the current policy, i.e.: Target network soft update; in, It is the target performance function. It is the policy gradient operator. It is the mathematical expectation. It was the first commentator network. It is a policy network. It is the action gradient; To maintain the stability of the training process, a soft update mechanism is used for the target network parameters: in, This is a soft update coefficient; In step S6, when allocating communication and computing resources to each user, a basic share mechanism is introduced: the total uplink bandwidth, total downlink bandwidth, and total MEC computing capacity of the system are multiplied by the basic share coefficient to obtain the minimum guaranteed resources for each user, and the remaining resources are allocated according to the resource allocation strategy.

Images