Multi-task domain adaptive method for calculation unloading

Abstract

The invention discloses a multi-task domain self-adaptive method for computing unloading. The method comprises the steps that according to a scene that a user processes a target task in itself or unloads the target task to an edge server for processing, a target function is constructed, multiple sets of source domain data are obtained, each set of source domain data comprises sample data and labels of multiple users, and the labels are initial classification values and initial regression values; inputting the multiple groups of source domain data into the feedforward neural network model to obtain a trained feedforward neural network model; acquiring multiple groups of target domain data, and enhancing the multiple groups of target domain data to obtain multiple groups of enhanced target domain data; wherein each group of target domain data comprises sample data of a plurality of users. The technical problems that an existing task unloading model is low in generalization ability, high in privacy risk, large in calculation amount and long in calculation time in different domain environments are solved.

Description

Technical Field

[0001] The present invention relates to the field of resource computing technology, and in particular to a multi-task domain adaptation method for computing offloading. Background Art

[0002] With the rapid development of Multi-access Edge Computing (MEC) technology, its role in reducing network latency and improving computing efficiency has been widely recognized. Multi-access edge computing can effectively alleviate the performance bottleneck of mobile devices due to limited computing resources by deploying computing resources at the edge of the network; especially in the context of the popularization of mobile communications and smart terminals, MEC can offload computing-intensive tasks from resource-constrained terminal devices to edge servers through task offloading, thereby improving overall computing efficiency and reducing the load pressure on terminal devices.

[0003] In the MEC environment, the task offloading problem can usually be divided into two sub-problems: task offloading decision-making and computing resource allocation; in order to cope with the dynamically changing network environment, swarm intelligence networks have gradually become an important means of edge computing; by integrating the collective intelligence of multiple mobile users and edge devices, edge intelligent networks achieve collaborative decision-making and resource sharing, thereby optimizing task offloading decisions in complex and changing network environments; however, due to the constant changes in the network environment, the data distribution in the target domain is often significantly different from that in the source domain, which makes it difficult for traditional mathematical model-based task offloading methods to be effectively generalized in new environments.

[0004] In order to improve the generalization ability of task offloading models between different domains, artificial intelligence technologies such as Deep Reinforcement Learning (DRL) have also been widely used in MEC systems in recent years, hoping to learn the optimal task offloading strategy by dynamically interacting with the wireless environment; however, traditional DRL methods have poor adaptability when facing unexpected disturbances and require comprehensive retraining, resulting in high application costs in new environments.

[0005] Although existing technologies have made some progress in MEC task offloading, there are still some significant shortcomings and deficiencies: Poor adaptability: Traditional computing task offloading models usually rely on pre-trained mathematical models or neural networks for decision-making. These models overfit the source domain data during the training process. When the target domain data deviates significantly from the source domain, the generalization ability of the model is weak, resulting in a significant decrease in reasoning accuracy.

[0006] Privacy issues: In MEC systems, source domain data usually involves sensitive information of users. Traditional methods need to access source domain data during the reasoning process, which may lead to user privacy leakage. Moreover, accessing source domain data also brings a large storage and computing burden.

[0007] High retraining cost: In order to adapt to changes in different network environments, existing technologies usually require retraining of models, which not only consumes a lot of computing resources but also takes a long time, making it difficult to meet the real-time and high-efficiency requirements of the MEC system.

[0008] Low resource utilization efficiency: As the existing models fail to effectively consider the dynamic changes of the network environment, they suffer from low resource utilization efficiency in terms of resource allocation and task offloading decisions. This not only reduces the computing resource utilization of edge servers, but may also increase task delays and affect user experience. Summary of the invention

[0009] The embodiments of the present invention provide a multi-task domain adaptation method for computation offloading, so as to at least solve the technical problems of low generalization ability, high privacy risk, large computation amount and long computation time of the existing task offloading model in different domain environments.

[0010] According to one aspect of an embodiment of the present invention, a multi-task domain adaptation method for computational offloading is provided. The method may include: constructing an objective function according to a scenario in which a user processes a target task by themselves or offloads it to an edge server for processing; obtaining multiple sets of source domain data, each set of source domain data includes sample data and labels of multiple users, and the labels are initial classification values ​​and initial regression values; inputting the multiple sets of source domain data into a feedforward neural network model to obtain a trained feedforward neural network model; obtaining multiple sets of target domain data, enhancing the multiple sets of target domain data to obtain multiple sets of enhanced target domain data; wherein each set of target domain data includes sample data of multiple users; constructing a teacher-student model, wherein the teacher-student model includes a student model and a teacher model, and the trained feedforward neural network model is used as the teacher model and the student model respectively; inputting the enhanced first set of target domain data into the teacher model to obtain pseudo labels for the sample data of each user, wherein the pseudo labels include a first classification prediction value and a first regression prediction value; A set of target domain data is input into the student model to obtain an initial prediction value of the sample data of each user, wherein the initial prediction value includes a second classification prediction value and a second regression prediction value; based on the initial prediction value and the pseudo label, a total loss value is obtained; based on the total loss value, the model parameters of the student model are updated, and iterations are repeated. When the total loss value is minimized, the model parameters of the teacher model are updated with the model parameters of the student model when the total loss value is minimized, and the model parameters of the student model when the total loss value is minimized are randomly restored to obtain the restored model parameters, and the restored model parameters are used as the student model corresponding to the second set of target domain data, and iterations are repeated to obtain a trained student model; each set of target domain data is input into the trained target student model to obtain a target prediction value for each set of target domain data; based on the target prediction value of each set of target domain data and the target domain data corresponding to the target prediction value, the objective function is minimized to minimize the user's computing task cost.

[0011] Optionally, the loss function of the trained feedforward neural network model is expressed as:

[0012] in, is the total loss value of the trained feedforward neural network model, is the weight of the classification loss of the trained feedforward neural network model, is the classification loss value of the trained feedforward neural network model, is the weight of the regression loss of the trained feedforward neural network model, is the regression loss value of the trained feedforward neural network model.

[0013] Optionally, the classification loss value of the trained feedforward neural network model is expressed as:

[0014] Among them, M is the total number of source domain data groups, is one of multiple sets of source domain data, is the classification prediction value of each set of source domain data in multiple sets of source domain data output by the trained feedforward neural network model. is the initial classification value of each set of source domain data in multiple sets of source domain data.

[0015] Optionally, the regression loss of the trained feedforward neural network model is expressed as:

[0016] in, is the regression loss of the trained feedforward neural network model.

[0017] Optionally, obtaining a total loss value based on the initial prediction value and the pseudo label includes: obtaining a cross entropy loss value based on the first classification prediction value and the second classification prediction value; obtaining a mean square error loss value based on the first regression prediction value and the second regression prediction value; and obtaining a total loss value based on the cross entropy loss value and the mean square error loss value.

[0018] Optionally, the expression for obtaining the total loss value based on the cross entropy loss value and the mean square error loss value is:

[0019] in, represents the total loss value, is the cross entropy loss value, is the mean square error loss value, is the first hyperparameter, is the second hyperparameter.

[0020] Optionally, the model parameters of the student model when the total loss value is minimized are randomly restored, and the expression of the restored model parameters is obtained as follows:

[0021] in, are the restored model parameters, is the mask tensor, are the model parameters of the feedforward neural network model, represents the convolution operation, The model parameters of the student model when the total loss value is minimized for random recovery.

[0022] Beneficial effects of the present invention: (1) By introducing the teacher-student architecture, there is no need to access source domain data during the inference phase. This not only effectively protects user privacy and avoids the leakage of sensitive information, but also reduces the burden of storage and computing, making the system more lightweight and secure. The design of this architecture reduces the dependence on source domain data, thereby significantly reducing the overhead of data transmission and storage.

[0023] (2) By adopting a multi-task learning framework, the present invention can simultaneously handle task offloading decisions and computing resource allocation problems; compared with the traditional single-task model, this framework improves the resource utilization efficiency of the system, enabling the edge server to more reasonably allocate computing resources in a dynamically changing environment, thereby reducing task delays and improving user experience; at the same time, the joint training of classification and regression tasks enables the model to better cope with complex environments and improve adaptability.

[0024] (3) The student model is continuously updated during the inference phase through a dynamic adaptive adjustment mechanism to adapt to changes in the target domain data. This mechanism reduces the reliance on comprehensive retraining, allowing the model to adapt to new environments more quickly, improving generalization capabilities and computational efficiency. This means that the present invention can be quickly deployed and adjusted in different network environments, reducing the cost of system maintenance and updates.

[0025] (4) By combining the cross entropy loss of the classification task and the mean square error loss of the regression task, an optimized weighted loss function is designed, so that the model can strike a balance between the accuracy of task offloading decisions and resource allocation; by reasonably adjusting the weights, the model can take into account multiple objectives at the same time during the optimization process, improving the overall performance, thereby making edge computing perform better in multi-task scenarios.

[0026] (5) By introducing a random recovery mechanism into the student model, the present invention effectively prevents the catastrophic forgetting problem that may occur during the long-term adaptive process. This mechanism partially restores the parameters of the student model to the initial state, ensuring that the model can still maintain its ability to understand the original task during the long-term learning process, thereby achieving low-cost continuous learning. This design not only improves the adaptability of the model, but also significantly reduces the cost of retraining, allowing the system to remain efficient and stable during long-term operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 is a flowchart of a multi-task domain adaptation method for computation offloading according to an embodiment of the present invention; Figure 2It is a framework diagram of a multi-task domain adaptation method for computational offloading according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0029] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] Example 1 According to an embodiment of the present invention, a multi-task domain adaptation method for computational offloading is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system comprising at least one set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0031] Figure 1 is a flowchart of a multi-task domain adaptation method for computation offloading according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps: Step S101 : constructing an objective function according to a scenario in which a user processes a target task by himself or offloads the target task to an edge server for processing.

[0032] In the technical solution provided in step S101 of the present invention, in a single-server multi-user MEC system, it is assumed that Mobile users, that is , the N-dimensional offloading decision vector of mobile users to the edge server (Mobile Edge Server, referred to as MES) is expressed as , respectively Indicates that the user offloads the task to the edge server. Indicates that the user processes tasks locally, and the N-dimensional vector of MES computing resource allocation is , and ensure , that is, MES allocates computing resources to users who need resources, and the computing resources allocated to all users do not exceed the computing resources of MES itself.

[0033] set up Indicates user The CPU cycle frequency, Indicates processing task The required CPU cycle frequency, Indicated in The total cost of local processing on It indicates the energy efficiency parameter that mainly depends on the hardware chip architecture. Indicates emphasis on computing latency and power consumption; local computing The weighted cost calculation method is:

[0034] set up Indicate to upload The data size, Assign to user for MES of computing resources, express After being processed by the edge server, the size of the processing result returned to the user. express from The data rate achieved by the wireless uplink to the MES, express From MES to The data rate achieved by wireless downlink, for The power consumption uploaded from the user to the edge server, Execute for edge servers The power consumption, The power consumption of the edge server when it returns the processing result to the user. The weighted cost calculation method for unloading to MES is:

[0035] The total cost of the MEC system can be defined as the weighted sum of the costs of all mobile users:

[0036] The computational offloading problem is formulated as a joint optimization problem of offloading decision and computational resource allocation, aiming to minimize the overall cost of the system. Under the fluctuation conditions of the network and the constraints of computational resources, the offloading decision and computational resource allocation are formulated as a mixed integer nonlinear programming problem. The joint offloading and computational resource allocation are formulated as a weighted sum cost minimization problem, that is, the objective function: .

[0037] Step S102, obtaining multiple groups of source domain data, each group of source domain data includes sample data and labels of multiple users, and the labels are initial classification values ​​and initial regression values.

[0038] In the technical solution provided in the above step S102 of the present invention, multiple groups of source domain data: first, randomly generate user input features such as data volume, calculation cycle, local calculation frequency, channel gain, etc. according to a set distribution (such as normal distribution); then use the optimization tool GEKKO to construct an offloading decision and resource allocation problem, the goal is to minimize the total cost (including local calculation, transmission and server calculation costs) and meet the delay constraint; then generate each user's offloading decision and resource allocation ratio as the target variable according to the optimization result; finally, save the generated input features and optimization results as a tabular data file for machine learning model training, and record the configuration parameters of the generated data for reproduction.

[0039] Step S103, inputting multiple groups of source domain data into the feedforward neural network model to obtain a trained feedforward neural network model.

[0040] In the technical solution provided in the above step S103 of the present invention, the feedforward neural network model is trained on multiple groups of source domain data to obtain a trained feedforward neural network model.

[0041] Step S104, acquiring multiple groups of target domain data, and enhancing the multiple groups of target domain data to obtain multiple groups of enhanced target domain data; wherein each group of target domain data includes sample data of multiple users.

[0042] In the technical solution provided in the above step S104 of the present invention, multiple groups of target domain data: randomly generate input features of nodes according to a distribution different from that set in the source domain (such as normal distribution), such as data volume, calculation cycle, local calculation frequency, channel gain, etc., and save the generated input features as a tabular data file for use in machine learning.

[0043] Step S105, constructing a teacher-student model, wherein the teacher-student model includes a student model and a teacher model, and the trained feedforward neural network model is used as the teacher model and the student model respectively.

[0044] In the technical solution provided in the above step S105 of the present invention, Figure 2is a framework diagram of a multi-task domain adaptation method for computation offloading according to an embodiment of the present invention, such as Figure 2 The teacher model and the student model in .

[0045] Step S106, inputting the enhanced first set of target domain data into the teacher model to obtain pseudo labels for the sample data of each user, wherein the pseudo labels include a first classification prediction value and a first regression prediction value.

[0046] In the technical solution provided in step S106 of the present invention, if Figure 2 As shown, multiple groups of target domain data are enhanced, and the enhanced first group of target domain data is input into the teacher model to obtain the pseudo label of each user's sample data, that is, the classification prediction value (first classification prediction value) and regression prediction value (first regression prediction value) behind the teacher model.

[0047] Step S107, inputting the first set of target domain data into the student model to obtain an initial prediction value of each user's sample data, wherein the initial prediction value includes a second classification prediction value and a second regression prediction value.

[0048] In the technical solution provided in step S107 of the present invention, if Figure 2 As shown, the student model processes the first set of target domain data to obtain the initial prediction value of each user's sample data, which is Figure 2 The classification prediction value (second classification prediction value) and regression prediction value (second regression prediction value) behind the middle school student model.

[0049] Step S108, obtaining a total loss value based on the initial prediction value and the pseudo label.

[0050] In the technical solution provided in the above step S108 of the present invention, the initial prediction value and the pseudo label are calculated to obtain a total loss value.

[0051] Step S109, based on the total loss value, update the model parameters of the student model, repeat the iteration, when the total loss value is the minimum, use the model parameters of the student model when the total loss value is the minimum to update the model parameters of the teacher model, randomly restore the model parameters of the student model when the total loss value is the minimum, obtain the restored model parameters, use the restored model parameters as the student model corresponding to the second set of target domain data, repeat the iteration, and obtain the trained student model.

[0052] In the technical solution provided in step S109 of the present invention, if Figure 2As shown in , based on the total loss value, the model parameters of the student model are updated, the total loss value is back-propagated, and the gradient is calculated. When the total loss value is minimized, the model parameters of the student model when the total loss value is minimized are used to update the model parameters of the teacher model. The model parameters of the student model when the total loss value is minimized are randomly restored to obtain the restored model parameters. The restored model parameters are used as the student model corresponding to the second set of target domain data. When all Epoch training is completed, the trained student model is obtained, as shown in Figure 2 As shown, the model parameters of the student model when the total loss value is minimized are used to update the model parameters of the teacher model. This is the model parameters of the student model when the total loss value is minimized without random recovery.

[0053] The expression for updating the model parameters of the teacher model with the model parameters of the student model when the total loss value is minimized is:

[0054] in, for The model parameters of the teacher model at the time step, that is, the model parameters of the updated teacher model, is used to represent a coefficient that controls the proportion of weights provided to the teacher model during the averaging process, for The model parameters of the teacher model at the time step, that is, the model parameters of the teacher model that have not been updated, for The model parameters of the student model at the time step, that is, the model parameters of the student model when the total loss value is the smallest.

[0055] Step S110, input each set of target domain data into the trained target student model to obtain the target prediction value of each set of target domain data.

[0056] In the technical solution provided in the above step S110 of the present invention, each set of target domain data is input into the trained target student model to obtain the target prediction value of each set of target domain data.

[0057] Step S111, based on the target prediction value of each group of target domain data and the target domain data corresponding to the target prediction value, minimize the objective function to minimize the user's computing task cost.

[0058] In the technical solution provided by step S111 of the present invention, the target prediction value of each group of target domain data and the target domain data corresponding to the target prediction value obtained in step S110 are substituted into In the process of solving the problem, the minimum cost of the user's computing task is obtained.

[0059] The above method of this embodiment is further introduced below.

[0060] As an optional implementation mode, in step S103, the loss function of the trained feedforward neural network model is expressed as:

[0061] in, is the total loss value of the trained feedforward neural network model, is the weight of the classification loss of the trained feedforward neural network model, is the classification loss value of the trained feedforward neural network model, is the weight of the regression loss of the trained feedforward neural network model, is the regression loss value of the trained feedforward neural network model.

[0062] As an optional implementation method, the classification loss value of the trained feedforward neural network model is expressed as:

[0063] Among them, M is the total number of source domain data groups, is one of multiple sets of source domain data, is the classification prediction value of each set of source domain data in multiple sets of source domain data output by the trained feedforward neural network model. is the initial classification value of each set of source domain data in multiple sets of source domain data.

[0064] As an optional implementation mode, the regression loss expression of the trained feedforward neural network model is:

[0065] in, is the regression loss of the trained feedforward neural network model.

[0066] As an optional implementation method, step S108, the total loss value is obtained based on the initial prediction value and the pseudo label, including: obtaining a cross entropy loss value based on the first classification prediction value and the second classification prediction value; obtaining a mean square error loss value based on the first regression prediction value and the second regression prediction value; obtaining a total loss value based on the cross entropy loss value and the mean square error loss value.

[0067] In this embodiment, the first classification prediction value and the second classification prediction value are calculated to obtain a cross entropy loss value; the first regression prediction value and the second regression prediction value are calculated to obtain a mean square error loss value; the cross entropy loss value and the mean square error loss value are calculated to obtain a total loss value.

[0068] As an optional implementation manner, the expression for obtaining the total loss value based on the cross entropy loss value and the mean square error loss value is:

[0069] in, represents the total loss value, is the cross entropy loss value, is the mean square error loss value, is the first hyperparameter, is the second hyperparameter.

[0070] In this embodiment, the cross entropy loss value, the mean square error loss value, the first hyperparameter, and the second hyperparameter are calculated to obtain a total loss value.

[0071] Among them, the expression of the cross entropy loss value is:

[0072] in, is the first classification prediction value, that is, the first classification prediction value predicted by the student model for the sample data of the i-th user. is the second classification prediction value, that is, the second classification prediction value predicted by the teacher model for the sample data of the i-th user.

[0073] The expression of mean square error loss value is:

[0074] in, is the first regression prediction value, that is, the first regression prediction value predicted by the student model for the sample data of the i-th user. is the second regression prediction value, that is, the second regression prediction value predicted by the teacher model for the sample data of the i-th user.

[0075] As an optional implementation manner, in step S109, the model parameters of the student model when the total loss value is minimized are randomly restored, and the expression of the restored model parameters is obtained as follows:

[0076] in, are the restored model parameters, is the mask tensor, are the model parameters of the feedforward neural network model, represents the convolution operation, The model parameters of the student model when the total loss value is minimized for random recovery.

[0077] In this embodiment, the mask tensor By a probability The Bernoulli distribution is generated, that is, .

[0078] In an embodiment of the present invention, an objective function is constructed according to a scenario in which a user processes a target task by himself or offloads it to an edge server for processing; multiple sets of source domain data are obtained, each set of source domain data includes sample data and labels of multiple users, and the labels are initial classification values ​​and initial regression values; the multiple sets of source domain data are input into a feedforward neural network model to obtain a trained feedforward neural network model; multiple sets of target domain data are obtained, and the multiple sets of target domain data are enhanced to obtain multiple sets of enhanced target domain data; each set of target domain data includes sample data of multiple users; a teacher-student model is constructed, wherein the teacher-student model includes a student model and a teacher model, and The trained feedforward neural network model is used as the teacher model and the student model respectively; the enhanced first set of target domain data is input into the teacher model to obtain the pseudo-label of each user's sample data, wherein the pseudo-label includes the first classification prediction value and the first regression prediction value; the first set of target domain data is input into the student model to obtain the initial prediction value of each user's sample data, wherein the initial prediction value includes the second classification prediction value and the second regression prediction value; based on the initial prediction value and the pseudo-label, the total loss value is obtained; based on the total loss value, the model parameters of the student model are updated, and the iteration is repeated. When the total loss value is minimized, the model parameters of the student model with the minimum total loss value are used. The model parameters of the teacher model are updated randomly, and the model parameters of the student model when the total loss value is minimized are randomly restored to obtain the restored model parameters. The restored model parameters are used as the student model corresponding to the second set of target domain data, and repeated iterations are performed to obtain a trained student model; each set of target domain data is input into the trained target student model to obtain the target prediction value of each set of target domain data; based on the target prediction value of each set of target domain data and the target domain data corresponding to the target prediction value, the objective function is minimized to minimize the user's computing task cost, which solves the problems of low generalization ability, high privacy risk, large computational complexity and computing cost of existing task offloading models in different domain environments. In order to solve the technical problem of long computing time, we have achieved efficient task offloading and computing resource allocation through teacher-student architecture, multi-task learning framework, dynamic adaptive adjustment, optimized loss function design and low-cost continuous learning. In the inference stage, the teacher model guides the student model to achieve continuous adaptation without the need for source domain data, protecting user privacy and reducing computing costs. Multi-task learning integrates task offloading decisions and resource allocation, and jointly optimizes to improve adaptability. Dynamic adjustment enables the student model to adapt to changes in the target domain, and the optimized weighted loss function takes into account both classification and regression performance. In addition, the random recovery mechanism prevents catastrophic forgetting in long-term self-learning.

[0079] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0080] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0082] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0083] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0084] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-task domain adaptation method for computation offloading, characterized in that: include: Construct the objective function based on the scenario where the user processes the target task on their own or offloads it to the edge server for processing; Acquire multiple sets of source domain data, each set of source domain data includes sample data and labels of multiple users, where the labels are initial classification values ​​and initial regression values; Inputting multiple sets of source domain data into the feedforward neural network model to obtain a trained feedforward neural network model; Acquire multiple sets of target domain data, and enhance the multiple sets of target domain data to obtain multiple sets of enhanced target domain data; wherein each set of target domain data includes sample data of multiple users; Constructing a teacher-student model, wherein the teacher-student model includes a student model and a teacher model, and using the trained feedforward neural network model as the teacher model and the student model respectively; Inputting the enhanced first set of target domain data into the teacher model to obtain pseudo labels of sample data of each user, wherein the pseudo labels include a first classification prediction value and a first regression prediction value; Inputting the first set of target domain data into the student model to obtain an initial prediction value of each user's sample data, wherein the initial prediction value includes a second classification prediction value and a second regression prediction value; Based on the initial prediction value and pseudo label, the total loss value is obtained; Based on the total loss value, update the model parameters of the student model, repeat the iteration, when the total loss value is the minimum, use the model parameters of the student model when the total loss value is the minimum to update the model parameters of the teacher model, randomly restore the model parameters of the student model when the total loss value is the minimum, obtain the restored model parameters, use the restored model parameters as the student model corresponding to the second set of target domain data, repeat the iteration, and obtain the trained student model; Input each set of target domain data into the trained target student model to obtain the target prediction value of each set of target domain data; Based on the target prediction value of each set of target domain data and the target domain data corresponding to the target prediction value, the objective function is minimized to minimize the user's computing task cost.

2. The method according to claim 1, characterized in that The loss function of the trained feedforward neural network model is expressed as: in, is the total loss value of the trained feedforward neural network model, is the weight of the classification loss of the trained feedforward neural network model, is the classification loss value of the trained feedforward neural network model, is the weight of the regression loss of the trained feedforward neural network model, is the regression loss value of the trained feedforward neural network model.

3. The method according to claim 2, characterized in that The expression of the classification loss value of the trained feedforward neural network model is: Among them, M is the total number of source domain data groups, is one of multiple sets of source domain data, is the classification prediction value of each set of source domain data in multiple sets of source domain data output by the trained feedforward neural network model. is the initial classification value of each set of source domain data in multiple sets of source domain data.

4. The method according to claim 3, characterized in that The expression of the regression loss of the trained feedforward neural network model is: in, is the regression loss of the trained feedforward neural network model.

5. The method according to claim 1, characterized in that Based on the initial prediction value and pseudo label, the total loss value is obtained, including: Based on the first classification prediction value and the second classification prediction value, a cross entropy loss value is obtained; Based on the first regression prediction value and the second regression prediction value, a mean square error loss value is obtained; Based on the cross entropy loss value and the mean square error loss value, the total loss value is obtained.

6. The method according to claim 5, characterized in that The expression of the total loss value based on the cross entropy loss value and the mean square error loss value is: in, represents the total loss value, is the cross entropy loss value, is the mean square error loss value, is the first hyperparameter, is the second hyperparameter.

7. The method according to claim 1, characterized in that The model parameters of the student model when the total loss value is minimized by random recovery are obtained, and the expression of the restored model parameters is: in, are the restored model parameters, is the mask tensor, are the model parameters of the feedforward neural network model, represents the convolution operation, The model parameters of the student model when the total loss value is minimized for random recovery.

8. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

9. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

10. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when being executed.

Images