Resource Allocation Method for Dual-Time-Scale MuAR Systems within a 6G Network Framework

Abstract

This invention proposes a dual-timescale MuAR system resource allocation method within a 6G network framework. Based on this framework, it jointly optimizes long-term computational resource allocation and short-term channel and subframe configuration, minimizing the positional offset of virtual objects across different display devices while ensuring end-to-end latency below 16 milliseconds. To overcome the limitations of traditional MA-DRL methods, this invention employs a completely decentralized learning strategy, where each agent autonomously trains a localized model, enabling rapid adaptation to instantaneous environmental changes while maintaining decision-making responsiveness. A graph attention network is integrated into this decentralized learning framework to explicitly simulate the hidden bidirectional interference dependencies between agents, promoting coordinated optimization. A Bayesian optimization-enhanced deep deterministic policy gradient algorithm (DDPG) is developed, utilizing the joint prior distribution of computational requirements and resource allocation decisions to guide the exploration, thereby improving the learning efficiency of DDPG.

Description

Technical Field

[0001] This invention belongs to the field of 6G mobile communication technology, specifically a resource allocation method for a dual-timescale mobile multi-user augmented reality system under a 6G network framework. Background Technology

[0002] Mobile Augmented Reality (AR) overlays virtual objects onto the physical environment using wireless optical or video see-through displays, extending the perception of the virtual world into the real world. Multi-user Augmented Reality (MuAR) systems revolutionize collaborative experiences by facilitating real-time interaction with virtual objects through a shared augmented environment. Current 5G architecture's resource allocation mechanisms cannot support MuAR real-time interaction scenarios, particularly in addressing issues such as significant virtual object positional offsets, low efficiency of asymmetric network traffic transmission, high overhead of centralized scheduling synchronization, and insufficient interference coordination capabilities. Therefore, next-generation mobile communication networks are needed to enable fast and accurate data sharing among AR users. MuAR systems also pose challenges to the underlying network: strict control of rendering latency (typically below 20 milliseconds per frame) is required to prevent visual artifacts, while maintaining millisecond-level consistency in end-to-end latency across all collaborating users. Since time differences between users can cause virtual object positional offsets, disrupting collaborative spatial perception and shared environment consistency, resolving these two challenges is crucial for maintaining spatial alignment accuracy.

[0003] In the MuAR system, end-to-end latency primarily occurs during 3D registration and rendering. Offloading these tasks to a Mobile Edge Computing (MEC) server effectively alleviates the computational bottleneck. However, asymmetric data traffic arises during the upload of raw video captured by the terminal device to the MEC server and the download of rendered video to the terminal device. Adaptively configuring the uplink / downlink channel direction using Dynamic Time-Division Duplexing (DTDD) technology can effectively address this asymmetric traffic issue. Furthermore, any modifications initiated by a user to a shared virtual object must be propagated to the views of all participants in real time, leading to positional offset issues caused by cross-device state updates.

[0004] Multi-Agent Deep Reinforcement Learning (MA-DRL) enables autonomous decision-making under time-varying network conditions and is often applied to resource allocation problems. However, traditional MA-DRL methods typically employ a centralized training with decentralized execution (CTDE) approach. This means that agents share historical states and actions during the centralized training phase, and then execute the pre-trained network based on local environmental observations. This approach has significant limitations. 1) In dynamic environments where real-time resource allocation occurs across multiple base stations (BSs), frequent policy synchronization incurs excessive communication overhead and latency penalties. 2) As the number of BSs increases, computational complexity grows exponentially, leading to scalability issues. 3) During centralized training, the model over-relies on global information, but in distributed execution, each BS operates solely based on local observations, lacking cross-cell state awareness. This results in a mismatch between training and deployment, weakening policy effectiveness. 4) Centralized training may implicitly assume a stable environment, but in distributed execution, each agent's local environment may constantly change due to policy variations of other agents, further exacerbating model performance degradation. These limitations prevent the MA-DRL method from being directly applied to the multidimensional resource joint optimization problem in DTDD-enhanced MuAR systems. Summary of the Invention

[0005] To address the challenge of minimizing location offset rate in mobile MuAR systems within a 6G network framework while ensuring latency requirements, this invention proposes a Two-timescale Distributed Resource Allocation (TDRA) framework to jointly optimize long-term Computing Resource Allocation (CRA) and short-term Channel Assignment and Subframe Configuration (CASC).

[0006] To achieve the above objectives, the present invention adopts the following solution:

[0007] The deep coupling and differing time-varying characteristics of computing resources with channel and subframe resources constitute a cross-temporal resource management problem. From a practical perspective, computing resource allocation is a large-scale decision-making process because the arrival times of computing tasks are relatively stable, with intervals measured in seconds. On the other hand, channel conditions and interference fluctuate at smaller time scales (e.g., milliseconds). To address this temporal decoupling problem, the resource allocation framework proposed in this invention is a dual-time-scale one. Specifically, the large-scale window size is based on the AR device's video refresh cycle (16 milliseconds), while the small-scale interval is based on the duration of dynamic TDD subframes (1 millisecond). This hierarchical architecture enables differentiated synchronous management of heterogeneous resources across different time domains.

[0008] To overcome the limitations of traditional MA-DRL methods, this invention employs a completely decentralized learning strategy, where each agent autonomously trains a localized model, enabling rapid adaptation to transient environmental changes while maintaining decision-making responsiveness. However, due to the limitation of partial observation, it has a fundamental limitation in capturing global features. This limitation hinders effective coordination when resolving complex dependencies and may lead to decision conflicts between agents. To address this issue, this invention integrates a Graph Attention Network (GAT) into the completely decentralized learning strategy, explicitly simulating hidden interfering dependencies between agents to promote coordinated optimization. Furthermore, applying MA-DRL methods to handle multidimensional resource allocation problems with large state and action spaces results in computational intensity and inefficiency. To address this, this invention develops the Bayesian Optimization-Enhanced Deep Deterministic Policy Gradient (BO-DDPG) algorithm, which utilizes the joint prior distribution of computational requirements and computational resource allocation decisions to guide exploration, thereby improving the learning efficiency of DDPG.

[0009] The specific steps of the method of the present invention are as follows:

[0010] S1: Define the video refresh cycle of the AR device as a large time scale. Within each large time scale, apply the Bayesian optimization-enhanced depth deterministic policy gradient algorithm to make computational resource allocation decisions and minimize the computational latency of 3D registration and rendering tasks.

[0011] S2: Define the duration of a subframe as a small time scale. Within each small time scale, apply a bidirectional graph attention network and a deep Q-network to make channel allocation and subframe configuration decisions in order to minimize the position offset rate of virtual objects displayed on different display devices.

[0012] Specifically, in S1, the deep deterministic policy gradient algorithm enhanced by Bayesian optimization includes:

[0013] S1-1: Statistical modeling of the conditional probability of events based on historical samples;

[0014] S1-2: Optimize the utility function related to the occurrence of the event. Based on the given posterior distribution, output the optimal computational resource allocation decision using gradient descent.

[0015] S1-3: Training a DNN is a function approximator that obtains the optimal computational resource allocation decision based on the current strategy. Based on the Bellman equation, respectively... and Status - Action Value and

[0016] S1-4: By comparing two state-action values and Select the state-action value with the highest value as the final output action.

[0017] Specifically, S1-1: The process of statistically modeling the conditional probability of events based on historical samples includes:

[0018] and

[0019] Bayesian optimization: computational resource requirements are defined as triples. in and These represent the number of CPU cycles required to perform 3D registration, host rendering, and resolver rendering, respectively. Host rendering refers to the rendering of virtual objects initiated by the AR user when running as a host, while resolver rendering refers to the rendering of virtual objects initiated by other hosts when running as a resolver.

[0020] The distribution of the number of CPU cycles required for 3D registration is simulated as a Pareto distribution, expressed as:

[0021] Where a, b, and c represent shape, scale, and threshold, respectively;

[0022] Therefore, its cumulative distribution function can be calculated in the following way:

[0023]

[0024] The number of CPU cycles required for host rendering is represented by a truncated log-normal distribution.

[0025] Where μ represents the mean and δ represents the standard deviation;

[0026] Therefore, its cumulative distribution function can be calculated in the following way:

[0027]

[0028] The distribution of the total number of CPU cycles required for parser rendering is simulated as an exponential distribution, as shown below:

[0029] Where λ represents the interaction strength;

[0030] Therefore, its cumulative distribution function can be calculated in the following way:

[0031]

[0032] Specifically, S1-2: The process of optimizing the utility function related to the occurrence of the event, and outputting the optimal computational resource allocation decision using gradient descent based on a given posterior distribution, includes:

[0033] Define a function that maps CRA variables to compute resource requirements, as shown below:

[0034] Among them Ψ n，m It is a CRA variable, i.e., BSB. n For AR users U m The decision variable for allocated computational resources, ∈, is the error term, which is treated as Gaussian noise with a mean of 0;

[0035] According to Bayes' theorem, the posterior distribution With respect to the prior distribution and likelihood function of g(Ψ) Related, as shown below:

[0036] in It is a historical sample set;

[0037] Define a utility function to describe the expected improvement in the function value g(Ψ), expressed as: g * (Ψ) represents the past maximum function value;

[0038] The gradient descent algorithm is used to solve the utility function maximization problem. By maximizing the utility function, the optimal CRA decision for the next time slot is found.

[0039] Specifically, the trained DNN in S1-3 is a function approximator used to obtain the optimal computational resource allocation decision based on the current strategy. The process includes:

[0040] The problem of minimizing computational latency is formulated as a Markov decision process, consisting of a state space, an action space, and a reward function.

[0041] Treating the base station (BS) as an intelligent agent, B n Elements in the state space Indicates AR user U m The computational resource requirements for the three types of tasks; AR user U m Action space includes in and These represent the proportions of computational resources allocated to perform 3D registration, host rendering, and resolver rendering, respectively; the reward function is defined as the total negative value, i.e. in and These represent the computational latency for performing 3D registration, host rendering, and resolver rendering, respectively.

[0042] The framework consists of a training DNN and a target DNN; the training DNN is a function approximator used to obtain the optimal action a based on the current policy π(θ). * (t) and the state-action value of the action. As shown below:

[0043]

[0044] Target DNNs are used to evaluate the best action based on the following factors:

[0045]

[0046] Where θ and ω are the weight parameters of the training DNN and the target DNN, respectively; the weight parameters of the training DNN are updated through backpropagation, and its error loss is:

[0047]

[0048] Where, N R It is the size of the random sampling batch of the replay memory; the target DNN copies the weight parameters of the trained DNN at fixed steps.

[0049] Specifically, in S2, the process of applying a bidirectional graph attention network and a deep Q-network to make channel allocation and subframe configuration decisions includes: modeling the mobile multi-user augmented reality system as a network consisting of N... B The directed graph composed of subgraphs captures the bidirectional interference dependencies between agents through a bidirectional graph attention network; then, a deep Q-network generates a joint optimization decision for channel allocation and subframe configuration to minimize the average position offset rate.

[0050] Specifically, the mobile multi-user augmented reality system is modeled as a system consisting of N B The process of capturing bidirectional interference dependencies between agents in a directed graph composed of subgraphs using a bidirectional graph attention network includes:

[0051] Model the mobile multi-user augmented reality system as a system consisting of N B The process of capturing bidirectional interference dependencies between agents in a directed graph composed of subgraphs using a bidirectional graph attention network includes:

[0052] Model the mobile multi-user augmented reality system as a system consisting of N B A directed graph composed of subgraphs, using This indicates that the nth subgraph is composed of... It means that, among them Represents AR users, ε:={e m,m′} represents the disturbance that node m receives from node m′;

[0053] The interference received by node m from other nodes and the interference caused by node m to other nodes are respectively input into the MLP layer. The MLP layer extracts the latent features of each node and converts them into high-dimensional features, represented as:

[0054]

[0055] in and These are the weight parameters that need to be trained in an MLP, where σ(·) represents the ReLU activation function, and I n，m and These represent the interference that node m receives from other nodes and the interference that node m causes to other nodes, respectively. The coefficients between nodes can be implemented through multi-head attention, as shown below:

[0056] α m,m′ =ATT(Wh m ,Wh m′ );

[0057] Where ATT(·) is the self-attention mechanism, W is the corresponding weight parameter, and this formula represents the importance of node m′ to node m;

[0058] Masking and normalized attention coefficients can be implemented in the following ways:

[0059]

[0060] Where softmax(·) is the softmax function and LeakyReLu(·) is the LeakyReLu function, the state embedding is obtained by concatenating Q independent attention mechanisms, and its value is:

[0061]

[0062] Where ||(·) represents a join operation, and Q represents the number of attention heads;

[0063] The state embedding is passed to the fully connected layer for feature fusion, and its calculation formula is as follows:

[0064]

[0065] Specifically, the process of generating joint optimization decisions for channel allocation and subframe configuration through a deep Q-network includes:

[0066] The state space of partially observable Markov decisions is used express;

[0067] In a deep Q-network, the trained DNN updates its weight parameters by minimizing a mean squared error loss function, which is:

[0068]

[0069] in, Given a mini-batch size, the weight vector μ(t+1) can be solved using gradient descent, i.e.:

[0070]

[0071] Where Γ represents the learning rate.

[0072] Beneficial effects: Compared with the prior art, the present invention can achieve at least the following technical effects:

[0073] 1. This invention proposes a dual-time-scale resource allocation framework to jointly optimize long-term computing resource allocation and short-term channel and subframe configuration, thereby minimizing the position offset rate while ensuring latency requirements.

[0074] 2. This invention adopts a completely decentralized learning strategy, in which each agent autonomously trains a localized model, thereby enabling it to quickly adapt to instantaneous environmental changes while maintaining decision-making responsiveness. The Graph Attention Network (GAT) is integrated into the completely decentralized learning strategy to explicitly simulate the hidden interference dependencies between agents, thereby promoting coordinated optimization.

[0075] 3. This invention develops the Bayesian Optimization-Enhanced Deep Deterministic Policy Gradient (BO-DDPG) algorithm, which uses the joint prior distribution of computational requirements and computational resource allocation decisions to guide the exploration, thereby improving the learning efficiency of DDPG. Attached Figure Description

[0076] Figure 1 : Schematic diagram of VST AR system;

[0077] Figure 2 MuAR System Workflow Diagram;

[0078] Figure 3 :D-TDD frame structure diagram;

[0079] Figure 4 :BiGATDQN framework;

[0080] Figure 5 TD-RA working framework;

[0081] Figure 6 Convergence performance graphs of BODDPG, DDPG, and BO algorithms for the CRA problem;

[0082] Figure 7 : Relationship between latency and number of AR users for BO-DDPG, DDPG and BO algorithms;

[0083] Figure 8 : Relationship between latency and computational capability of BO-DDPG, DDPG and BO algorithms;

[0084] Figure 9 : Relationship between latency and interaction strength of BO-DDPG, DDPG and BO algorithms;

[0085] Figure 10 Convergence performance of different algorithms in CASC sub-schemes;

[0086] Figure 11 The relationship between average offset rate and the number of AR users in different algorithms;

[0087] Figure 12 The relationship between average offset rate and bandwidth in different algorithms;

[0088] Figure 13 The relationship between average offset rate and cross strength in different algorithms;

[0089] Figure 14The variation of location offset rate with the number of AR users under different resource allocation schemes (i.e., TD-RA, TD-RA woCRA, TD-RA woCA, and TD-RAw.o.SC);

[0090] Figure 15 The relationship between the position offset rate and the computational energy for different resource allocation schemes (i.e., TD-RA, TD-RA woCRA, TD-RA woCA, and TD-RAw.o.SC).

[0091] Figure 16 The relationship between location offset rate and bandwidth for different resource allocation schemes (i.e., TD-RA, TD-RA woCRA, TD-RA woCA, and TD-RAw.o.SC). Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] Based on visual presentation methods, AR systems can be divided into two types: Optical See-Through (OST) and Video See-Through (VST). This invention studies the VST AR system, which works by using a camera to capture a real-world scene, combining it with virtual objects, and displaying it on a head-mounted display (HMD). The hardware components and workflow of the VST AR system are as follows: Figure 1 As shown. Its specific workflow is (1) First, the system captures video through the camera, and then generates a three-dimensional point cloud based on the visual data as a digital representation of the physical environment. (2) Perform a three-dimensional registration operation to accurately overlay the virtual objects generated by the server onto the real environment. The three-dimensional registration operation includes feature extraction, feature matching and pose estimation. Pose estimation is to reposition and describe the objects in the coordinate system from the observer's perspective, determine which virtual objects the user can see, and define the precise spatial relationship between them. (3) In order to obtain a realistic visual experience, the physical environment and virtual objects should share lighting information. The rendering process uses camera lens and position data to estimate the ambient lighting of the scene, and then performs virtual reality synthesis, including shadow generation, environmental occlusion and reflection / refraction simulation. (4) Finally, the HMD is used to display the image to the AR user.

[0094] The 3D point cloud in the above process (1) is used to describe the high-fidelity details of the physical environment, but its massive amount of data causes transmission delays that exceed the acceptable threshold for real-time data sharing. Since AR users observe the physical environment within a limited area on the same side, this invention adopts a lightweight multi-user augmented reality system framework, using a two-dimensional planar representation to replace the 3D point cloud, which can effectively reduce the problem of excessive delays caused by excessive data volume. In this MuAR system, there are two important roles: the host and the resolver. It should be noted that the host and the resolver are equivalent concepts, and their roles can be interchanged.

[0095] like Figure 2 As shown, the MuAR workflow mainly includes 7 steps: (1) The host acquires video and generates two-dimensional surface data of the physical environment based on the semantic-based quadtree method; (2) The host inputs virtual objects and performs a three-dimensional registration operation to create and maintain local real-text coordinates; (3) The host renders virtual objects based on real-time lighting; (4) To share virtual objects, the host inserts the virtual objects into the quadtree data and transmits them to the parser; (5) The parser extracts semantic features from the received quadtree data and matches them with the local feature library; (6) The parser constructs a matrix to represent the spatial relationship of the Field of Views (FoVs) between the host and the parser; (7) The parser renders the virtual objects shared by the host. Positional offsets caused by different device processing capabilities should be corrected to ensure that AR users have a consistent experience. Therefore, positional offset correction should be performed after step 7.

[0096] This invention considers a DTDD-based edge computing system consisting of BS (Browser Base Station) and AR (Augmented Reality) users. It is assumed that AR users are randomly distributed throughout the network and connected to the nearest BS. Each BS is equipped with an MEC (Multi-access Edge Computing) server, providing network services while handling complex computational tasks. BSs are connected to each other and to a central cloud server via a wired core network. Each BS caches a feature library of its physical environment and downloads raw data of virtual objects from the central cloud server. The BS's index set is used... It means that N B This refers to the number of BS (Block Layout). In a small cell, BS B n Employing an Orthogonal Frequency Division Multiple Access (OFDMA) architecture, through N C The number of channels is N A Provide services to AR users. In the nth small cell, the index sets of AR users and channels are respectively used as... and It means. BS B nExecution of channel allocation set matrix Representation, where vector If BSB n For AR users U m Channel C allocation i Then c n，m，i =1, otherwise c n，m，i =0.

[0097] When an AR user acts as a host, it uploads its locally generated data to the relevant BS. When acting as a parser, it downloads data from other AR users' data from the BS. To handle asymmetric traffic between uplink (UL) and downlink (DL) channels, this invention employs a DTDD frame structure, such as... Figure 3 As shown, it defines a frame time as 10ms, consisting of 10 subframes, each with a time of 1ms. Each subframe consists of 14 OFDM symbols, used for UL transmission, DL transmission, and the Guard Period (GP). The GP is simply a configuration of two adjacent symbols located between subframes in different transmission directions. A complete subframe is entirely used for either UL or DL. The frame set uses... This indicates that the subframe set of frame t is represented by... Where N F =10. BS B n In channel C i The subframe configuration set on is represented as If channel C i Subframe F on k BBS n Configured to send AR users U m DL, then f i，t，k =1, otherwise f i，t，k =0.

[0098] Since the execution of computational tasks is sequential, while channel states and traffic patterns are time-varying, we propose a dual-timescale resource allocation scheme: a large timescale is used for computational resource allocation, and a small timescale is used for channel allocation and subframe configuration. n For AR users U m The joint optimization of channel allocation and subframe configuration is defined as tuple Φ n，m :={c n,m，i ,f i，t,k}

[0099] (1) Transmission Model

[0100] To improve spectral efficiency, this experiment allows transmission links within different cells to reuse the same channel, which may lead to Same-Link Interference (SLI) and Cross-Link Interference (CLI) in D-TDD systems. SLI occurs when two AR devices in adjacent cells transmit data in the same direction; CLI occurs when two AR devices in adjacent cells transmit data in opposite directions.

[0101] As the host, AR user U m In subframe F k via channel C i To BSB n Upload its registration data or virtual object model; UL transmission rate is:

[0102]

[0103] Where |C i | is channel C i bandwidth, P n，m AR user U m The transmission power, g mn [i] is AR user U m To BSB n Channel gain, σ 2 It is the additive Gaussian noise power. It's BSB n Received interference includes SLI from AR users in neighboring cells. and CLI from the adjacent BS The calculation formula is:

[0104]

[0105] Where, N′ M It is with AR user U m The number of AR users in neighboring BS n′ using the same channel, P n′，m′ and P n′ These are AR users U′ m And the transmission power of neighboring BS n′, g m′n AR user U′ m To B n Channel gain, g n′n It is from BS n′ to BSB n Channel gain.

[0106] As a parser, AR user U m In subframe F k via channel Ci From BSB n Download the calculation results of your own task and the virtual object models of other AR users. The DL transfer rate is:

[0107]

[0108] Among them, P n It's BSB n The transmission power, g nm [i] is BSB n To AR user U m Channel gain. AR user U m The received interference is from SLI from a neighboring BS. CLI from neighboring small cell AR users Composition, its value is:

[0109]

[0110] Among them, g n′m and g m′m These are BS n′ and AR user U′ respectively. m To AR user U m Channel gain.

[0111] (2) Computational Model

[0112] Since 3D registration and rendering are resource-intensive tasks, they are all performed by the MEC server. As the host, AR user U... m First, 3D registration is performed to obtain the current orientation and position, and then rendering of the virtual object initiated by it begins. As the resolver, AR user U... m Render virtual objects initiated by other hosts based on its own orientation and location information. Define triples. As the l-th large timescale period, BSB n For AR users U m Decision variables for allocating computational resources. and These represent the proportions of computational resources allocated to perform 3D registration, host rendering, and resolver rendering, respectively. The computational latency for performing 3D registration is also mentioned. Computation latency of host rendering Computational latency of parser rendering They are respectively:

[0113]

[0114]

[0115] Computational resource requirements are defined as triples in and These represent the number of CPU cycles required to perform 3D registration, host rendering, and resolver rendering, respectively. n It means BSB n The total number of CPU cycles on the MEC server needs to be considered in relation to BSB. n The decision variables for allocating computational resources should satisfy the following constraints:

[0116]

[0117] (3) Delay Model

[0118] Taking the HoloLens 2 AR device as an example, its display refresh rate and camera capture rate are both synchronized at 60Hz. This means that the end-to-end latency of each video frame must not exceed 16ms, otherwise artifacts or jitter may occur. Therefore, we set 16ms as the large-scale period, and a subframe is considered a small-scale period. During the AR application initialization process, after the AR device captures the first video frame through the camera, it retrieves the current System Frame Number (SFN) through the Master Information Block (MIB) broadcast information. It then enters a standby phase until the large-scale lifetime begins, ensuring synchronization with the MuAR system.

[0119] Assuming AR user U m In the Subframe begins to move towards BSB n Upload a frame of video, and Subframe upload complete. subframe and If subframes are located in the same frame, then the following conditions must be met:

[0120]

[0121] Where |T s | indicates a 1ms upload latency:

[0122]

[0123] like If the subframe is in frame t+1, then the following conditions must be met:

[0124]

[0125] Upload latency is:

[0126]

[0127] like The subframe is located in frame t+2, and subframe and If the interval between subframes does not exceed 16ms, then the following conditions are met:

[0128]

[0129] Upload latency is:

[0130]

[0131] Assuming the rendered virtual object starts from the first... Downloading begins at the subframe, and at the... End of subframe. subframe and If subframes are located in the same frame, then the following conditions must be met:

[0132]

[0133] Download latency is:

[0134]

[0135] like If the subframe is in frame t+1, then the following conditions must be met:

[0136]

[0137] Download latency is:

[0138]

[0139] like The subframe is located in frame t+2, and subframe and If the interval between subframes does not exceed 16ms, then the following conditions are met:

[0140]

[0141] Download latency is:

[0142]

[0143] Therefore, AR user U m The end-to-end delay experienced by a single video frame can be expressed as:

[0144]

[0145] (4) Position offset model

[0146] Latency consistency among AR users is crucial because inconsistent latency reduces spatial alignment accuracy and causes the same virtual object to appear shifted from the viewpoint of different users. The relationship between positional shift rate and AR user latency differences can be modeled as follows:

[0147]

[0148] Here, ρ1 serves as a global scaling factor, determining the overall magnitude of the position offset rate. ρ2 adjusts the initial amplitude of the exponential function modeling the sensitivity to delay differences. ρ3 adjusts the nonlinear sensitivity to delay differences, with larger ρ3 values ​​amplifying PO. n，m Response to delay variations. ρ4 controls the degree of nonlinearity of the model response curve. Finally, ρ5 serves as a baseline offset term to ensure PO n,m It retains physically meaningful values ​​even under zero-latency conditions. To meet the normalization requirements, the factors should satisfy ρ1≥0, ρ2>0, ρ3<0, ρ4>0, ρ5≥0 and

[0149] (5) Optimization Objective

[0150] End-to-end latency of video frames is a key indicator of the Quality of Experience (QoE) in AR, and keeping it below 16 milliseconds is sufficient to meet user needs. Furthermore, positional offset rate is a key indicator for evaluating the consistency of interaction among AR users in the MuAR system. Therefore, the optimization objective of this paper is to minimize the positional offset rate under the constraint of latency below 16 milliseconds. To achieve this objective, we propose a dual-time-scaled resource management framework, where channel allocation, subframe configuration, and computational resource allocation are jointly optimized. The positional offset rate minimization problem can be formulated as follows:

[0151]

[0152] Among these constraints, constraint C1 means that the end-to-end delay cannot exceed 16 milliseconds. Constraints C2 and C3 ensure that the channel allocation and subframe configuration parameters are binary, respectively. Constraint C4 limits the computational resource allocation ratio for each BS to a range of 0 to 1, excluding 0. Constraint C5 shows the computational capability limit for each BS.

[0153] (6) Solution

[0154] To address the problem of minimizing position offset rate, a TD-RA scheme is proposed. The TD-RA scheme framework consists of two layers: an upper layer is a large-time-scale CRA sub-scheme, and a lower layer is a small-time-scale CASC sub-scheme, as shown below. Figure 5As shown. Each large time period consists of 16 hourly periods, and these smaller time periods are aligned with subframes. Each BS equipped with a MEC server acts as a proxy, implementing the resource allocation scheme in a distributed manner.

[0155] First, within each large-scale lifetime, each AR device captures one video frame. The corresponding computational requirements are forwarded to the relevant BS (Base Station) and then shared between adjacent BSs via fiber optic cables. The computational requirements and CASC (Content Execution Control) decision from the last small time period of the previous large time period are input as state to the BO-DDPG (Board of DDPG) network. The BO-DDPG network learns the characteristics of the computational requirements and then outputs the CRA (Content Assistance Decision) with the minimum end-to-end latency based on the given CASC decision. Within each small time-scale interval, the CRA decision of the current large time-scale lifetime, the amount of data in the uploaded video frame, the downloaded output after rendering, and information such as signal strength and bidirectional interference are input to the BiGAT-DQN (Bi-GAT-DQN) network. Finally, BiGAT is used to learn bidirectional interference characteristics, and DQN is used to make the CASC decision.

[0156] CRA sub-solution based on BO-DDPG

[0157] At each large timescale, the BO-DDPG CRA subroutine is formulated as a computational latency minimization problem, as follows:

[0158]

[0159] Problem II is a nonlinear programming problem involving strongly coupled polynomial functions. and DDPG, as a model-free algorithm in a continuous action space, offers a general solution to the problem of minimizing computational latency. However, when used as a parser, each AR user needs to collect the rendering computational resource requirements of other users to initialize virtual objects, and these requirements arrive randomly. Therefore, directly applying DDPG to computational resource allocation sub-schemes is quite challenging. Integrating Bayesian optimization (BO) into the DDPG framework improves DDPG's learning efficiency by providing informed action exploration directions, thus addressing the aforementioned problem.

[0160] (1) Bayesian optimization

[0161] The distribution of the number of CPU cycles required for 3D registration is simulated as a Pareto distribution (PD), as shown below:

[0162]

[0163] Where a, b, and c represent shape, scale, and threshold, respectively. Its cumulative distribution function can then be calculated as follows:

[0164]

[0165] The distribution of CPU cycles required for host rendering exhibits a severe tail characteristic, and its truncated lognormal distribution (TLD) is shown below:

[0166]

[0167] Where μ is the mean and δ is the standard deviation; its cumulative distribution function can then be calculated as follows:

[0168]

[0169] The distribution of the total number of CPU cycles required for parser rendering is simulated as an exponential distribution, as shown below:

[0170]

[0171] Where λ is the interaction strength; then, its cumulative distribution function can be calculated as follows:

[0172]

[0173] Define a function that maps CRA variables to compute resource requirements, as shown below:

[0174]

[0175] Where ∈ represents the error term, which is considered as Gaussian noise with a mean of 0. According to Bayes' theorem, the posterior distribution... With respect to the prior distribution and likelihood function of g(Ψ) Related, as shown below:

[0176]

[0177] in This is the historical sample set. Then, a utility function is defined to describe the expected improvement in the function value g(Ψ), as follows:

[0178]

[0179] Among them, g * (Ψ) represents the past maximum function value. The gradient descent algorithm is used to solve the utility function maximization problem. By maximizing the utility function, the optimal CRA decision for the next time slot can be found.

[0180] (2)DDPG

[0181] The computational latency minimization problem is reformulated as a Markov decision process, consisting of a state space, an action space, and a reward function. In the CRA sub-scheme, the base station is treated as the agent, and BSB... n Elements in the state space Indicates AR user U m The computational resource requirements for the three types of tasks; AR user U m Action space includes in and These represent the proportions of computational resources allocated to perform 3D registration, host rendering, and resolver rendering, respectively; the reward function is defined as the total negative value, i.e.

[0182] The DDPG framework consists of a training DNN and a target DNN. The training DNN is a function approximator used to obtain the optimal action a based on the current policy π(θ). * (t) and the state-action value of the action. As shown below:

[0183]

[0184] Target DNNs are used to evaluate the best action based on the following factors:

[0185]

[0186] Where θ and ω are the weight parameters of the training DNN and the target DNN, respectively. The weight parameters of the training DNN are updated through backpropagation, and the error loss is:

[0187]

[0188] Where, N R This is the size of the random sampling batch in the replay memory. The target DNN copies the weight parameters of the trained DNN at fixed steps.

[0189] (3) BO-DDPG

[0190] Bayesian optimization provides a mathematical framework for statistically modeling the conditional probabilities of events based on historical samples, and then optimizing the utility function associated with the occurrence of the event to determine the optimal action. It enhances the role of semi-supervised learning by providing a more informed direction for action exploration, especially in the early stages of training. Based on a given posterior distribution, the optimal CRA decision is output via gradient descent. The optimal CRA decision made by training the DNN This is represented by the state-action value derived from the Bellman equation. and Finally, by comparing the two state-action values, the final output action of DDPG is determined.

[0191] The specific process of BO-DDPG is as follows: (1) First, the parameters are initialized; (2) CRA decision is obtained by Bayesian optimization and training DNN respectively; (3) Based on the two actions, BS calculates their respective state-action values ​​and selects the action with the larger state-action value for output; (4) BS executes the selected operation on the system and records the transition to the replay memory.

[0192] CASC sub-scheme based on BiGAT-DQN

[0193] Considering that interactions between AR users are dynamic and that each BS (Base Station) cannot obtain global environmental knowledge, we formulate the CASA (Content Observable Markov Decision Process) problem as a distributed partially observable Markov decision process (PO-MDP). Each BS acquires local observation data and makes CASA decisions in a distributed manner. The PO-MDP consists of...<o,s,a,r> The state space is composed of σo representing the observation, s representing the state, a representing the action, and r representing the reward. The action space is represented by a = {c n,m,i f i，t，k} represents the condition. Given a CRA decision, the reward function can be expressed as a series of single CASA problems, as shown below:

[0194]

[0195] The complex spatial relationships between Base Stations (BSs) make solving the CASC problem extremely difficult in DTDD-based distributed MuAR systems. Specifically, the following challenges need to be considered: 1) Insufficient environmental information observed by each agent; 2) Distributed policy updates by agents lead to an uncertain and non-steady learning environment; 3) Interference between agents is bidirectional, and different types of interference have varying degrees of importance. Specifically, when making decisions, agents need to consider interference from neighboring cells, as well as the interference their own decisions cause to neighboring cells. Furthermore, considering factors such as available channel capacity and channel state, the impact of interference of the same value can differ.

[0196] In this work, we propose a cooperation mechanism between agents, using BiGAT-DQN to capture the relationships between agents. The MuAR system can be modeled as a system consisting of N... B A directed graph composed of subgraphs, using This indicates that the nth subgraph is composed of... It means that, among them Represents AR users, ε:={e m，m′} represents the disturbance that node m receives from node m′. For example... Figure 4 As shown, the interference received by node m from other nodes and the interference caused by node m to other nodes are respectively input into the MLP layer. Then, the MLP layer extracts the latent features of each node and transforms them into high-dimensional features, represented as follows:

[0197]

[0198] in and These are the weight parameters that need to be trained in an MLP, and σ(·) represents the ReLU activation function. n，m and These represent the interference that node m receives from other nodes and the interference that node m causes to other nodes, respectively. The coefficients between nodes can be implemented using multi-head attention, as shown below:

[0199] α m,m′ =ATT(Wh m , Wh m′ (41)

[0200] Where ATT(·) is the self-attention mechanism, and W is the corresponding weight parameter. This formula represents the importance of node m′ to node m. The mask and normalized attention coefficients can then be implemented as follows:

[0201]

[0202] Where softmax(·) is the softmax function, and LeakyReLU(·) is the LeakyReLU function. The coefficient α m，m′ and This can be obtained through equations (36) and (37). Furthermore, the state embedding can also be obtained by concatenating Q independent attention mechanisms, with the following value:

[0203]

[0204] Where ||(·) represents the connection operation, and Q represents the number of attention heads. Furthermore, the state embedding is passed to the fully connected layer for feature fusion, and its calculation formula is:

[0205]

[0206] The state space of PO-MDP is used Similar to the traditional DQN algorithm, the BiGAT-DQN algorithm employs a training DNN and a target DNN, where the training DNN takes a specific action in the current state based on (29) and (30), and the target DNN evaluates the action based on (31). The training DNN in the DQN algorithm updates the weight parameters by minimizing the mean squared error loss function, which is:

[0207]

[0208] in, This refers to the mini-batch size. The weight vector μ(t+1) can be solved using gradient descent, i.e.:

[0209]

[0210] Where Γ represents the learning rate.

[0211] TD-RA convergence analysis

[0212] From a lifecycle perspective, Problem I is a long-term cumulative function. However, Problems II and III are one-off problems, requiring decisions to be made only on large or small timescales. Considering the stochastic processes of other BSCs, the following theorem can be introduced to verify the convergence of the TD-RA scheme.

[0213] Theorem 1: For the long-running TD-RA problem, if the average position offset is ξ, then the probability that traversing the position offset is feasible is 1, i.e.

[0214]

[0215] in,

[0216] Proof: If a sequence of positive integers s t If it is convergent, then it satisfies the following condition:

[0217]

[0218] in, is a sequence of positive integers s t The average value.

[0219]

[0220] Among them, 0 <M0<∞。

[0221] Furthermore, Abel's theorem states that:

[0222]

[0223] Therefore, we have:

[0224]

[0225] Based on this, we define long-term returns as γ is the discount factor. Then, we can obtain:

[0226]

[0227] The proof is complete.

[0228] The present invention will now be described in detail with reference to the accompanying drawings and simulation examples.

[0229] The detailed steps of the CRA solution based on BO-DDPG are as follows:

[0230] Step 1: Determine and enter computing resource requirements and

[0231] Step 2, initialize the number of BS N B AR user count N A The computing power W of each MEC server n The parameters of the Pareto distribution, truncated log-normal distribution, and exponential distribution are: a, b, c, δ, σ, and λ, and historical samples. The weight parameters θ and ω of the training DNN and the target DNN.

[0232] Step 3, agent n, based on historical samples Estimate posterior probability

[0233] Step 4: Maximize the utility function according to equation (28) and use the gradient descent algorithm to obtain the optimal CRA decision.

[0234] Step 5, according to equations (29) and (30), utilize the state and Obtain the CRA decision made by the trained DNN

[0235] Step 6, Input Action and To the target DNN.

[0236] Step 7: Evaluate the state-action value using the target DNN. and

[0237] Step 8: The final action output by DDPG is determined based on the higher state-action value.

[0238] Step 9: Store (s(t), a(t), r(t), s(t+1)) into the playback memory.

[0239] Step 10: Update the weights of the trained DNN according to equation (32).

[0240] Step 11, steps 3 to 10 constitute one frame of operation. Repeat steps 3 to 10 T times, and every T... s Step 1, assign θ to ω.

[0241] Step 12: All the above steps constitute a single process. Repeat the above steps until the end.

[0242] Next, we will present simulation results to evaluate the effectiveness of the TD-RA scheme in the D-TDD MuAR system. In this system, 10 BSs are uniformly distributed, and 300 AR users are randomly distributed within a 2×2 square kilometer area. The key parameters of the system are shown in Table I. The parameter value of the position offset rate model in Equation (20) is ρ1=-5.247e -10 ,ρ2=5.1,ρ3=0.9521,ρ4=45.96,ρ5=-8.648e -5 A GAT consists of four attention heads, each composed of a single-layer feedforward neural network activated by the LeakyReLU activation function. DDPG employs an input layer, three hidden layers, and an output layer, with 512, 128, and 256 neurons respectively. DQN consists of four fully connected layers, with 128, 128, and 64 neurons respectively.

[0243]

[0244] The convergence performance of the BODDPG, DDPG, and BO algorithms on the CRA problem is as follows: Figure 6 As shown, it is clear that BO-DDPG converges faster and achieves a higher average cumulative reward compared to other schemes. This is because BO improves the learning efficiency of DDPG by estimating computational resource requirements. Furthermore, compared to DDPG and BO, BODDPG exhibits a more stable and smoother learning curve in both the early learning and convergence phases. This indicates that BO can effectively reduce stochastic exploration in DDPG. It can also be seen that BO converges to the lowest average cumulative reward. This is because each BS achieves the optimal CRA decision using BO based on the distribution of historical CRA decisions within the average cumulative reward over a large timescale in the CASC sub-scheme. It does not consider the dynamic nature of the network environment.

[0245] The relationship between latency and the number of AR users for BO-DDPG, DDPG, and BO algorithms is as follows: Figure 7As shown, the latency of all three algorithms increases with the number of AR users, with BO-DDPG achieving the lowest latency. When the number of users exceeds 121, the computational latency of BO becomes 0, and when the number of users exceeds 169, the computational latency of DDPG becomes 0. This is because in this invention, when the end-to-end latency for a user exceeds 16ms, the computational latency is set to 0, indicating that the computational resource decision has failed. This demonstrates that, under conditions of limited computational resources, our proposed BODDPG-based CRA sub-scheme can serve more AR users compared to DDPG and BO.

[0246] The relationship between latency and computational power for BO-DDPG, DDPG, and BO algorithms is as follows: Figure 8 As shown, the latency of the three algorithms is inversely proportional to the computing power, with BO-DDPG exhibiting the lowest latency. This indicates that, under the same computing power, BO-DDPG has the best latency performance.

[0247] The relationship between latency and interaction strength of BO-DDPG, DDPG, and BO algorithms is as follows: Figure 9 As shown, the latency of the three algorithms increases with the increase of interaction intensity, with BO-DDPG achieving the lowest latency. This indicates that, under conditions of limited computing resources, our proposed BODDPG-based CRA sub-scheme meets the interaction requirements of more AR users compared to DDPG and BO.

[0248] Figure 10 The convergence performance of different algorithms for solving the CASC problem is demonstrated. It can be observed that the BiGAT-DQN algorithm converges almost completely in approximately 1800 rounds. The GAT-DQN with GAT converges almost completely in 100, 200, 300, 400, and 500 MHz (0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1 rounds, while the traditional DQN algorithm only converges almost completely in approximately 2100 rounds. The BiGAT-DQN algorithm converges almost completely in 1800 rounds without MAH, while the traditional DQN algorithm only converges almost completely in approximately 2100 rounds.

[0249] The BiGAT-DQN algorithm achieves the highest average cumulative reward because ICI can effectively reduce bidirectional and multidimensional interference characteristics captured by the BiGAT mechanism. The convergence value of BiGAT-DQN with the algorithm is similar to that of BiGAT-DQN without the algorithm, but BiGAT-DQN converges faster. This is because BiGAT-DQN without the algorithm allows the agent to distinguish whether changes in the wireless channel state are due to its own behavior or the behavior of other agents that enhances its learning efficiency. Without the help of BiGAT and the multi-head attention mechanism, the traditional DQN algorithm converges to the lowest average cumulative reward.

[0250] Figure 11-13 This shows the relationship between the average offset rate and the number of AR users, bandwidth, and interaction intensity in different algorithms, specifically the relationship between BiGAT DQN, GAT-DQN, BiGATDQN, and DQN. Figure 11 It can be seen that the average offset rate of all algorithms increases with the number of AR users. This means that the higher the AR user density in a cell, the lower the spectral efficiency. Figure 12 As the intensity of interaction on each agent increases, the average offset of all algorithms decreases because the amount of interaction data to be processed and transmitted increases. Figure 13 As channel bandwidth increases, the average offset rate of all algorithms decreases because higher bandwidth leads to faster transmission rates. Compared to other algorithms, our proposed BiGAT-DQN achieves the lowest location offset rate under varying conditions of AR user number, bandwidth, and interaction intensity. These results demonstrate that by analyzing the characteristics of ICI between AR users, BiGAT-DQN can fully utilize channel resources and minimize the location offset rate caused by heterogeneous device processing capabilities.

[0251] Figure 14 The table shows how the location offset rate varies with the number of AR users under different resource allocation schemes (i.e., TD-RA, TD-RAw.o.CRA, TD-RAw.o.CA, and TD-RAw.o.SC). It can be seen that the location offset increases with the number of AR users for all methods, with our proposed TD-RA consistently outperforming the other schemes. This indicates that TD-RA is more suitable for dense network environments compared to other schemes.

[0252] Figure 15The table shows the variation of location offset rate with computing capacity under different resource allocation schemes (i.e., TD-RA, TD-RA woCRA, TD-RA woCA, and TD-RA w.o.SC). It can be seen that, under the same computing capacity, the location offset rate of our proposed TD-RA scheme is consistently superior to other schemes, especially when the computing capacity is small, its superior performance is even more pronounced. This indicates that the TD-RA scheme is suitable for various computing capacity environments, especially those with limited computing power.

[0253] Figure 16 The figure shows the variation of location offset rate with bandwidth under different resource allocation schemes (i.e., TD-RA, TD-RA woCRA, TD-RA woCA, and TD-RA w.o.SC). As the figure shows, regardless of whether the bandwidth is high or low, the location offset rate of our proposed TD-RA scheme is consistently lower than that of other schemes, especially when the bandwidth is small, its superior performance is even more pronounced. This indicates that the TD-RA scheme is suitable for various bandwidth environments, especially those with limited bandwidth.

Claims

1. A resource allocation method for a dual-timescale MuAR system within a 6G network framework, characterized in that, Includes the following steps: S1: Set the video refresh cycle of the AR device to a large time scale. Within each large time scale, apply the Bayesian optimization-enhanced depth deterministic strategy gradient algorithm to make computational resource allocation decisions and minimize the computational latency of 3D registration and rendering tasks. S2: Set the duration of a subframe to a small time scale. Within each small time scale, apply a bidirectional graph attention network and a deep Q-network to make channel allocation and subframe configuration decisions, so as to minimize the position offset rate of virtual objects in different display devices.

2. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 1, characterized in that, In S1, the deep deterministic policy gradient algorithm enhanced with Bayesian optimization includes: S1-1: Statistical modeling of the conditional probability of events based on historical samples; S1-2: Optimize the utility function related to the occurrence of the event. Based on the given posterior distribution, output the optimal computational resource allocation decision using gradient descent. S1-3: Training a DNN is a function approximator that obtains the optimal computational resource allocation decision based on the current strategy. And based on the Bellman equation, respectively, and Status - Action Value and S1-4: By comparing two state-action values and Select the state-action value with the highest value as the final output action.

3. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 2, characterized in that, S1-1: The process of statistically modeling the conditional probability of an event based on historical samples includes: Bayesian optimization: computational resource requirements are defined as triples. in and These represent the number of CPU cycles required to perform 3D registration, host rendering, and resolver rendering, respectively; where host rendering refers to the rendering of virtual objects initiated by the AR user when running as a host; and resolver rendering refers to the rendering of virtual objects initiated by other hosts when running as a resolver. The distribution of the number of CPU cycles required for 3D registration is simulated as a Pareto distribution, expressed as: Where a, b, and c represent shape, scale, and threshold, respectively; Its cumulative distribution function can be calculated in the following way: The number of CPU cycles required for host rendering can be represented by a truncated log-normal distribution as follows: Where μ represents the mean and δ represents the standard deviation; Its cumulative distribution function can be calculated in the following way: The distribution of the total number of CPU cycles required for parser rendering is simulated as an exponential distribution, as shown below: Where λ represents the interaction strength; Its cumulative distribution function can be calculated in the following way:

4. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 2, characterized in that, S1-2: The process of optimizing the utility function related to the occurrence of an event, and outputting the optimal computational resource allocation decision using gradient descent based on a given posterior distribution, includes: Define a function that maps CRA variables to compute resource requirements, as shown below: Among them Ψ n，m It is a CRA variable, i.e., BSB. n For AR users U m The decision variable for allocated computational resources, ∈, is the error term, which is treated as Gaussian noise with a mean of 0; According to Bayes' theorem, the posterior distribution With respect to the prior distribution and likelihood function of g(Ψ) Related, as shown below: in It is a historical sample set; Define a utility function to describe the expected improvement in the function value g(Ψ), expressed as: g * (Ψ) represents the past maximum function value; The gradient descent algorithm is used to solve the utility function maximization problem. By maximizing the utility function, the optimal CRA decision for the next time slot is found.

5. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 2, characterized in that, In S1-3, training the DNN is a function approximator that obtains the optimal computational resource allocation decision based on the current strategy. The process includes: The problem of minimizing computational latency is formulated as a Markov decision process, consisting of a state space, an action space, and a reward function. Treating the base station as an intelligent agent, the nth base station B n Elements in the state space Indicates AR user U m The computational resource requirements for three types of tasks: 3D registration, host rendering, and resolver rendering; AR user U m Action space includes in and These represent the proportions of computing resources allocated to perform 3D registration, host rendering, and resolver rendering, respectively; the reward function is defined as the negative of the total computational latency, i.e. in and These represent the computational latency for performing 3D registration, host rendering, and resolver rendering, respectively. The framework consists of a training DNN and a target DNN; the training DNN is a function approximator used to obtain the optimal action a based on the current policy π(θ). * (t) and the state-action value of the action. As shown below: Target DNNs are used to evaluate the best action based on the following factors: Where θ and ω are the weight parameters of the training DNN and the target DNN, respectively; the weight parameters of the training DNN are updated through backpropagation, and its error loss is: Where, N R It is the size of the random sampling batch of the replay memory; the target DNN copies the weight parameters of the trained DNN at fixed steps.

6. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 1, characterized in that, In S2, the process of making channel allocation and subframe configuration decisions using a bidirectional graph attention network and a deep Q-network includes: S2-1: Model the mobile multi-user augmented reality system as a system consisting of N B A directed graph composed of subgraphs is used to capture bidirectional interference dependencies between agents through a bidirectional graph attention network; S2-2: A joint optimization decision for channel allocation and subframe configuration is generated through a deep Q network to minimize the average position offset rate.

7. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 6, characterized in that, In S2-1, the mobile multi-user augmented reality system is modeled as a system consisting of N... B The process of capturing bidirectional interference dependencies between agents in a directed graph composed of subgraphs using a bidirectional graph attention network includes: Model the mobile multi-user augmented reality system as a system consisting of N B A directed graph composed of subgraphs, using This indicates that the nth subgraph is composed of... It means that, among them Represents AR users, ε:={e m，m′ } represents the disturbance that node m receives from node m′; The interference received by node m from other nodes and the interference caused by node m to other nodes are respectively input into the MLP layer. The MLP layer extracts the latent features of each node and converts them into high-dimensional features, represented as: in and These are the weight parameters that need to be trained in an MLP, where σ(·) represents the ReLU activation function, and I n，m and These represent the interference that node m receives from other nodes and the interference that node m causes to other nodes, respectively. The coefficients between nodes can be implemented through multi-head attention, as shown below: α m，m′ =ATT(Wh m Wh m′ ); Where ATT(·) is the self-attention mechanism, W is the corresponding weight parameter, and this formula represents the importance of node m′ to node m; Masking and normalized attention coefficients can be implemented in the following ways: Where softmax(·) is the softmax function and LeakyReLU(·) is the LeakyReLU function, the state embedding is obtained by concatenating Q independent attention mechanisms, and its value is: Where ||(·) represents a join operation, and Q represents the number of attention heads; The state embedding is passed to the fully connected layer for feature fusion, and its calculation formula is as follows:

8. The resource allocation method for a dual-timescale MuAR system under a 6G network framework according to claim 6, characterized in that, The process of generating joint optimization decisions for channel allocation and subframe configuration through a deep Q-network includes: The state space of partially observable Markov decisions is used express; In a deep Q-network, the trained DNN updates the weight parameter μ by minimizing the mean squared error loss function, which is: in, Given a mini-batch size, the weight vector μ(t+1) can be solved using gradient descent, i.e.: Where Γ represents the learning rate.

Images