A task offloading method based on a stackelberg game driven by a LLM

Abstract

This invention discloses a task offloading method driven by LLM and based on Stackelberg game theory, belonging to the field of distributed computing and artificial intelligence applications. The method includes: first, LLM analyzes the task and predicts resource vectors; the platform infers the private cost function of the computing power provider using Bayesian methods and IRL; second, a Stackelberg game model of the platform and computing power provider is constructed, and the equilibrium is solved through backward induction; the platform sets personalized prices, and the computing power provider provides optimal resources; qualified resource screening and LLM scoring prevent computing power providers from concealing or misreporting resources; finally, the LLM enters the "reasoning-action-scoring" stage to schedule tasks, achieving a closed-loop optimization system. The advantages of this invention compared to existing technologies are: this invention can systematically solve the problem of dual information asymmetry caused by the uncertainty of tasks and computing power providers, achieving an efficient computing power provider incentive mechanism and task scheduling mechanism.

Description

Technical Field

[0001] This invention relates to an incentive and scheduling framework for distributed computing networks. Specifically, it utilizes artificial intelligence and game theory to systematically solve the incentive and scheduling problem of resource allocation in socialized edge computing networks, belonging to the field of distributed computing and artificial intelligence applications. Background Technology

[0002] With the booming development of applications such as the Internet of Things (IoT) and Augmented Reality (AR), the computing paradigm is shifting from centralized cloud computing to distributed edge computing. This shift has given rise to a vast but underutilized pool of resources—a "socialized" computing resource network composed of devices such as personal computers and smartphones. Therefore, effectively integrating this idle computing power can bring dual benefits in terms of economy and performance.

[0003] However, building an efficient social computing market requires overcoming a major challenge: dual information asymmetry. This asymmetry manifests on two levels: task uncertainty and computing power provider uncertainty. Task uncertainty means that users are unaware of the resources required to complete their tasks, and computing power platforms, upon receiving a task, cannot know in advance, precisely, the actual resource requirements for its execution. Computing power provider uncertainty means that the platform also cannot know the private cost function of each social edge node in the network. As rational economic participants, computing power providers have an incentive to hide or overstate their true costs to obtain greater profits. This dual uncertainty renders existing economic models and algorithms, such as VCG auctions and simple machine learning predictions, ineffective. They either assume that task requirements and computing power provider information are known, or that computing power providers are honest, making it impossible to handle both types of uncertainty simultaneously. Currently, there is no scientific solution to address the dual information asymmetry problem in "social" computing to achieve efficient incentive mechanisms for computing power providers and task scheduling mechanisms. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes an LLM-driven task offloading method based on Stackelberg game theory. The specific details are as follows:

[0005] Step 1: Analyze task resource requirements and construct a task dependency graph based on LLM;

[0006] LLM parses the computational tasks submitted by users, identifies key algorithms, computational patterns, dependencies, and resource-intensive operations within the tasks, and outputs a multi-dimensional predicted resource vector based on the above analysis. and task dependency graph and based on Funds required for the forecast task :

[0007] ,

[0008] in, For the predicted CPU cycles, Peak memory usage, Total data I / O For the required GPU size, Score the parallelizability potential of the task.

[0009] ,

[0010] in, It is a collection of subtasks obtained by LLM after decomposing the original task. Each subtask Each is associated with a subtask resource requirement vector predicted by LLM. .vector The structure and the total resource vector defined above Correspondingly. It is a set of dependency edges representing the order of subtasks.

[0011] Step 2: Estimate the computing power provider cost function using a "two-stage" hybrid strategy based on Bayesian inference and IRL;

[0012] 1. To address the "cold start" problem in cost estimation, for newly added computing power providers, the platform utilizes a Bayesian approach, combining the provider's equipment type and prior information on the cost functions of previous computing power providers in the region, to quickly generate an initial cost estimate. The platform defines a set of cost function parameters. And set a prior probability distribution for new computing power providers. .

[0013] 1) When prior information is lacking, the distribution is set to uniform, i.e., for all... ;

[0014] 2) When prior information is available, the probability mass is assigned to the parameter that is closest to the prior information. Specifically, the platform constructs a device feature vector based on the device type and location coordinates of the computing power provider. and position vector Calculate the mixed distance between historical computing power providers and newly added computing power providers. :

[0015] ,

[0016] The cost parameters of the historical computing power provider with the smallest distance value are assigned to the newly added computing power provider. For Euclidean distance, Weighting coefficients ( ).

[0017] 2. Computing power providers In pricing The actual amount of resources provided , constitute observation data Using Bayes' theorem, according to Calculate the posterior probability distribution and update the cost function parameters. :

[0018] .

[0019] The platform is based on the posterior probability distribution To form the final estimate of the cost function. .

[0020] 3. Computing power providers With computing power platform at every historical moment All interactions will generate state-movement trajectory data. As the interaction data becomes richer, the platform optimizes the cost function estimation using the IRL algorithm. At this point, the probability of an action is modeled as:

[0021] ,

[0022] in, It is an estimate of the utility of the computing power provider. This is the normalization constant. These are learnable parameters. The learning process is achieved by maximizing the log-likelihood of the observed data:

[0023] .

[0024] Step 3: Construct a Stackelberg game between the platform and the computing power provider, and solve for the equilibrium using backward induction;

[0025] 1. The platform's goal is to select and publish a personalized set of pricing functions. In order to maximize its own utility The platform's utility equals the total revenue collected from task initiators. Subtract the total incentive cost paid to all edge nodes:

[0026] .

[0027] Each computing power provider As a follower in the game, upon receiving the exclusive pricing function published by the platform... Then, select the amount of resources you want to provide. In order to maximize its net income Its net benefit equals the reward it receives minus the private cost of providing the resources. :

[0028] ,

[0029] .

[0030] 2. As the number of interactions increases, the learning from the "two-stage" hybrid strategy... It will become increasingly closer to the true nature of computing power providers. Therefore, through Solve Obtain the estimated computing power provider Optimal response function .

[0031] 3. The platform will provide response functions for all predicted computing power providers. Substitute your own utility function : Meanwhile, the platform's global optimization issues can be broken down into: This is a separate optimization problem. The platform provides nodes for each computing power. Solve the following problems separately to design the optimal personalized price:

[0032] ,

[0033] .

[0034] Each computing power provider According to exclusive and its true cost To make its actual optimal resource supply :

[0035] ,

[0036] Finally, the strategy combination is obtained. .

[0037] 4. The platform calculates... and Deviation between And set an acceptable deviation threshold. To establish a resource pool :

[0038] ,

[0039] This resource pool initially screens out computing power providers whose resource deviations are within the threshold and whose computing power providers meet the minimum resource requirements of the sub-tasks. At the same time, it allows for appropriate fluctuations in the cost of computing power resource providers and prevents them from exaggerating or concealing information in order to obtain greater profits.

[0040] Step 4: LLM uses the task dependency graph to enter the "reasoning-action-scoring" stage to allocate tasks and provide feedback for optimization;

[0041] 1. The platform maintains the communication latency matrix of computing power providers in the qualified resource supply pool. Reliability rating of computing power providers The details are as follows:

[0042] ,

[0043] ,

[0044] Among them, when When =0, Delay matrix Medium computing power providers Communication latency to the platform, when and hour, Delay matrix Medium computing power providers To computing power providers Communication latency, For LLM computing power providers Historical rating values, initial Set it to 100.

[0045] 2. LLM acts as a heuristic solver, entering the "reasoning-action-scoring" cycle:

[0046] 1) LLM obtains the task dependency graph Subtask vectors ,resource Pricing Delay matrix Reliability rating of computing power providers information.

[0047] 2) Based on the information obtained, LLM determines the priority of computing power providers. :

[0048] ,

[0049] in These correspond to the weights of latency, reliability, and payment to computing power providers, respectively, with all weights placed on... Take values ​​within the range, and satisfy the following conditions: .

[0050] 3) LLM analysis of the resource requirements of currently executable subtasks and according to priority Assigned to China satisfies The LLM continuously monitors the task execution status, and if a task fails, the LLM, based on its understanding of the feedback information (such as insufficient video memory), reassigns the task to a new computing power provider.

[0051] 4) After the task is completed, the computing power provider is evaluated based on the LLM. Task completion rating Update its reliability score ,in Weight value, this value in Within the range.

[0052] 3. Construct resource prediction data pairs in the format of <task description, actual resource consumption>, and task inference data pairs in the format of <task dependency graph, task completion status>. Use these data to fine-tune the LLM model to optimize its prediction and decision-making capabilities. Simultaneously, incorporate the data generated during the game (…). This constitutes new "state-action" trajectory data, which is used to update the cost estimation function in the hybrid strategy and achieve closed-loop optimization of the system.

[0053] In summary, the advantages of this invention are as follows:

[0054] 1) To address the "uncertainty" of user-submitted tasks, this invention utilizes an LLM (Limited Resource Management) model to parse the required resource vectors and subtask dependency graphs from the submitted tasks. Simultaneously, the LLM acts as a high-level heuristic solver, assigning subtasks to the currently optimal computing power provider. The LLM monitors the task execution status, and when a subtask fails, it can reassign the task to a new qualified computing power provider based on its understanding of the feedback. After task completion, data is generated for fine-tuning the LLM, and data generated during the game process is also used to update the cost inference model, achieving closed-loop optimization of the system.

[0055] 2) This invention utilizes a dual guarantee mechanism of real-time resource screening and LLM scoring to avoid fraud by computing power providers. Specifically, by judging whether the resources provided by the computing power provider are within the threshold and meet the task resource requirements, qualified computing power resource providers are initially screened. This mechanism allows for appropriate fluctuations in provider costs and provided resources, while effectively preventing providers from providing fewer resources in order to obtain greater profits, thereby screening out qualified computing power providers. LLM scoring is introduced for long-term reputation management. When scheduling tasks, LLM will select computing power providers based on this reliability score, thereby "punishing" dishonest or poorly performing computing power providers and incentivizing them to maintain honest and high-quality service. Attached Figure Description

[0056] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0057] Figure 1 This is a schematic diagram of an embodiment of the present invention;

[0058] Figure 2 This is the overall flowchart of the present invention;

[0059] Figure 3 This is a schematic diagram of the equilibrium between the computing power platform and the computing power provider based on the Stackelberg game in this invention;

[0060] Figure 4 This is a schematic diagram of LLM task allocation in an embodiment of the present invention. Detailed Implementation

[0061] This invention designs an LLM-driven task offloading method based on Stackelberg game theory, such as... Figure 1 As shown, the specific implementation method is as follows:

[0062] Step 1: Analyze task resource requirements and construct a task dependency graph based on LLM:

[0063] Users submit computing tasks to the computing power platform, and the computing power platform receives the tasks. Subsequently, the computing platform uses LLM to analyze the computing tasks submitted by users, identifying key algorithms, computing patterns, dependencies, and resource-intensive operations within the tasks. Based on this analysis, it outputs a multi-dimensional predicted resource vector. and task dependency graph And the funds required for the prediction task :

[0064]

[0065] in, For the predicted CPU cycles, Peak memory usage, For the required GPU size, Total data I / O Score the parallelizability potential of the task.

[0066]

[0067] in, It is a collection of subtasks obtained by LLM after decomposing the original task. Each subtask Each is associated with a subtask resource requirement vector predicted by LLM. .vector The structure and the total resource vector defined above Correspondingly. It is a set of dependency edges representing the order of subtasks.

[0068] Step 2: Estimating the computing power provider cost function using a two-stage hybrid strategy based on Bayesian inference and IRL:

[0069] 1. For newly added computing power providers, due to the lack of behavioral data, the platform uses a Bayesian method, combined with the provider's equipment type and prior information on the cost functions of previous computing power providers in the region, to quickly form an initial cost estimate, thereby solving the "cold start" problem of cost inference. The platform first defines a set of cost function parameters. And set a prior probability distribution for new computing power providers. .

[0070] 2) When prior information is lacking, the distribution is set to uniform, i.e., for all... ;

[0071] 3) When prior information is available, the probability mass is assigned to the parameter that is closest to the prior information. Specifically, the platform constructs a device feature vector based on the device type and location coordinates of the computing power provider. and position vector Calculate the mixed distance between historical computing power providers and newly added computing power providers. :

[0072]

[0073] The cost parameters of the historical computing power provider with the smallest distance value are assigned to the newly added computing power provider. For Euclidean distance, Weighting coefficients ( ).

[0074] 2. Computing power providers In pricing The actual amount of resources provided , constitute observation data Using Bayes' theorem, according to Calculate the posterior probability distribution and update the cost function parameters. :

[0075]

[0076] The platform is based on the posterior probability distribution To form the final estimate of the cost function. .

[0077] 3. Computing power providers With computing power platform at every historical moment All interactions will generate state-movement trajectory data. As data becomes more abundant, the platform utilizes this data to optimize cost function estimation using the IRL algorithm. The goal of the IRL algorithm is to find a cost function... This makes the observed historical actions under this cost function... Able to maximize the utility function of computing power providers That is, the cost function is closest to the actual cost function of the computing power provider. At this point, the probability of an action is modeled as:

[0078]

[0079] in, It is an estimate of the utility of the computing power provider. These are learnable parameters. The learning process is achieved by maximizing the log-likelihood of the observed data:

[0080]

[0081] Step 3: Construct a Stackelberg game between the platform and the computing power provider, and solve for the equilibrium using backward induction;

[0082] like Figure 3 As shown in the process, the computing power platform acts as the leader and computing power provider. As followers, platforms maximize their own utility through personalized pricing and incentivize computing power providers, who in turn provide corresponding resources based on the platform's pricing to maximize their own utility.

[0083] 1. The platform's goal is to select and publish a personalized set of pricing functions. In order to maximize its own utility The platform's utility equals the total revenue collected from task initiators. Subtract the total incentive cost paid to all edge nodes:

[0084]

[0085] Each computing power provider As a follower in the game, upon receiving the exclusive pricing function published by the platform... Then, they will choose the amount of resources they want to provide. In order to maximize its net income Its net benefit equals the reward it receives minus the private cost of providing the resources. :

[0086]

[0087]

[0088] 2. As the number of interactions increases, the learned strategies from the hybrid approach... It will become increasingly closer to the true nature of computing power providers. Therefore, through Solve Obtain the estimated computing power provider Optimal response function .

[0089] 3. The platform personalizes pricing based on anticipated follower responses, and the platform uses all predicted individual response functions. Substitute your own utility function Because the platform can provide nodes for each computing power Set up independent And each computing power provides nodes response Only depend on Therefore, the platform's global optimization problem can be decomposed into: This is a separate optimization problem. The platform provides nodes for each computing power. Solve the following problems separately to design the optimal personalized price:

[0090]

[0091]

[0092] Once the platform determines the optimal set of discriminatory pricing strategies... It will give each exclusive They are published separately to the corresponding computing power providing nodes. Each computing power provider According to exclusive and its true cost To make its actual optimal resource supply :

[0093]

[0094] Finally, the strategy combination is obtained. .

[0095] 4. The platform calculates... and Deviation between And set an acceptable deviation threshold. To establish a resource pool :

[0096]

[0097] This resource pool initially screens out computing power providers whose resource deviations are within the threshold and whose computing power providers meet the minimum resource requirements of the sub-tasks. At the same time, it allows for appropriate fluctuations in the cost of computing power resource providers and prevents them from exaggerating or concealing information in order to obtain greater profits.

[0098] Step 4: LLM uses the task dependency graph to enter the "reasoning-action-scoring" stage to allocate tasks and provide feedback for optimization;

[0099] 1. For example Figure 4 As shown, after obtaining the resource requirement vector of the task... and qualified resource supply pool Afterwards, the system enters the task execution scheduling phase. To reduce the latency caused by communication required for tasks and the reassignment of tasks in case of failure, the platform maintains a communication latency matrix for computing power providers in the qualified resource supply pool. Reliability rating of computing power providers The details are as follows:

[0100]

[0101]

[0102] in, Delay matrix Medium computing power providers Communication latency to the platform, when hour, Delay matrix Medium computing power providers To computing power providers Communication latency, for Based on historical task completion data, LLM evaluates computing power providers. The initial rating Set it to 100.

[0103] 2. [This will involve] complex dependencies. Multiple subtasks Optimally allocate Medium-power nodes present an NP-hard combinatorial optimization problem. Therefore, in this step, LLM is used as an advanced heuristic solver to generate high-quality approximate scheduling solutions.

[0104] 1) LLM obtains input information, which includes the task dependency graph. Subtask vectors ,resource Pricing Delay matrix Reliability rating of computing power providers information.

[0105] 2) Based on the information obtained, LLM determines the priority of computing power providers. :

[0106]

[0107] in These correspond to the weights of latency, reliability, and payment to computing power providers, respectively, with all weights placed on... Take values ​​within the range, and satisfy the following conditions: .

[0108] 3) LLM analysis of the resource requirements of currently executable subtasks and according to priority Assigned to China satisfies The LLM continuously monitors the task execution status, and if a task fails, the LLM, based on its understanding of the feedback information (such as insufficient video memory), reassigns the task to a new computing power provider.

[0109] 4) After the task is completed, the computing power provider is evaluated based on the LLM. Task completion rating Update its reliability score ,in Weight value, this value in Within the range.

[0110] 3. Construct resource prediction data pairs in the format of <task description, actual resource consumption>, and task inference data pairs in the format of <task dependency graph, task completion status>. Use these data to fine-tune the LLM model to optimize its prediction and decision-making capabilities. Simultaneously, incorporate the data generated during the game (…). This constitutes new "state-action" trajectory data, which is used to update the cost estimation function in the hybrid strategy and achieve closed-loop optimization of the system.

[0111] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A task offloading method based on Stackelberg game driven by LLM, characterized in that: Includes the following steps: S1: Based on LLM, analyze the resource and task dependency graph required for the task; S2: Infer the private cost function of the computing power provider through Bayesian and reinforcement learning; S3: Construct a Stackelberg game model between the computing power platform and the computing power provider; S4: Tasks are allocated by LLM to achieve an efficient incentive mechanism for computing power providers and a task scheduling mechanism.

2. The task offloading method based on Stackelberg game driven by LLM according to claim 1, characterized in that: S1 specifically includes: S1.1: Based on LLM, perform semantic understanding on user-submitted computing tasks to identify key algorithms, computing patterns, data dependencies, and resource-intensive operations; S1.2: Based on the recognition results, generate a multi-dimensional resource requirement vector including CPU cycle requirements, peak memory usage, total data I / O, GPU usage requirements, and parallelization potential score; S1.3: Generate a task dependency graph and predict the funds required for each task.

3. The task offloading method based on Stackelberg game driven by LLM according to claim 1, characterized in that: S2 specifically includes: S2.1: Use a Bayesian method to combine equipment type and location information to form an initial cost estimate for newly added computing power providers; S2.2: For computing power providers with existing interaction records, inverse reinforcement learning is used to infer cost preference parameters from historical supply behavior; S2.3: Dynamically update the cost function parameters based on the actual supply response of each interaction.

4. The task offloading method based on Stackelberg game driven by LLM according to claim 1, characterized in that: Specifically, S3 includes: S3.1: Set the computing power platform as the leader in the game and the computing power provider as the follower, and design a personalized pricing strategy; S3.2: Each computing power provider independently determines the amount of resources to be supplied based on the received pricing strategy; S3.3: By solving the optimal response function of the followers, the pricing strategy is optimized in reverse, and reliable computing power providers are selected based on the deviation between the actual supply and the predicted supply.

5. The task offloading method based on Stackelberg game driven by LLM according to claim 1, characterized in that: S4 specifically includes: S4.1: Maintain a pool of qualified computing power providers, with each provider in the pool associated with communication latency, reliability score, and cost. S4.2: LLM acquires key input information and determines scheduling priorities; S4.3: Assign executable subtasks according to priority and monitor their execution status in real time; S4.4: When a subtask fails or times out, the LLM reallocation mechanism is triggered; S4.5: After the task is completed, a feedback score is generated based on the completion quality, timeliness, and resource efficiency, and the reliability score is updated.

6. The task offloading method based on Stackelberg game driven by LLM according to claim 5, characterized in that: The key input information in S4.2 includes the task dependency graph, subtask resource requirements, and communication latency, reliability score, and cost of each provider. The scheduling priority is determined by a weighted average of three factors: communication latency, reliability score, and the financial cost that the platform needs to pay to the computing power provider.

7. The task offloading method based on Stackelberg game driven by LLM according to claim 5, characterized in that: The reliability score in S4.5 is updated based on a moving average method. The new score is obtained by merging the historical score and the current task feedback score according to a preset ratio.

8. The task offloading method based on Stackelberg game driven by LLM according to claim 1, characterized in that: It also includes closed-loop optimization steps, specifically including: Collect data on the discrepancy between task descriptions and actual resource consumption to form training samples for task resource prediction; Record the task dependency graph and the final execution result to form training samples for task scheduling decisions; The LLM was fine-tuned using the above samples to improve its task parsing and scheduling capabilities. The states and action trajectories generated in the game are used to update the cost function estimation model in S2.

Images