A hierarchical joint optimization method for energy-saving generative communication for multi-user applications

Abstract

This invention discloses a hierarchical joint optimization method for energy-efficient generative communication in multi-user scenarios. The method includes: constructing a multi-constraint joint optimization problem with the objective of minimizing total system energy consumption for uplink generative communication scenarios involving multiple heterogeneous user terminals and a single edge server; decoupling the outer layer into a two-layer iterative architecture of outer compression ratio optimization and inner resource allocation; solving the inner layer using sequential convex approximation and alternating direction multiplier method; and transforming the compression ratio optimization problem into a monotonic optimization problem by splitting the compression ratio variable in the outer layer, then iteratively converging using block coordinate descent and external approximation algorithms; finally, outputting the optimal configuration and executing transmission. This invention effectively solves the problems of deep coupling of multi-dimensional resources, complex optimization solutions, high computational overhead, and difficulty in converging to a high-quality solution in multi-user generative communication. While satisfying latency and content-aware quality constraints, it minimizes total system energy consumption and improves computational efficiency and system scalability in multi-user scenarios.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a hierarchical joint optimization method for multi-user energy-saving generative communication. Background Technology

[0002] With the development of 6G networks, intelligent applications (such as digital twins, remote surgery, and virtual reality) are generating massive amounts of image and video data, posing serious challenges to the capacity and energy efficiency of communication systems. Traditional communication paradigms, based on Shannon's information theory, focus on bit-level error-free transmission without considering the specific meaning and task relevance of the transmitted data, leading to inefficiency and significant resource waste when transmitting large-scale multimedia content. To address this issue, generative communication, as an emerging paradigm, has attracted widespread attention. Its core lies in using generative models to learn the semantic distribution of data, extracting and transmitting only deep features relevant to the receiving end's task, thereby significantly reducing transmission redundancy and improving spectral efficiency.

[0003] Despite the immense potential of generative communication architectures, existing technologies still face significant limitations in resource allocation within multi-user environments. Traditional optimization methods, such as alternating optimization, typically group variables, decomposing the original problem into several easily manageable subproblems. However, when system variables are highly coupled, these methods struggle to account for the joint effects between variables, leading to a severe deterioration in solution quality and even difficulty in achieving stable convergence when applied directly.

[0004] To this end, this invention proposes a hierarchical decoupling optimization framework, which aims to comprehensively consider the differences in semantic service requirements, terminal computing power and channel conditions among heterogeneous users. Under the premise of strictly meeting the constraints of latency and content-aware quality, it minimizes the total energy consumption of the system by jointly optimizing the compression ratio, local and edge computing frequencies, transmission power and transmission bandwidth, thereby providing a feasible resource management solution for the actual deployment of large-scale, high-energy-efficiency generative communication systems. Summary of the Invention

[0005] The purpose of this invention is to provide a hierarchical joint optimization method for energy-saving generative communication for multi-user systems. This method can solve the problems of high resource coupling, difficult optimization solutions, high energy consumption, and slow convergence in multi-user generative communication. Under the premise of strictly satisfying latency and content quality constraints, it minimizes the total energy consumption of the system and improves the computational efficiency and scalability in multi-user scenarios.

[0006] To achieve the above objectives, this invention provides a hierarchical joint optimization method for multi-user energy-saving generative communication, comprising the following steps: S1. System Modeling and Construction of the Original Joint Optimization Problem: For an uplink generative communication scenario involving multiple heterogeneous user terminals and an edge server, establish an energy consumption model, a latency model, and a generated content awareness quality model, and construct a joint optimization problem with the goal of minimizing the total system energy consumption while satisfying system operation constraints. S2. Decoupling and initialization of hierarchical problems: Decompose the joint optimization problem into an outer compression ratio optimization sub-problem and an inner computation and communication resource allocation sub-problem. Construct a two-layer nested iterative architecture in which the outer layer generates candidate compression ratio vectors and the inner layer finds the minimum total system energy consumption under a given compression ratio and feeds it back to the outer layer. S3. Solve the inner layer resource allocation subproblem: Introduce the spectral efficiency variable for substitution, use the sequential convex approximation to handle non-convex constraints, and combine the alternating direction multiplier method to solve the inner layer subproblem, and obtain the total system energy consumption and resource allocation results under the corresponding compression ratio; S4. Solve the outer compression ratio optimization subproblem: Split the compression ratio variable to decouple its dual impact on system performance, transform the outer problem into a monotonic optimization problem, update the compression ratio using block coordinate descent combined with the outer approximation algorithm, and call step S3 to iterate until convergence; S5. Output the optimal resource allocation scheme: Output the converged compression ratio and corresponding resource allocation parameters, configure the parameters of the user terminal and the edge server, and execute generative data transmission.

[0007] Preferably, the energy consumption model in S1 includes a local coding energy consumption model, a transmission energy consumption model, and an edge decoding energy consumption model; The latency models include local encoding latency model, transmission latency model, and edge decoding latency model; System operation constraints include end-to-end latency constraints, generated content-aware quality constraints, total computing power constraints of edge servers, total system bandwidth constraints, and physical constraints of variables.

[0008] Preferably, the objective function of the joint resource optimization problem constructed in S1 is to minimize the total energy consumption of the system. The formula is as follows: ; in, For the first The amount of raw data per user For the first Compression rate per user, For the first Local computing frequency per user Assign the edge server to the first The computation frequency of each user For the first The number of floating-point operations that a user terminal processor can perform per clock cycle The number of floating-point operations that an edge server processor can perform per clock cycle. and These represent the number of floating-point operations per bit of data for the encoder and decoder, respectively. The effective switching capacitor coefficient of the user terminal processor. The effective switching capacitor coefficient of the edge server processor. For the first Transmit power of each user For the first Bandwidth allocated to each user For the first Channel gain for each user For noise power spectral density, This represents the total number of heterogeneous user terminals.

[0009] Preferably, S3 specifically includes: S31. Introduce a spectral efficiency variable for variable substitution; S32. Use the sequential convex approximation method to handle non-convex power constraints; S33. Parallel solution is performed using the alternating direction multiplier method to obtain the total system energy consumption and resource allocation results under the corresponding compression ratio.

[0010] Preferably, S31 specifically includes: S311, Introduction of the first Spectrum efficiency variable for individual users The formula is: ; Based on the inverse solution of this spectral efficiency variable, the first... The explicit expression for the transmit power of an individual user is given by the formula: ; S312. Based on the inverse solution of the transmit power expression, the first... Wireless transmission power consumption per user The original form, based on transmit power and bandwidth, is rewritten in terms of spectral efficiency. It is a convex function with a single independent variable, as follows: ; in, For The first variable Wireless transmission power consumption per user; S313. Transform the content-aware quality constraint into a lower bound constraint on spectral efficiency: ; in, To generate content-aware quality threshold and the first Semantic compression rate per user The minimum spectral efficiency requirement is determined jointly.

[0011] Preferably, S32 specifically includes: In the During the subsequent convex approximation iterations, the non-convex power constraint is applied: ; in, For the first Maximum hardware transmit power limit for each user terminal; Equivalent rewrite as: ; At the previous bandwidth iteration point Perform a first-order Taylor expansion on the function on the right side to construct a linear convex approximation constraint: .

[0012] Preferably, S33 specifically includes: S331. To decouple multi-user shared resource constraints, a global auxiliary variable is introduced. and To satisfy: ; in, For the first A global auxiliary variable for user bandwidth. For the first A global auxiliary variable for the frequency of edge computing for each user. The total available bandwidth resources of the system. Total computing power available for edge servers; S332. Construct the augmented Lagrangian function and solve it iteratively through three stages: local variable update, global variable update, and dual variable update. ; ; ; ; Among them, superscript Indicates the first The next ADMM iteration, with superscript or The variables represent the values ​​of the corresponding variables in the corresponding ADMM iteration rounds; The set of local variables is defined as follows: ; For the global auxiliary variable set, defined as ; The set of scaled dual variables is defined as follows: ,in, For the scaling form dual variable corresponding to the bandwidth consistency constraint, For the scaled dual variable corresponding to the frequency consistency constraint in edge computing; To augment the Lagrange function, where This represents the set of penalty parameters, which includes the penalty parameters corresponding to bandwidth consistency constraints. Penalty parameters corresponding to edge computing frequency consistency constraints ; The global feasible region is composed of the non-negativity constraints of variables, the upper limit of the total bandwidth of the system, and the upper limit of the total computing power of the edge servers.

[0013] S333. Execute the core solution process of the alternating direction multiplier method, including local variable update, global variable projection update and dual variable update operations, and combine time delay constraints to solve for the optimal value of each variable. S334. Set the iteration termination condition: When the deviation between global and local variables and the update amount of dual variables are both less than the preset convergence threshold during the iteration process, stop the iteration and output the total system energy consumption and corresponding resource allocation results obtained in the current iteration; if the termination condition is not met, return to S333 to continue the iteration update until the convergence requirement is met.

[0014] Preferably, S333 specifically includes: S3331. During the local variable update process, for any user The local time delay constraint is expressed as: in, , , These are intermediate coefficients used to simplify local delay constraints, corresponding to the edge decoding delay term, local encoding delay term, and wireless transmission delay term, respectively, and satisfying the following: For the first The maximum end-to-end latency that a user can tolerate.

[0015] The local computational frequency is obtained by solving the Lagrange duality decomposition. Edge computing frequency Transmission bandwidth With spectral efficiency .

[0016] Among them, local computing frequency The updated value is: in, For the first Maximum local computing frequency limit per user terminal For the first Non-negative Lagrange multipliers corresponding to local delay constraints for each user.

[0017] Edge computing frequency The unique positive real root of the following cubic equation is obtained by solving it: in, The ADMM penalty parameter is the one corresponding to the edge computing frequency consistency constraint. This is a constant in the local update of edge computing frequency.

[0018] For the transmission bandwidth With spectral efficiency The resulting two-dimensional convex subproblem is solved using an alternating block coordinate descent method. (Fixed transmission bandwidth) At that time, by solving the following about the spectral efficiency The one-dimensional equation is obtained Candidate update values: By combining the lower bound constraint of spectral efficiency and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the spectral efficiency. The updated value.

[0019] Fixed spectral efficiency At that time, by solving the following about the transmission bandwidth The cubic equation is obtained Candidate update values: By combining the feasible bandwidth interval and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the transmission bandwidth. The updated value.

[0020] in, The ADMM penalty parameter is the one corresponding to the bandwidth consistency constraint. This is a constant in the local bandwidth update. Alternate execution of spectral efficiency. Update and transmission bandwidth Update until the preset local convergence condition is met, and update the Lagrange multipliers through a one-dimensional search. This allows the local delay constraint to satisfy the complementary relaxation condition, resulting in local update results for the user's local computing frequency, edge computing frequency, transmission bandwidth, and spectral efficiency.

[0021] S3332. During the global variable update process, the transmission bandwidth variable and edge computing frequency variable obtained in the local variable update stage are aggregated and projected onto a simplex set that satisfies non-negativity of elements and a limited sum, respectively, to obtain the update results of the global auxiliary variable of bandwidth and the global auxiliary variable of edge computing frequency. S3333. During the dual variable update process, based on the consistency error between local variables and global auxiliary variables, update the dual variables corresponding to the bandwidth consistency constraint and the dual variables corresponding to the edge computing frequency consistency constraint.

[0022] Preferably, the specific steps for splitting the compression ratio variable in S4 include: for each user compression ratio Perform variable splitting and introduce the first compressed variable. Second compression variable ; in, Used to characterize the beneficial effects of increased compression ratio on data volume, latency, and energy consumption. Used to characterize the adverse effects of increased compression rate on the perceived quality constraints of generated content; Introducing consistency constraints, the formula is: ; Substituting the second compression variable yields the intermediate variable. The formula is: ; By decoupling the dual effects of compression ratio through variable splitting, the non-monotonic outer-layer optimization problem is transformed into a standard monotonic optimization problem, as shown in the formula: .

[0023] Because the outer objective function and constraints are related to and It has monotonicity, when there is When a feasible solution is found, it can be advanced to the boundary along the monotonically improving direction. Furthermore, it does not degrade the objective function value, so the optimal solution lies on the boundary, and the inequality constraint does not change the optimal solution corresponding to the original compression ratio consistency relationship.

[0024] Preferably, in S4, updating the compression ratio using block coordinate descent combined with an external approximation algorithm specifically includes: pairing the two-dimensional variables corresponding to each user... As a variable block, a two-dimensional monotonic optimization subproblem is solved under the condition that the variable blocks of other users are fixed. The variable blocks of all users are updated cyclically using the block coordinate descent method. Each two-dimensional monotonic optimization subproblem is solved using the external approximation algorithm, and the iteration is repeated until convergence. In each iteration, the inner resource allocation solver of step S3 is called to calculate the objective function value.

[0025] Therefore, the hierarchical joint optimization method for multi-user energy-saving generative communication described above has the following beneficial effects: (1) By using the variable splitting technique, the original energy consumption optimization problem with non-monotonic characteristics is transformed into a monotonic optimization problem. Combined with the external approximation algorithm, a high-quality compression ratio is found, which significantly reduces the total energy consumption of the system compared with the alternating optimization algorithm.

[0026] (2) The ADMM-based solution algorithm decomposes the large-scale optimization problem with deep coupling of multiple variables into multiple parallel low-dimensional sub-problems, avoiding the high computational overhead of traditional centralized solution algorithms and greatly improving the system's scalability in multi-user scenarios.

[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0028] Figure 1 A flowchart of a hierarchical joint optimization method for multi-user energy-saving generative communication provided in an embodiment of the present invention; Figure 2 The convergence curve of the inner ADMM resource allocation algorithm provided in this embodiment of the invention; Figure 3 A comparison chart of the total system energy consumption of the proposed method provided in the embodiments of the present invention and the existing alternating optimization method. Detailed Implementation

[0029] Example The following combines the inclusion This paper describes in detail an uplink generative communication system consisting of heterogeneous mobile user terminals and an edge server, where the user set is denoted as . The edge server is responsible for collecting information such as channel status, service latency requirements, content-aware quality requirements, and local computing power and power limits of the terminal reported by each user. It then uniformly executes the hierarchical optimization process of this invention, obtains the collaborative configuration results, and distributes them to each user terminal.

[0030] This invention employs a two-layer nested architecture. The outer layer optimizes the compression ratio for each user, while the inner layer optimizes the local computing frequency, edge computing frequency, transmission bandwidth, and transmit power under a given compression ratio. Each time the outer layer generates a candidate compression ratio vector, it calls the inner layer solver once to obtain the minimum total energy consumption achievable by the system under that compression ratio vector. The outer layer then continues to update the compression ratio based on this feedback result until the convergence condition is met.

[0031] like Figure 1 As shown, this invention discloses a hierarchical joint optimization method for multi-user energy-saving generative communication, the steps of which include: S1. System Modeling and Construction of the Original Joint Optimization Problem: For an uplink generative communication scenario involving multiple heterogeneous user terminals and an edge server, establish an energy consumption model (including local encoding energy consumption model, transmission energy consumption model, and edge decoding energy consumption model), a latency model (including local encoding latency model, transmission latency model, and edge decoding latency model), and a generated content-aware quality model. Construct a joint optimization problem with the goal of minimizing the total system energy consumption and satisfying system operation constraints (including end-to-end latency constraints, generated content-aware quality constraints, edge server total computing power constraints, system total bandwidth constraints, and variable physical constraints).

[0032] The objective function of the constructed joint resource optimization problem is to minimize the total energy consumption of the system. The formula is as follows: ; in, For the first The amount of raw data per user For the first Compression rate per user, For the first Local computing frequency per user Assign the edge server to the first The computation frequency of each user For the first The number of floating-point operations that a user terminal processor can perform per clock cycle The number of floating-point operations that an edge server processor can perform per clock cycle. and These represent the number of floating-point operations per bit of data for the encoder and decoder, respectively. The effective switching capacitor coefficient of the user terminal processor. The effective switching capacitor coefficient of the edge server processor. For the first Transmit power of each user For the first Bandwidth allocated to each user For the first Channel gain for each user For noise power spectral density, This represents the total number of heterogeneous user terminals.

[0033] In this embodiment, S1 specifically refers to: S11, Local Coding Model Construction No. Local encoding delay for individual users The model is as follows: ; No. Local encoding energy consumption per user The model is as follows: ; In this embodiment, the first Compression rate per user This is achieved by controlling the number of features output from the encoder's final layer; the compressed feature data to be transmitted is [amount missing]. ; S12, Wireless Transmission Model Construction The system uses Frequency Division Multiple Access (FDMA) technology for transmission.

[0034] Transmission rate Given by Shannon's formula: ; Transmission delay The formula is: ; No. Wireless transmission power consumption per user The formula is: ; S13, Edge Decoding Model Construction Decoding latency The model is as follows: ; Decoding power consumption The model is as follows: ; S14. Generate a content-aware quality model Using the Multi-Scale Structural Similarity Index (MS-SSIM) as the metric, a system is constructed that includes the first... Compression rate per user and received signal-to-noise ratio The utility fitting function is given by the following formula: ; in, For the first A user-perceived quality function for generated content. These are the fitting parameters.

[0035] S15. Construction of Joint Optimization Problem A joint resource optimization problem is established with the objective of minimizing the total energy consumption of the system, and the formula is: ; in, To find the minimum total system energy consumption by using the compression ratio, transmit power, local computing frequency, transmission bandwidth, and edge allocation computing frequency of each user as optimization variables; The constraints include: End-to-end delay constraints: ; Generative content-aware quality constraints: ; Total server computing power constraint: ; Total system bandwidth constraints: ; Variable physical constraints: , , , , , ; in, Indicates user The minimum required generated content perceived quality.

[0036] S2. Decoupling and initialization of hierarchical problems: Decompose the joint optimization problem into an outer compression ratio optimization sub-problem and an inner computation and communication resource allocation sub-problem. Construct a two-layer nested iterative architecture in which the outer layer generates candidate compression ratio vectors and the inner layer finds the minimum total system energy consumption under a given compression ratio and feeds it back to the outer layer. During the initialization phase, the edge server generates an initial compression ratio candidate vector that meets the constraints, which serves as the starting point for the outer iteration.

[0037] S3. Solving the inner-layer resource allocation subproblem: Introducing a spectral efficiency variable for substitution, employing a sequential convex approximation to handle non-convex constraints, and combining the alternating direction multiplier method to solve the inner-layer subproblem, yielding the total system energy consumption and resource allocation results under the corresponding compression ratio; specifically including: S31. Introduce a spectral efficiency variable for variable substitution; specifically including: S311, Introduction of the first Spectrum efficiency variable for individual users The formula is: ; Based on the inverse solution of this spectral efficiency variable, the first... The explicit expression for the transmit power of an individual user is given by the formula: ; S312. Based on the inverse solution of the transmit power expression, the first... Wireless transmission power consumption per user The original form, based on transmit power and bandwidth, is rewritten in terms of spectral efficiency. It is a convex function with a single independent variable, as follows: ; in, For The first variable Wireless transmission power consumption per user; Combining local encoding, wireless transmission, and edge decoding, the complete end-to-end latency constraint is: ; S313. Transform the content-aware quality constraint into a lower bound constraint on spectral efficiency: ; in, To generate content-aware quality threshold and the first Semantic compression rate per user The minimum spectral efficiency requirement is determined jointly.

[0038] After the above variable substitutions, the objective function and all other constraints of the inner-layer optimization problem, except for the hardware maximum transmit power constraint, are transformed into convex form.

[0039] S32. A sequential convex approximation method is used to handle non-convex power constraints; specifically including: In the During the subsequent convex approximation iterations, the non-convex power constraint is applied: ; in, For the first Maximum hardware transmit power limit for each user terminal; Equivalent rewrite as: ; Because of the right side about The function is convex, and directly processing this constraint will still lead to a non-convex feasible region. Therefore, at the bandwidth iteration point in the previous round... Perform a first-order Taylor expansion on the function on the right side to construct a linear convex approximation constraint: .

[0040] S33. Parallel solution is performed using the alternating direction multiplier method to obtain the total system energy consumption and resource allocation results under the corresponding compression ratio. Specifically, this includes: S331. To decouple multi-user shared resource constraints, a global auxiliary variable is introduced. and To satisfy: ; in, For the first A global auxiliary variable for user bandwidth. For the first A global auxiliary variable for the frequency of edge computing for each user. The total available bandwidth resources of the system. Total computing power available for edge servers; S332. Construct the augmented Lagrangian function and solve it iteratively through three stages: local variable update, global variable update, and dual variable update. ; ; ; ; Among them, superscript Indicates the first The next ADMM iteration, with superscript or The variables represent the values ​​of the corresponding variables in the corresponding ADMM iteration rounds; The set of local variables is defined as follows: ; For the global auxiliary variable set, defined as ; The set of scaled dual variables is defined as follows: ,in, For the scaling form dual variable corresponding to the bandwidth consistency constraint, For the scaled dual variable corresponding to the frequency consistency constraint in edge computing; To augment the Lagrange function, where This represents the set of penalty parameters, which includes the penalty parameters corresponding to bandwidth consistency constraints. Penalty parameters corresponding to edge computing frequency consistency constraints ; The global feasible region is composed of the non-negativity constraints of variables, the upper limit of the total bandwidth of the system, and the upper limit of the total computing power of the edge servers.

[0041] S333: Execute the core solution process of the alternating direction multiplier method, including local variable update, global variable projection update, and dual variable update operations.

[0042] S3331, Local variable update: For any user The local time delay constraint is defined as: ; in, , , All of these are intermediate variables used to simplify local delay constraints: Related to the amount of data decoded at the edge. ; Related to the amount of locally encoded data, ; Related to the amount of data transmitted, ; For the first The maximum end-to-end latency that a user can tolerate; The local computation frequency, edge computation frequency, and bandwidth-spectral efficiency coupling variables are solved using Lagrange dual decomposition; the optimal solution for the local computation frequency is: ; in, For the first Maximum local computing frequency limit per user terminal For the first Lagrange multipliers constrained by user latency; The marginal calculation frequency is obtained by solving the unique positive real root of the following cubic equation: ; in, The ADMM penalty parameter is the one corresponding to the edge computing frequency consistency constraint. This is a constant in the local update of edge computing frequency.

[0043] For the transmission bandwidth With spectral efficiency The resulting two-dimensional convex subproblem is solved using an alternating block coordinate descent method. (Fixed transmission bandwidth) At that time, by solving the following about the spectral efficiency The one-dimensional equation is obtained Candidate update values: By combining the lower bound constraint of spectral efficiency and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the spectral efficiency. The updated value.

[0044] Fixed spectral efficiency At that time, by solving the following about the transmission bandwidth The cubic equation is obtained Candidate update values: By combining the feasible bandwidth interval and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the transmission bandwidth. The updated value.

[0045] in, The ADMM penalty parameter is the one corresponding to the bandwidth consistency constraint. This is a constant in the local bandwidth update. Alternate execution of spectral efficiency. Update and transmission bandwidth Update until the preset local convergence condition is met, and update the Lagrange multipliers through a one-dimensional search. This allows the local delay constraint to satisfy the complementary relaxation condition, resulting in local update results for the user's local computing frequency, edge computing frequency, transmission bandwidth, and spectral efficiency.

[0046] S3332, Global Variable Update: After aggregating the transmission bandwidth variable and edge computing frequency variable obtained in the local variable update stage, project them onto a simplex set that satisfies non-negativity of elements and a limited sum, to obtain the update results of the global auxiliary variable of bandwidth and the global auxiliary variable of edge computing frequency.

[0047] S3333, Dual Variable Update: Based on the consistency error between local variables and global auxiliary variables, update the dual variables corresponding to the bandwidth consistency constraint and the dual variables corresponding to the edge computing frequency consistency constraint.

[0048] S334. Set the iteration termination condition: When the deviation between global and local variables and the update amount of dual variables are both less than the preset convergence threshold during the iteration process, stop the iteration and output the total system energy consumption and corresponding resource allocation results obtained in the current iteration; if the termination condition is not met, return to S333 to continue the iteration update until the convergence requirement is met.

[0049] The linearized base point of the convex approximation sequence is updated using the bandwidth variable obtained in this round of solution, and S32 is returned to enter the next round of SCA iteration; the entire inner layer solution process is repeated until the difference between the total system energy consumption of two adjacent iterations is lower than the preset global convergence threshold of SCA, the iteration terminates, and finally the total system energy consumption and corresponding resource allocation results under the current compression ratio are output.

[0050] like Figure 2 As shown, the parallel solution algorithm based on ADMM used in this invention has fast convergence characteristics, and can usually achieve the preset convergence accuracy requirement within 10-20 iterations.

[0051] S4. Solving the outer compression ratio optimization subproblem: The compression ratio variable is split to decouple its dual impact on system performance, transforming the outer problem into a monotonic optimization problem. This specifically includes: For each user compression ratio Perform variable splitting and introduce the first compressed variable. Second compression variable ; in, Used to characterize the beneficial effects of increased compression ratio on data volume, latency, and energy consumption. Used to characterize the adverse effects of increased compression rate on the perceived quality constraints of generated content; Introducing consistency constraints, the formula is: ; Substituting the second compression variable yields the intermediate variable. The formula is: ; By decoupling the dual effects of compression ratio through variable splitting, the non-monotonic outer-layer optimization problem is transformed into a standard monotonic optimization problem, as shown in the formula: .

[0052] Because the outer objective function and constraints are related to and It has monotonicity, when there is When a feasible solution is found, it can be advanced to the boundary along the monotonically improving direction. Furthermore, it does not degrade the objective function value, so the optimal solution lies on the boundary, and the inequality constraint does not change the optimal solution corresponding to the original compression ratio consistency relationship.

[0053] The compression ratio is updated using a combination of block coordinate descent and external approximation algorithms, specifically including: The two-dimensional variable pairs corresponding to each user As a variable block, a two-dimensional monotonic optimization subproblem is solved under the condition that the variable blocks of other users are fixed. The variable blocks of all users are updated cyclically using the block coordinate descent method. Each two-dimensional monotonic optimization subproblem is solved using the external approximation algorithm, and the iteration is repeated until convergence. In each iteration, the inner resource allocation solver of step S3 is called to calculate the objective function value.

[0054] S5. Output the converged compression ratio and corresponding resource allocation parameters (including local computing frequency, edge computing frequency, transmission bandwidth, and transmit power); The edge server sends configuration parameters to each user, and each user adjusts the number of output features of the encoder's last layer according to the optimal compression ratio, performs generative coding according to the optimal local computing frequency, and transmits the compressed generative features according to the optimal transmit power and allocated bandwidth. The edge server decodes and reconstructs the data of each user according to the optimal edge computing frequency, completes the parameter configuration of the user terminal and the edge server, and executes generative data transmission. Under the premise of meeting the end-to-end latency constraints of each user and the perceptual quality constraints of the generated content, the total system energy consumption is reduced.

[0055] like Figure 3 As shown, compared with the traditional alternating optimization method, the hierarchical joint optimization method proposed in this invention can achieve lower total system energy consumption in different user number scenarios, and the energy consumption advantage becomes more significant as the number of users increases.

[0056] Therefore, the present invention adopts the above-mentioned hierarchical joint optimization method for multi-user energy-saving generative communication, which can effectively solve the problems of deep coupling of multi-dimensional resources, complex optimization solutions, large computational overhead, and difficulty in converging to a high-quality solution in multi-user generative communication. Under the premise of satisfying the constraints of latency and perceived quality of generated content, it can minimize the total energy consumption of the system, while improving the computational efficiency and system scalability in multi-user scenarios.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hierarchical joint optimization method for multi-user energy-saving generative communication, characterized by the following steps: include: S1. System Modeling and Construction of the Original Joint Optimization Problem: For an uplink generative communication scenario involving multiple heterogeneous user terminals and an edge server, establish an energy consumption model, a latency model, and a generated content awareness quality model, and construct a joint optimization problem with the goal of minimizing the total system energy consumption while satisfying system operation constraints. S2. Decoupling and initialization of hierarchical problems: Decompose the joint optimization problem into an outer compression ratio optimization sub-problem and an inner computation and communication resource allocation sub-problem. Construct a two-layer nested iterative architecture in which the outer layer generates candidate compression ratio vectors and the inner layer finds the minimum total system energy consumption under a given compression ratio and feeds it back to the outer layer. S3. Solve the inner layer resource allocation subproblem: Introduce the spectral efficiency variable for substitution, use the sequential convex approximation to handle non-convex constraints, and combine the alternating direction multiplier method to solve the inner layer subproblem, and obtain the total system energy consumption and resource allocation results under the corresponding compression ratio; S4. Solve the outer compression ratio optimization subproblem: Split the compression ratio variable to decouple its dual impact on system performance, transform the outer problem into a monotonic optimization problem, update the compression ratio using block coordinate descent combined with the outer approximation algorithm, and call step S3 to iterate until convergence; S5. Output the optimal resource allocation scheme: Output the converged compression ratio and corresponding resource allocation parameters, configure the parameters of the user terminal and the edge server, and execute generative data transmission.

2. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 1, characterized in that: The energy consumption model in S1 includes the local coding energy consumption model, the transmission energy consumption model, and the edge decoding energy consumption model; The latency models include local encoding latency model, transmission latency model, and edge decoding latency model; System operation constraints include end-to-end latency constraints, generated content-aware quality constraints, total computing power constraints of edge servers, total system bandwidth constraints, and physical constraints of variables.

3. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 2, characterized in that: The objective function of the joint resource optimization problem constructed in S1 is to minimize the total energy consumption of the system. The formula is as follows: ； in, For the first The amount of raw data per user For the first Compression rate per user, For the first Local computing frequency per user Assign the edge server to the first The computation frequency of each user For the first The number of floating-point operations that a user terminal processor can perform per clock cycle The number of floating-point operations that an edge server processor can perform per clock cycle. and These represent the number of floating-point operations per bit of data for the encoder and decoder, respectively. The effective switching capacitor coefficient of the user terminal processor. The effective switching capacitor coefficient of the edge server processor. For the first Transmit power of each user For the first Bandwidth allocated to each user For the first Channel gain for each user For noise power spectral density, This represents the total number of heterogeneous user terminals.

4. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 1, characterized in that, S3 specifically includes: S31. Introduce a spectral efficiency variable for variable substitution; S32. Use the sequential convex approximation method to handle non-convex power constraints; S33. Parallel solution is performed using the alternating direction multiplier method to obtain the total system energy consumption and resource allocation results under the corresponding compression ratio.

5. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 4, characterized in that, S31 specifically includes: S311, Introduction of the first Spectrum efficiency variable for individual users The formula is: ； Based on the inverse solution of this spectral efficiency variable, the first... The explicit expression for the transmit power of an individual user is given by the formula: ； S312. Based on the inverse solution of the transmit power expression, the first... Wireless transmission power consumption per user The original form, based on transmit power and bandwidth, is rewritten in terms of spectral efficiency. It is a convex function with a single independent variable, as follows: ； in, For The first variable Wireless transmission power consumption per user; S313. Transform the content-aware quality constraint into a lower bound constraint on spectral efficiency: ； in, To generate content-aware quality threshold and the first Semantic compression rate per user The minimum spectral efficiency requirement is determined jointly.

6. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 5, characterized in that, S32 specifically includes: In the During the subsequent convex approximation iterations, the non-convex power constraint is applied: ； in, For the first Maximum hardware transmit power limit for each user terminal; Equivalent rewrite as: ； At the previous bandwidth iteration point Perform a first-order Taylor expansion on the function on the right side to construct a linear convex approximation constraint: 。 7. The hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 6, characterized in that, S33 specifically includes: S331. To decouple multi-user shared resource constraints, a global auxiliary variable is introduced. and To satisfy: ； in, For the first A global auxiliary variable for user bandwidth. For the first A global auxiliary variable for the frequency of edge computing for each user. The total available bandwidth resources of the system. Total computing power available for edge servers; S332. Construct the augmented Lagrangian function and solve it iteratively through three stages: local variable update, global variable update, and dual variable update. ； ； ； ； Among them, superscript Indicates the first The next ADMM iteration, with superscript or The variables represent the values ​​of the corresponding variables in the corresponding ADMM iteration rounds; The set of local variables is defined as follows: ; For the global auxiliary variable set, defined as ; The set of scaled dual variables is defined as follows: ,in, For the scaling form dual variable corresponding to the bandwidth consistency constraint, For the scaled dual variable corresponding to the frequency consistency constraint in edge computing; To augment the Lagrange function, where This represents the set of penalty parameters, which includes the penalty parameters corresponding to bandwidth consistency constraints. Penalty parameters corresponding to edge computing frequency consistency constraints ; The global feasible region is composed of the non-negativity constraints of variables, the upper limit of the total bandwidth of the system, and the upper limit of the total computing power of the edge servers. S333. Execute the ADMM iterative solution process. In the local variable update phase, the resource variable sub-problems corresponding to each user are solved according to user division, and can be processed in parallel on the edge server side. In the global variable update phase, the global auxiliary variable of bandwidth and the global auxiliary variable of edge computing frequency are projected onto the global feasible region. During the dual variable update phase, the corresponding dual variable is updated based on the consistency error between the local variable and the global auxiliary variable. S334. Set the iteration termination condition: When the deviation between global and local variables and the update amount of dual variables are both less than the preset convergence threshold during the iteration process, stop the iteration and output the total system energy consumption and corresponding resource allocation results obtained in the current iteration; if the termination condition is not met, return to S333 to continue the iteration update until the convergence requirement is met.

8. A hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 7, characterized in that, S333 specifically includes: S3331, Local variable update: For any user The local time delay constraint is expressed as: in, , , These are intermediate coefficients used to simplify local delay constraints, corresponding to the edge decoding delay term, local encoding delay term, and wireless transmission delay term, respectively, and satisfying the following: For the first The maximum end-to-end latency that a user can tolerate; The local computational frequency is obtained by solving the Lagrange duality decomposition. Edge computing frequency Transmission bandwidth With spectral efficiency ; Among them, local computing frequency The updated value is: in, For the first Maximum local computing frequency limit per user terminal For the first The non-negative Lagrange multipliers corresponding to the local delay constraints of each user; Edge computing frequency The unique positive real root of the following cubic equation is obtained by solving it: in, The ADMM penalty parameter is the one corresponding to the edge computing frequency consistency constraint. This is a constant in the local update of edge computing frequency; For the transmission bandwidth With spectral efficiency The two-dimensional convex subproblem is solved using the block coordinate descent method with alternating solutions; fixed transmission bandwidth. At that time, by solving the following about the spectral efficiency The one-dimensional equation is obtained Candidate update values: By combining the lower bound constraint of spectral efficiency and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the spectral efficiency. The updated value; Fixed spectral efficiency At that time, by solving the following about the transmission bandwidth The cubic equation is obtained Candidate update values: By combining the feasible bandwidth interval and the linearized power constraint, the obtained candidate values ​​are projected onto the feasible interval to obtain the transmission bandwidth. The updated value; in, The ADMM penalty parameter is the one corresponding to the bandwidth consistency constraint. For constants in local bandwidth updates; spectral efficiency is performed alternately. Update and transmission bandwidth Update until the preset local convergence condition is met, and update the Lagrange multipliers through a one-dimensional search. This allows the local delay constraints to satisfy the complementary relaxation condition, resulting in local update results for the user's local computing frequency, edge computing frequency, transmission bandwidth, and spectral efficiency. S3332, Global Variable Update: After aggregating the transmission bandwidth variable and edge computing frequency variable obtained in the local variable update stage, project them onto a simplex set that satisfies non-negativity of elements and a limited sum, to obtain the update results of the global auxiliary variable of bandwidth and the global auxiliary variable of edge computing frequency. S3333, Dual Variable Update: Based on the consistency error between local variables and global auxiliary variables, update the dual variables corresponding to the bandwidth consistency constraint and the dual variables corresponding to the edge computing frequency consistency constraint.

9. A hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 1, characterized in that: In S4, the specific splitting of compression ratio variables includes: for each user compression ratio Perform variable splitting and introduce the first compressed variable. Second compression variable ; in, Used to characterize the beneficial effects of increased compression ratio on data volume, latency, and energy consumption. Used to characterize the adverse effects of increased compression rate on the perceived quality constraints of generated content; Introducing consistency constraints, the formula is: ； Substituting the second compression variable yields the intermediate variable. The formula is: ； By decoupling the dual effects of compression ratio through variable splitting, the non-monotonic outer-layer optimization problem is transformed into a standard monotonic optimization problem, as shown in the formula: ； Because the outer objective function and constraints are related to and It has monotonicity, when there is When a feasible solution is found, it can be advanced to the boundary along the monotonically improving direction. Furthermore, it does not degrade the objective function value, so the optimal solution lies on the boundary, and the inequality constraint does not change the optimal solution corresponding to the original compression ratio consistency relationship.

10. A hierarchical joint optimization method for multi-user energy-saving generative communication according to claim 9, characterized in that: S4 uses a combination of block coordinate descent and external approximation algorithm to update the compression ratio, specifically including: adjusting the two-dimensional variables corresponding to each user... As a variable block, a two-dimensional monotonic optimization subproblem is solved under the condition that the variable blocks of other users are fixed. The variable blocks of all users are updated cyclically using the block coordinate descent method. Each two-dimensional monotonic optimization subproblem is solved using the external approximation algorithm, and the iteration is repeated until convergence. In each iteration, the inner resource allocation solver of step S3 is called to calculate the objective function value.

Images