An energy-aware 6g network random quantization federated learning communication and computation integrated resource allocation method

Abstract

The application discloses an energy-aware 6G network random quantization federated learning communication and computation integrated resource allocation. The application quantizes the local model parameters of each user, aiming to meet the energy consumption and quantization error constraints, and minimize the total convergence time of federated learning. Among them, the quantization error is used to ensure that the model accuracy is maintained at an ideal level, and the minimum target is set in actual time units, not just for global communication rounds. The application makes distributed intelligence run efficiently, reliably and sustainably in a resource-constrained wireless environment through resource fusion and intelligent scheduling, accelerating the original design of 6G wireless network endogenous intelligence.

Description

Technical Field

[0001] This invention relates to the field of 6G network technology, specifically to an energy-aware 6G network random quantization federated learning communication and computing integrated resource allocation method. Background Technology

[0002] With the advent of the 6G era in communication technology, network-intrinsic intelligence has become a core development direction. 6G deeply integrates artificial intelligence from external supplementary applications into the network's intrinsic fundamental capabilities, driving highly self-optimizing, intelligent self-healing, and extreme energy efficiency. However, this vision faces severe challenges. On the one hand, the network environment is dynamic and complex; on the other hand, massive numbers of terminal devices are widely distributed, generating data with privacy sensitivity and heterogeneous characteristics. Against this backdrop, Federated Learning (FL), as a privacy-preserving distributed machine learning framework, collaborates with multiple clients to train models without aggregating raw data, becoming key to building 6G distributed intelligence. However, traditional federated learning still faces two major difficulties in deploying in real wireless environments: first, frequent model interactions bring huge communication overhead, crowding out scarce spectrum resources and generating high energy consumption; second, the significant heterogeneity of devices in terms of computing power, data distribution, and energy reserves leads to low training efficiency or even global failure.

[0003] To address these challenges, firstly, model compression and quantization techniques are crucial for improving federated learning efficiency. By reducing the accuracy of model update propagation, the communication burden can be significantly alleviated. The value of stochastic quantization lies not only in providing a flexible trade-off between communication efficiency and model accuracy, but also in its randomness, which produces a regularization-like effect to enhance model generalization capabilities. Secondly, network architecture itself is evolving from the traditional model of separating communication, computing, and storage resources to a design that integrates communication and computing (JSAC). Future 6G networks will support intelligent applications in a globally optimal manner through cross-layer and cross-domain resource joint scheduling. This requires stochastic optimization to dynamically and collaboratively allocate wireless resources such as transmit power, sub-channels, and quantization levels, as well as computing resources such as CPU cycles and task offloading decisions, online without relying on difficult-to-obtain prior statistical information.

[0004] Applying quantization techniques and federated learning to 6G network energy efficiency resource allocation essentially involves distributed intelligent model training and network resource management decisions. Simultaneously, energy awareness becomes a key constraint in technology design. Whether in wireless communication or local model computation, energy consumption directly determines the actual battery life of terminal devices, impacting the greenness and sustainability of the entire network operation. This invention breaks down fragmented design, deeply collaborating stochastic quantization, energy awareness, and federated learning within a unified communication and computing framework, achieving a leap from "step-by-step processing" to "global linkage." In terms of performance, adaptive quantization and joint resource scheduling ensure model accuracy while reducing communication energy consumption and shortening system convergence time. The energy awareness mechanism protects vulnerable devices, improving the fairness and overall robustness of heterogeneous network training. In terms of application, it provides an engineerable and scalable solution for massive, heterogeneous, and energy-sensitive IoT terminals to participate in distributed intelligent training.

[0005] This invention, operating in a 6G network environment with energy efficiency as its goal, employs stochastic quantization technology to overcome communication bottlenecks in federated learning. Through integrated resource allocation for communication and computation, it optimizes the quantization process, wireless transmission, and local computation across layers, achieving an optimal balance between privacy protection, learning efficiency, resource utilization, and energy consumption. This method can be deployed in 6G scenarios such as the Internet of Things (IoT), intelligent vehicle networks, and the Industrial Internet, which require high energy efficiency, strong privacy protection, and adaptive dynamic environments. Summary of the Invention

[0006] This invention proposes an energy-aware, integrated resource allocation method for communication and computation in stochastic quantization federated learning in 6G networks. This method satisfies energy constraints and quantization error tolerance conditions, jointly optimizes communication resources, computational resources, and quantization bit allocation, and minimizes the total convergence time of the federated learning task. This invention can be widely applied in 6G wireless network environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An energy-aware, stochastic quantization-based federated learning integrated communication and computing resource allocation method for 6G networks includes the following steps: S1. Construct a federated learning framework to collect state information from the device side; S2. Based on the status information, design a quantitative transmission mechanism to quantize the local model update amount generated by the device. S3. Based on the quantized local model update amount, construct the joint communication computation timing sequence to describe the total completion time and energy consumption of each round of training; S4. Based on the joint communication calculation time sequence, weigh the quantization error and aggregation impact, and give the quantized aggregation update; S5. Based on the quantized aggregate update, calculate the resource coordination optimization problem constrained by device energy, communication, and computing resources with the goal of minimizing the single-round completion time. S6. Based on the resource collaborative optimization problem, design an online solution implementation method for dual decomposition to dynamically solve the optimal quantization accuracy, transmission power, and computational resource allocation strategy of the system.

[0008] Preferably, S1 includes: Build a server and M A federated learning framework composed of multiple end users defines users. m Local loss function and global objective function Using learning rate And small-batch data updates, and collection of device-side status information; among which, users m In the i The model parameters after the first round of local training are: The amount of local model updates uploaded to the server is: in, Represents the current global model; Indicates user In the The small batch of samples selected in the step; user In the In the round of global federated training, the first The local model parameter vector obtained after step local stochastic gradient descent; This represents stochastic gradient estimation based on mini-batch samples.

[0009] Preferably, S2 includes: Introducing quantization operators , will users m In the i Wheelset local model update vector Each component adopts Each quantized bit is randomly quantized, and the lower and upper bounds of the quantization amplitude range are defined as 0 and 0, respectively. and in accordance with The system performs uniform segmentation, determines the quantization value of each component according to random rules, ensuring that the quantizer is unbiased for each component, and adjusts the quantization value according to the number of quantization bits. Calculate the amount of data uploaded by user m. .

[0010] Preferably, S3 includes: Construct the total completion time for each round of federated learning tasks. It consists of a local computing phase and an uplink transmission phase; among which, according to the user m allocated CPU frequency Take calculation time and its energy consumption According to the user m Transmission power and uplink transmission duration Computational communication energy consumption The minimum lower bound of the delay required to satisfy the communication constraints is obtained according to the Shannon capacity formula; uplink transmission is performed using time division multiple access, and the total communication delay is... The total latency is calculated as the sum of the transmission times of all users. It is determined by the slowest terminal.

[0011] Preferably, S4 includes: defining a user m In the i Wheel vector quantization error The server-side aggregation update relationship is as follows: in, This represents the percentage of the sample size. Represents the current global model; This represents the local model update vector; Represents the quantization operator; M Indicates the number of end users; Define aggregate quantization error: And a constraint is imposed on the second moment of the aggregated quantization error on the server side: in, Indicates the first The allowable aggregate quantization error tolerance of the round. It represents the mathematical expectation.

[0012] Preferably, S5 includes: using the first i Total completion time of each round of federal learning tasks Minimize as the optimization objective, calculate frequency Transmission power Transmission duration and the number of random quantization bits As a joint optimization variable, it satisfies the constraints of single-round energy budget, communication transmission rate, aggregated quantization error second moment, maximum CPU frequency, maximum transmission power, and the constraint that the number of quantization bits is a positive integer.

[0013] Preferably, S6 includes: to By performing continuous relaxation, the Lagrangian function of the single-round optimization problem is constructed. The coupling constraints are relaxed by constructing the dual problem. The original problem is decomposed into subproblems that optimize the computation frequency, transmission power and transmission duration, and subproblems that optimize the number of random quantization bits in combination with the aggregated quantization error tolerance. The Lagrangian multipliers corresponding to the energy budget constraint, communication constraint and aggregated error constraint are updated online based on the subgradient method or the alternating optimization method.

[0014] Preferably, the single-round energy budget is determined by the terminal in the first round. i The calculation frequency is determined by the remaining energy at the start of each round, the preset safe energy threshold, and the requirements for subsequent consecutive training rounds. Transmission power and the number of random quantization bits Feasible allocation range Compared with the prior art, the beneficial effects of the present invention are as follows: In one embodiment, the present invention has at least the following beneficial effects.

[0015] First, this invention incorporates the remaining energy of the terminal into single-round energy budget modeling, and calculates the feasible allocation range of frequency, transmission power and random quantization bits based on this constraint. Therefore, it can improve the adaptability of resource allocation decisions to the heterogeneous energy state of the terminal in scenarios where the terminal energy is limited and multiple rounds of continuous training are involved.

[0016] Second, this invention employs an unbiased random quantization mechanism and uses the second moment of aggregated quantization error as a training accuracy constraint, thereby explicitly transforming the model accuracy requirement into a measurable and adjustable constraint. This establishes a clear trade-off between communication overhead and model accuracy, providing an interpretable basis for optimizing the configuration of the number of random quantization bits.

[0017] Third, this invention aims to minimize the completion time of a single round of federated learning tasks by jointly optimizing terminal computing resources, wireless communication resources, and the number of random quantization bits. Compared to benchmark methods such as fixed bit allocation, equal time slot allocation, or equal energy allocation, this invention can more effectively shorten the latency of a single round of training and improve overall training efficiency while meeting the aggregate quantization error tolerance and terminal energy budget.

[0018] Fourth, the present invention adopts a dual decomposition and round-by-round online solution mechanism, which can dynamically generate near-optimal resource allocation results for the real-time changing channel state, terminal remaining energy and error tolerance requirements in the 6G network environment, thereby improving the feasibility and real-time adaptability of the method in actual wireless federated learning scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the overall structure and flow of an embodiment of the present invention; Figure 2 This is a flowchart of federated learning with local model quantization in an embodiment of the present invention; Figure 3 This is a diagram showing the duration structure of each round of global federated learning in an embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 This embodiment provides an energy-aware, stochastic quantization-based federated learning method for allocating communication and computing resources in a 6G network. Its overall structure and flow are as follows: Figure 1 As shown, it includes the following steps: S1. Construct a federated learning framework to collect state information from the device side.

[0024] First, a federated learning framework is constructed, oriented towards single server-multiple users. Local datasets and loss functions are defined, and learning rate and mini-batch data updates are adopted to collect state information from the device side.

[0025] In this embodiment, the federated learning framework is constructed using a server and It consists of 10 terminal users, and the user index is denoted as . Each user holds their own local dataset. The sample size is denoted as The server maintains global model parameters. And organize users to conduct iterative collaborative training.

[0026] In one embodiment of this example, the federated learning system operates in a 6G network environment. Compared to traditional wireless access networks, 6G networks offer greater bandwidth, lower latency, higher reliability, and stronger network-side intelligent scheduling capabilities, thus providing fundamental support for multiple terminals to concurrently participate in model training, parameter uploading, and global aggregation. In this embodiment, the high-quality wireless connectivity and intelligent resource orchestration capabilities provided by 6G networks are used as important network prerequisites for subsequent joint optimization of communication and computing resources during the system modeling phase.

[0027] Furthermore, in one embodiment of this example, "energy perception" refers to using the remaining energy of the terminal, computing energy consumption, and communication energy consumption as core perception quantities in the resource allocation process of each round of federated learning, so that the system can meet the training performance requirements while taking into account the terminal energy constraints and long-term operating capabilities.

[0028] In one embodiment of this example, the user The local loss function is defined as: in, Represents model parameters In the sample Error function on.

[0029] The total sample size of the system is defined as: The global objective function can then be expressed as: Accordingly, the federated learning task can be formulated as the following optimization objective: In one implementation, the first During global training, the server sends data to the user. Broadcast the current global model .user Execution based on this global model Step stochastic gradient descent, its local update can be written as: in, , Indicates the first Round learning rate, Indicates user In the The small batch of samples selected in the step user In the In the round of global federated training, the first The local model parameter vector obtained after a step of local stochastic gradient descent. This represents a stochastic gradient estimate based on a mini-batch of samples, satisfying: .

[0030] Based on the above recursive relationship, we can obtain the user In the The model parameters after the first round of local training are: .

[0031] Therefore, users The amount of local model updates uploaded to the server can be defined as: .

[0032] The server performs a weighted aggregation of the model updates uploaded by each user and obtains the next round of global model: Therefore, the global model update can be viewed as a weighted aggregation of the accumulated local gradient results of each terminal.

[0033] S2. Based on the status information, design a quantization transmission mechanism to quantify the local model update amount generated by the device.

[0034] Next, a quantization transmission mechanism was designed, introducing quantization operators to quantize the update components generated by the device and upload them to the server to participate in global model updates, thereby reducing communication load. In one implementation, to reduce communication overhead during federated learning, the server does not require users to upload high-precision model updates, but instead requires users to upload their quantization results. Let the quantization operator be denoted as... Then the server in the 1st The wheel model update can be written as: .

[0035] In one implementation, the user In the wheel adopted Quantize bits and update vectors for the local model. Each component is randomly quantized.

[0036] Federated learning with local model quantization, such as Figure 2 As shown. To avoid ambiguity during symbol recovery due to zero components and to ensure that the quantization interval covers all possible amplitudes, the lower and upper bounds of the quantization amplitude interval are first defined as follows: In the interval Above, according to ( If the maximum segment index corresponding to the quantization amplitude interval is divided uniformly, then the i-th segment... Each segment endpoint can be represented as: .

[0037] If a certain parameter component satisfy In one implementation, the quantization value is determined according to the following random rule: in, Indicates user In the During round-wise random quantization, the first round A small range of quantized amplitudes.

[0038] As can be seen from the above construction, this quantizer satisfies unbiasedness for each component, that is: in, It represents the mathematical expectation.

[0039] Further define the quantization error as: Then we have: And its conditional variance satisfies: Therefore, in one implementation, the entire quantized local model update can be represented as: .

[0040] In one implementation, if each component requires One amplitude bit and one symbol bit, and additional transmission interval range information is required. Bits, then user In the The amount of data uploaded in a round can be represented as: Therefore, a larger quantization bit count results in a larger amount of data uploaded while reducing quantization error; conversely, a smaller quantization bit count results in a smaller amount of data uploaded while reducing model update accuracy. Thus, the quantization bit count constitutes a key trade-off between communication efficiency and training accuracy.

[0041] S3. Based on the quantized local model update amount, construct the joint communication computation timing to describe the total completion time and energy consumption of each training round.

[0042] Construct a joint communication computation timing sequence to describe the total completion time (including total computation latency and total communication latency) and energy consumption of each training round.

[0043] In one implementation, the total completion time of each round of federated learning tasks consists of both the local computation phase and the uplink transmission phase. For users... In terms of its first Round execution The computation time required for each local iteration is: in, Indicates user In the CPU frequency allocated in rounds, This represents the number of CPU cycles required to process a single sample, therefore This represents the equivalent number of CPU cycles required to complete one local iteration. When using mini-batch training, this value can be understood as the equivalent value after adjusting for batch size and number of local data traversals.

[0044] The corresponding computational energy consumption can be expressed as: in, This refers to the energy consumption coefficient related to the terminal hardware.

[0045] In one implementation, the wireless federated learning system is deployed in a 6G network scenario. Compared to traditional cellular networks, 6G networks offer greater system bandwidth, lower transmission latency, higher link reliability, and stronger intelligent resource scheduling capabilities. Furthermore, they can be combined with novel capabilities such as ultra-dense access, integrated air-space-ground coverage, reconfigurable intelligent surfaces (RIS), integrated sensing and communication (ISAC), and AI-native networks to provide communication support for large-scale terminal collaborative learning.

[0046] In one embodiment of the present invention, the aforementioned 6G network capabilities are mainly reflected in two aspects: first, providing high-bandwidth, low-latency, and highly reliable transmission conditions for updating and uploading federated learning models; second, the network side can jointly coordinate terminal computing resources and wireless communication resources by combining terminal channel status, remaining energy, and training task requirements. Furthermore, in this embodiment, instead of independently modeling the underlying physical layers of key 6G technologies such as RIS and ISAC, their gain effects on wireless transmission and network scheduling are equivalently reflected in parameters such as channel gain, link reliability, and resource scheduling capabilities, thereby highlighting the joint optimization of communication and computing resources at the federated learning task level.

[0047] In one implementation, considering the feasibility requirements of latency analysis and resource allocation modeling in the federated learning model update and upload process, Time Division Multiple Access (TDMA) is used in the communication phase. TDMA serves only as an uplink access abstraction method to facilitate analysis and resource optimization, and does not preclude further extension to more general orthogonal or non-orthogonal multiple access mechanisms in 6G scenarios. In the The wheel's transmission power is Uplink transmission time is The corresponding communication energy consumption can be expressed as: .

[0048] According to Shannon's capacity formula, users In the The constraints for successful transmission are as follows: in, Indicates system bandwidth. Indicates user In the The channel gain of the wheel, Represents the noise power spectral density. Indicates user The amount of quantized data to be uploaded. From the above formula, we can further obtain the lower bound of the minimum latency required to satisfy the communication constraints: .

[0049] The duration structure of each round of global federated learning, such as... Figure 3 As shown. In one implementation, since all terminals perform local training in parallel, therefore the first... The total computation time of the round is determined by the slowest terminal, that is: TDMA ensures that uplink transmissions from each user occur sequentially, therefore the first... The total communication latency is: Therefore, the first The total completion time for each round of federated learning tasks can be expressed as: .

[0050] S4. Based on the joint communication calculation timing, weight the quantization error and aggregation impact, and give the quantized aggregation update.

[0051] Next, the weights are used to measure the error and the impact of aggregation, and the quantified aggregation update is given to ensure the stability of the federated aggregation results and the model convergence accuracy.

[0052] In one implementation, the impact of random quantization on model aggregation can be viewed as adding a zero-mean perturbation to the standard federated learning update. Define user. In the The vector quantization error of the wheel is: .

[0053] According to the component-wise unbiasedness, we have: Therefore, the server-side aggregation update relationship can be rewritten as: .

[0054] Let the aggregation quantization error be: The quantized federated learning update can then be represented as a standard FedAvg update with added perturbation terms. FedAvg refers to the Federated Averaging algorithm, which involves the server performing a weighted average aggregation of the local model updates uploaded by each terminal based on the proportion of samples from each terminal, in order to obtain the next round of global model.

[0055] In one implementation, to ensure the stability of the federated aggregation results and the model convergence accuracy after random quantization upload, a constraint is imposed on the second moment of the aggregation quantization error on the server side: in, Indicates the first The allowable aggregate quantization error tolerance.

[0056] Since the quantization operator satisfies component-wise unbiasedness, this constraint can limit the perturbation intensity introduced by random quantization while maintaining the unbiased characteristics of the aggregation update. Larger Corresponding to lower quantization accuracy but shorter communication latency, smaller This corresponds to higher quantization accuracy but also a greater communication burden. In some implementations, a relatively loose quantization error tolerance is used in the early stages of training, and this tolerance is gradually reduced in the later stages of training to balance the latency efficiency in the early stages of training and the model accuracy in the later stages of training.

[0057] S5. Based on the quantized aggregate update, calculate the resource coordination optimization problem constrained by device energy, communication, and computing resources, with the goal of minimizing the single-round completion time.

[0058] This step primarily involves calculating the minimum single-round completion time, with resource co-optimization being a problem constrained by multiple factors including device energy, communication, and computing resources. In one implementation, the communication computing resource allocation mentioned in this embodiment refers to the joint configuration and coordinated adjustment of the computing frequency, wireless transmission power, uplink transmission duration, and random quantization bit count of each participating terminal in federated learning, so as to achieve a comprehensive balance between single-round latency, energy consumption, and model accuracy in federated learning tasks.

[0059] In one implementation, the first Total completion time of each round of federal learning tasks Minimize as the optimization objective, calculate frequency Transmission power Transmission duration and quantization bit number As joint optimization variables, denoted as: The single-round resource optimization problem can then be expressed as: in, , and Representing users respectively The single-round energy budget, maximum CPU frequency, and maximum transmission power. In one embodiment, the single-round energy budget... By the terminal in the The remaining energy at the start of the round, the preset safe energy threshold, and the requirements of subsequent continuous training rounds are jointly determined, thereby explicitly mapping the "remaining terminal energy" to the available energy budget for this round, and further determining the feasible allocation range of calculation frequency, transmission power, and number of random quantization bits, reflecting the "energy sensing" characteristic described in this embodiment.

[0060] As can be seen from the above model, improving It can reduce local computation time, but increase computational energy consumption; improve It can reduce transmission time, but increase communication energy consumption; improve While random quantization errors can be reduced, the amount of data and transmission latency increase. Therefore, this embodiment optimizes the above variables uniformly to minimize the single-round completion time while meeting the terminal's remaining energy budget and aggregation error tolerance.

[0061] Furthermore, from the perspective of optimization structure, the aforementioned single-round resource optimization problem belongs to a nonlinear constrained optimization problem with integer variables. Among these, the quantization bit number variable... The variables are discrete integers, and the objective function includes a maximum value operator and a communication delay term. The constraints include... The problem involves nonlinear expressions such as the logarithmic rate function and quantization error statistics. Therefore, this problem can be classified into the category of Mixed-Integer Nonlinear Programming (MINLP), and in general, it manifests as a nonconvex optimization problem.

[0062] In one implementation, a dual decomposition method is used to solve the problem because the original problem exhibits significant multi-user coupling characteristics and direct solution is computationally expensive. On one hand, while each user's computational frequency, transmit power, transmission duration, and number of random quantization bits correspond to their respective terminal's local decision variables, they are also coupled through single-round completion time objectives, communication transmission constraints, and aggregated quantization error constraints. On the other hand, directly solving the original MINLP problem globally typically requires high computational complexity, making it difficult to meet the real-time requirements of round-by-round online resource allocation in federated learning.

[0063] Based on this, in one implementation, the integer variable is first... By performing continuous relaxation, energy constraints, communication constraints, and quantization error constraints are introduced into the Lagrange function, and the coupling constraints are relaxed by constructing a dual problem. After this process, the original problem can be decomposed into several lower-dimensional subproblems based on the user or variable type, and the computational resources, communication resources, and quantization bit count are optimized respectively. The consistency between the subproblems is then coordinated through iterative updates of the Lagrange multipliers. For the quantization bit count solution obtained after continuous relaxation, in one implementation, a feasible integer bit configuration can be restored by rounding down or combining constraint feasibility with a neighborhood search. This allows for obtaining an approximately optimal solution that satisfies the constraints while ensuring controllable solution complexity.

[0064] Therefore, it can be seen that the dual decomposition method used in this embodiment can effectively handle the nonlinear coupled optimization structure with continuous relaxation of integer variables, and is particularly suitable for online approximate solution by round in wireless federated learning scenarios.

[0065] S6. Based on the resource collaborative optimization problem, design an online solution implementation method for dual decomposition to dynamically solve the optimal quantization accuracy, transmission power, and computational resource allocation strategy of the system.

[0066] Finally, an online solution implementation method for dual decomposition is designed to dynamically solve for the optimal quantization accuracy, transmission power, and computational resource allocation strategy of the system.

[0067] In one implementation, considering that the above problem contains both continuous and integer variables, and that the channel state, terminal remaining energy state, and error tolerance requirements during training typically change with each round, a round-by-round online solution method is adopted. After continuous relaxation, the Lagrangian function for the single-round optimization problem can be constructed as follows: in, Indicates user In the The uplink transmission rate of the wheel, , and It is a Lagrange multiplier.

[0068] In one implementation, given the dual variables, the primal problem can be decomposed into the following subproblems: (1) Fixed number of random quantization bits Optimize calculation frequency Transmission power With transmission duration ; (2) Fixed computing and communication resource allocation, combined with aggregate quantization error tolerance to optimize the number of random quantization bits ; (3) Update the Lagrange multipliers corresponding to the energy budget constraint, communication constraint and aggregation error constraint online based on the subgradient method or alternating optimization method.

[0069] Through the aforementioned iterative process, for each round of real-time observed channel state, remaining terminal energy, and error tolerance requirements, an approximately optimal resource allocation result satisfying the constraints can be obtained. This method effectively achieves online joint optimization of communication computing resources in 6G network scenarios while maintaining controllable solution complexity.

[0070] Finally, system monitoring and evaluation were conducted, and quantization error tolerances were set under both IID and non-IID data distribution conditions to verify the dynamic adjustment capability and overall effectiveness.

[0071] Example 2 In one embodiment, to illustrate the effectiveness of the invention, model accuracy and training convergence time can be examined under different aggregation quantization error tolerances and different resource allocation strategies under both IID and non-IID data distribution conditions. Evaluation metrics may include test accuracy, training loss, and the time required for each training round. The evaluation content should include at least the following aspects: First, the impact of different aggregation quantization error tolerances on model accuracy and convergence time; Second, a comparison of the present invention with benchmark strategies such as fixed bit allocation, equal time slot allocation, and equal energy allocation; Third, the performance improvement of the adaptive aggregation quantization error tolerance strategy compared to the fixed error tolerance strategy.

[0072] Therefore, this invention achieves efficient federated learning communication and computing resource collaborative allocation for 6G network scenarios by transforming the model convergence accuracy requirement into an aggregate quantization error constraint and jointly designing it with the allocation of terminal computing resources and wireless communication resources.

Claims

1. A method for resource allocation integrating communication and computing in energy-aware 6G networks using stochastic quantization federated learning, characterized in that: Includes the following steps: S1. Construct a federated learning framework to collect state information from the device side; S2. Based on the status information, design a quantitative transmission mechanism to quantize the local model update amount generated by the device. S3. Based on the quantized local model update amount, construct the joint communication computation timing sequence to describe the total completion time and energy consumption of each round of training; S4. Based on the joint communication calculation time sequence, weigh the quantization error and aggregation impact, and give the quantized aggregation update; S5. Based on the quantized aggregate update, calculate the resource coordination optimization problem constrained by device energy, communication, and computing resources with the goal of minimizing the single-round completion time. S6. Based on the resource collaborative optimization problem, design an online solution implementation method for dual decomposition to dynamically solve the optimal quantization accuracy, transmission power, and computational resource allocation strategy of the system.

2. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S1 includes: Build a server and M A federated learning framework composed of multiple end users defines users. m Local loss function and global objective function Using learning rate And small-batch data updates, and collection of device-side status information; among which, users m In the i The model parameters after the first round of local training are: The amount of local model updates uploaded to the server is: in, Represents the current global model; Indicates user In the The small batch of samples selected in the step; user In the In the round of global federated training, the first The local model parameter vector obtained after step local stochastic gradient descent; This represents stochastic gradient estimation based on mini-batch samples.

3. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S2 includes: Introducing quantization operators , will users m In the i Wheelset local model update vector Each component adopts Each quantized bit is randomly quantized, and the lower and upper bounds of the quantization amplitude range are defined as 0 and 0, respectively. and in accordance with The system performs uniform segmentation, determines the quantization value of each component according to random rules, ensuring that the quantizer is unbiased for each component, and adjusts the quantization value according to the number of quantization bits. Calculate the amount of data uploaded by user m. .

4. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S3 includes: Construct the total completion time for each round of federated learning tasks. It consists of a local computing phase and an uplink transmission phase; among which, according to the user m allocated CPU frequency Take calculation time and its energy consumption According to the user m Transmission power and uplink transmission duration Computational communication energy consumption The minimum lower bound of the delay required to satisfy the communication constraints is obtained according to the Shannon capacity formula; uplink transmission is performed using time division multiple access, and the total communication delay is... The total latency is calculated as the sum of the transmission times of all users. It is determined by the slowest terminal.

5. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S4 includes: defining a user m In the i Wheel vector quantization error The server-side aggregation update relationship is as follows: in, This represents the percentage of the sample size. Represents the current global model; This represents the local model update vector; Represents the quantization operator; M Indicates the number of end users; Define aggregate quantization error: And a constraint is imposed on the second moment of the aggregated quantization error on the server side: in, Indicates the first The allowable aggregate quantization error tolerance of the round. It represents the mathematical expectation.

6. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S5 includes: [the following is a list of components] i Total completion time of each round of federal learning tasks Minimize as the optimization objective, calculate frequency Transmission power Transmission duration and the number of random quantization bits As a joint optimization variable, it satisfies the constraints of single-round energy budget, communication transmission rate, aggregated quantization error second moment, maximum CPU frequency, maximum transmission power, and the constraint that the number of quantization bits is a positive integer.

7. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 1, characterized in that, S6 includes: ... By performing continuous relaxation, the Lagrangian function of the single-round optimization problem is constructed. The coupling constraints are relaxed by constructing the dual problem. The original problem is decomposed into subproblems that optimize the computation frequency, transmission power and transmission duration, and subproblems that optimize the number of random quantization bits in combination with the aggregated quantization error tolerance. The Lagrangian multipliers corresponding to the energy budget constraint, communication constraint and aggregated error constraint are updated online based on the subgradient method or the alternating optimization method.

8. The energy-aware 6G network stochastic quantization federated learning communication and computing integrated resource allocation method according to claim 6, characterized in that, The single-round energy budget is determined by the terminal in the [number]th round. i The calculation frequency is determined by the remaining energy at the start of each round, the preset safe energy threshold, and the requirements for subsequent consecutive training rounds. Transmission power and the number of random quantization bits The feasible allocation range.

Images