A mobile crowdsourcing strategy optimization method and system based on multi-objective optimization

Abstract

The present invention relates to the intersection of crowd-sensing technology and artificial intelligence technology, and in particular to a mobile crowdsourcing strategy optimization method and system based on multi-objective optimization. The method comprises utilizing a crowdsourcing platform to publish task information, including publishing task type, maximum budget, and equipment requirements; pushing information based on the publication location of the task information; calculating worker information attributes based on the published task information and uploading it to the crowdsourcing platform; matching tasks and workers based on the worker information attributes and task information; and utilizing an initial pricing model to perform real-time dynamic pricing based on the actual situation of the current task and worker. The method combines the number of tasks completed by the worker on that day, the current reputation value, and the encryption of the current reputation and current rating of the organization. The present invention performs preliminary screening of all workers who have uploaded information to obtain a list of worker candidates and generate an initial solution set. This process narrows the search scope and reduces unnecessary computational effort.

Description

Technical Field

[0001] The present invention relates to the intersection of crowd-sensing technology and artificial intelligence technology, and in particular to a mobile crowdsourcing strategy optimization method and system based on multi-objective optimization. Background Art

[0002] Mobile crowd sensing is a paradigm for achieving collective intelligence through multiple sensors and devices in the IoT ecosystem. Powered by IoT devices, edge computing, and drone sensing, crowd sensing plays a vital role in urban sensing, environmental monitoring, and traffic management. Key challenges in mobile crowd sensing research include recruiting the right workers to complete tasks and distributing benefits among all parties (i.e., pricing). Existing recruitment strategies typically prioritize efficiency, quality, and coverage, but overlook the matching between workers and tasks, which directly impacts the quality of task completion. Furthermore, existing pricing methods, often employing tiered pricing, difficulty pricing, and fixed pricing, fail to fully account for differences between the historical and current status of each party, failing to fully incentivize participation and thus impacting overall profitability and participation.

[0003] Traditional mobile crowd-sensing systems have defects such as weak privacy protection, inefficient worker selection, lack of dynamic incentive mechanisms, and unreasonable profit distribution. The platform bears the dual responsibilities of data encryption and worker recruitment, resulting in insufficient privacy protection. The worker selection mechanism cannot effectively cope with the complexity of multi-tasks and is prone to falling into local optimality. The fixed incentive mechanism lacks flexibility and cannot motivate workers to participate in tasks with high quality. In addition, the profit distribution method is unreasonable and the interests of all parties are not effectively balanced. At this stage, a mobile crowdsourcing strategy optimization method and system based on multi-objective optimization is needed. Summary of the Invention

[0004] In order to solve the problems of weak privacy protection, low worker selection efficiency, lack of dynamic incentive mechanism and unreasonable benefit distribution in traditional mobile crowd-sensing systems, the present invention provides a mobile crowdsourcing strategy optimization method and system based on multi-objective optimization, which can accurately match workers with tasks and achieve efficient worker recruitment, maximize the completion effect of tasks, and ensure reasonable pricing to maximize the interests of all parties.

[0005] In a first aspect, the present invention provides a mobile crowdsourcing strategy optimization method based on multi-objective optimization, which adopts the following technical solutions:

[0006] A mobile crowdsourcing strategy optimization method based on multi-objective optimization, comprising:

[0007] Utilize crowdsourcing platforms to publish task information, including task type, maximum budget, and equipment requirements;

[0008] Pushing information based on the location where the task information is published, including extracting the geographical location where the task information is published and pushing the task based on the preset broadcast range;

[0009] Calculate worker information attributes based on the published task information and upload them to the crowdsourcing platform. The calculated worker information includes the distance attribute, reputation attribute, minimum income, and ability value attribute of the calculation task;

[0010] Perform task-worker matching based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table;

[0011] The optimal task allocation information table is input into the KDQN dynamic pricing model, and dynamic pricing is performed through the Stackelberg game to obtain the initial pricing model;

[0012] The initial pricing model is used to conduct real-time dynamic pricing based on the current tasks and the actual situation of the workers, including the number of tasks completed by the workers on that day, the current reputation value, the current reputation of the encryption agency and the current score.

[0013] Furthermore, the worker information attributes are calculated based on the published task information and uploaded to the crowdsourcing platform, including calculating the distance attribute based on the geographic information in the task information and the worker's current geographic information, and normalizing the calculated distance attribute. The distance attribute calculation formula is:

[0014] ,

[0015] in, is the radius of the Earth, is the dimensional difference between the worker's location and the task's posting location, is the dimension of the worker’s current geographical information, is the dimension of the task release location, is the longitude difference between the worker's location and the location where the task is posted.

[0016] Furthermore, the use of the EH-STLS algorithm for task-worker matching includes preliminary screening of worker information attributes of all uploaded information based on task information and worker information attributes to obtain a list of worker candidates, and generating an initial solution set based on the worker candidate list. The EH-STLS algorithm is used to perform a comprehensive evaluation on the initial solution set, so that each task is matched with only one worker at this stage, thereby generating a preliminary global optimal solution set.

[0017] Furthermore, the use of the EH-STLS algorithm for task-worker matching also includes performing dynamic crossover and mutation iteration on the basis of the initial solution set and while satisfying requirements and constraints, and further searching for high-quality Pareto solutions through variable domain local search after a certain number of iterative rounds to generate an optimal task allocation information table.

[0018] Furthermore, the optimal task allocation information table is input into the KDQN dynamic pricing model, including parsing and preprocessing the data in the task allocation information table, and constructing a task-worker allocation matrix, then extracting the task features and detailed attributes of the workers in the task allocation information table, and normalizing the extracted numerical data.

[0019] Furthermore, the dynamic pricing operation is performed through the Stackelberg game to obtain an initial pricing model, including using the backward induction method to solve the optimal strategy of followers, solving the optimal work effort and pricing strategy of workers and encryption agencies respectively, and then solving the optimal pricing strategy of the platform, and then determining the Stackelberg equilibrium solution, constructing the initial pricing model based on the equilibrium solution, and obtaining the model function by fitting the relationship between the pricing strategy and the task requirements and budget.

[0020] Furthermore, the initial pricing model is used to perform real-time dynamic pricing based on the actual situation of the current task and the worker, including obtaining the number of tasks completed on the day and the current reputation value in real time for the worker, and integrating it with the worker ability information in the initial pricing model into a real-time state vector for the worker. For the task, the task type, maximum budget and equipment requirements in the initial pricing model are combined to form a real-time state vector for the task, and finally summarized into a complete real-time state vector for pricing calculation.

[0021] In the second aspect, a mobile crowdsourcing strategy optimization system based on multi-objective optimization includes:

[0022] The data acquisition module is configured to: utilize the crowdsourcing platform to publish task information, including task type, maximum budget, and equipment requirements;

[0023] The push module is configured to: push information according to the publishing location of the task information, including extracting the geographical location where the task information is published and pushing the task according to a preset broadcast range;

[0024] An attribute module is configured to: calculate worker information attributes based on the published task information and upload the information to the crowdsourcing platform, wherein the calculated worker information includes calculating the distance attribute, reputation attribute, minimum income and ability value attribute of the task;

[0025] The feature extraction module is configured to: perform task-worker matching based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table;

[0026] The model module is configured to: input the optimal task allocation information table into the KDQN dynamic pricing model, perform dynamic pricing operations through the Stackelberg game, and obtain an initial pricing model;

[0027] The transformation module is configured to use the initial pricing model to perform real-time dynamic pricing based on the current task and the actual situation of the worker, including combining the number of tasks completed by the worker on the day, the current reputation value, and the encryption agency's current reputation and current score.

[0028] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for the mobile crowdsourcing strategy optimization method based on multi-objective optimization.

[0029] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the mobile crowdsourcing strategy optimization method based on multi-objective optimization.

[0030] In summary, the present invention has the following beneficial technical effects:

[0031] 1. This invention uses an enhanced heuristic search algorithm based on taboo and local search (EH-STLS) to match tasks with workers. By comprehensively considering task type, budget, and various constraints, it can accurately screen the most suitable workers for each task. Based on factors such as task location, equipment requirements, worker reputation and ability, tasks are assigned to workers in the area who have high-precision monitoring equipment and a good work record, greatly improving the accuracy of task allocation.

[0032] 2. In the EH-STLS algorithm, the present invention performs a preliminary screening of all workers who upload information to obtain a list of worker candidates and generate an initial solution set. This process narrows the search scope and reduces unnecessary computation. At the same time, the algorithm is used to perform a comprehensive evaluation on the initial solution set. Each task is matched with only one worker to generate a preliminary global optimal solution set, further focusing the search direction.

[0033] 3. In the KDQN model, the present invention parses and preprocesses the data in the task allocation information table, extracts task features and detailed worker attributes, and normalizes the numerical data, providing an accurate and standardized data basis for the pricing model.

[0034] 4. The present invention uses the backward induction method to solve the optimal strategy of followers in the Stackelberg game, determine the Stackelberg equilibrium solution and construct an initial pricing model. By fitting the relationship between the pricing strategy and the task requirements and budget, the model function is obtained, so that the pricing model has a scientific theoretical basis and strong adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is an overall system framework diagram of a mobile crowdsourcing strategy optimization method based on multi-objective optimization according to Example 1 of the present invention;

[0036] Figure 2 This is a comparison diagram of the KAN layer and the MLP layer of the reinforcement learning dynamic pricing model based on Stackelberg game in Example 1 of the present invention;

[0037] Figure 3 This is a graph showing the experimental results of worker task quality as the number of tasks changes in Example 1 of the present invention;

[0038] Figure 4 This is a graph showing the experimental results of task quality as the number of workers changes with the task in Example 1 of the present invention;

[0039] Figure 5 This is a pricing and benefit diagram for different numbers of tasks in Example 1 of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below with reference to the accompanying drawings.

[0041] Example 1

[0042] Reference Figure 1 , a mobile crowdsourcing strategy optimization method based on multi-objective optimization in this embodiment includes:

[0043] Utilize crowdsourcing platforms to publish task information, including task type, maximum budget, and equipment requirements;

[0044] Pushing information based on the location where the task information is published, including extracting the geographical location where the task information is published and pushing the task based on the preset broadcast range;

[0045] Calculate worker information attributes based on the published task information and upload them to the crowdsourcing platform. The calculated worker information includes the distance attribute, reputation attribute, minimum income, and ability value attribute of the calculation task;

[0046] Perform task-worker matching based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table;

[0047] The optimal task allocation information table is input into the KDQN dynamic pricing model, and dynamic pricing is performed through the Stackelberg game to obtain the initial pricing model;

[0048] The initial pricing model is used to conduct real-time dynamic pricing based on the current tasks and the actual situation of the workers, including the number of tasks completed by the workers on that day, the current reputation value, the current reputation of the encryption agency and the current score.

[0049] Specifically,

[0050] In the mobile crowd-sensing system, there are four main roles: task publishers, workers, platforms, and encryption agencies. For example, consider a task that requires data collection issued by a task publisher. The platform then broadcasts the task near the task location. After receiving the information uploaded by the workers, it selects the best worker through multi-objective optimization. The platform then puts the four parties together for dynamic pricing, which can greatly ensure the quality of task completion and the distribution of benefits to all parties involved.

[0051] like Figure 1 As shown, the present invention designs a mobile crowdsourcing efficiency strategy optimization framework based on multi-objective optimization and reinforcement learning game methods. In the worker recruitment and dynamic pricing framework, worker recruitment is achieved through a multi-objective optimization method, and dynamic pricing is determined through a reinforcement learning game method. Task benefit and equipment similarity in worker recruitment are calculated based on the task requirements published by the task publisher and the worker attributes uploaded by the worker. The calculation of task benefit and equipment similarity takes into account the worker's multiple ability indicators and the weight requirements of the task publisher. Subsequently, the platform considers each worker's task benefit and equipment similarity for different tasks and calculates an initial list of workers for each task. This initial list of workers is then used as input into the multi-objective optimization algorithm. In this case, the worker recruitment problem is modeled as a multi-objective optimization problem, aiming to maximize task benefit and meet task requirements while ensuring task completion quality and task requirements. Dynamic pricing is achieved by training an initial model based on historical pricing results from workers, the platform, and other factors. Based on this, dynamic pricing is determined through a pricing model that considers the worker's daily task completion number, current reputation value, and the current reputation and rating of the encryption organization. At this point, the dynamic pricing problem is modeled as a Stackelberg game problem, where the task publisher acts as the leader, and the platform, crypto agencies, and workers act as followers. The goal is to maximize the distribution of benefits to all parties while ensuring that the minimum benefits are met. The specific steps are as follows:

[0052] The first embodiment of the present invention provides an enhanced heuristic search method based on taboo and local search, which includes a task information broadcasting stage, a worker information uploading stage, and a task-worker matching stage.

[0053] Mission information broadcast phase:

[0054] Step 1: The task publisher sets the task type, maximum budget, and equipment requirements through the crowdsourcing platform and then publishes the task information.

[0055] Step 2: After receiving the task information, the crowdsourcing platform broadcasts the task information near the task point using the geographic location in the task information.

[0056] Step 3: After broadcasting the task information, wait for the workers to upload their own information in the information uploading phase, and then proceed to the next phase.

[0057] Worker information upload stage:

[0058] Step 4: Workers will judge the broadcasted tasks and only start calculating their own relevant information after determining the tasks they are interested in.

[0059] In step 4, the specific steps for workers to judge the broadcasted tasks are as follows:

[0060] Step 4.1: The worker compares the task type with his or her preference list, as follows:

[0061] ,

[0062] in, is the task type, A list of worker preferences.

[0063] Step 4.2: Determine the result of the previous step. If the result is 1, proceed to the next calculation operation; otherwise, ignore the task information.

[0064] Step 5: The worker calculates the distance attribute for the task based on the geographic information in the task information and the geographic information of the worker's current location. The distance attribute calculation formula is:

[0065] ,

[0066] in, is the radius of the Earth, is the dimensional difference between the worker's location and the task's posting location, is the dimension of the worker’s current geographical information, is the dimension of the task release location, is the longitude difference between the worker's location and the location where the task is posted.

[0067] Step 6: After calculating the distance attribute for the task, normalize it to facilitate subsequent task allocation.

[0068] Step 7: The worker will then calculate his or her reputation attribute based on the completion of the last N tasks. The worker will calculate his or her reputation attribute based on the completion of the last N tasks.

[0069] In this embodiment, the reputation attribute The calculation is as follows:

[0070] ,

[0071] in, Is a reputation base value, is the average reputation value of all workers, It is the sum of the reputation scores of the worker in the last N tasks.

[0072] Step 8: Workers also need to calculate their current minimum income as a reference to prevent the situation where income is lower than cost and protect their rights and interests. Workers calculate their current minimum income as shown in the following formula:

[0073] ,

[0074] in is the regional impact factor, which is determined by the local level and has a value between (0,1). is the profit obtained each time, and S is the number of historical records.

[0075] Step 9: Finally, the worker calculates his or her ability value attributes for the task of interest based on the task requirements and his or her own equipment. The worker calculates his or her ability value attributes for the task of interest based on the task requirements and his or her own equipment. The calculation is as follows:

[0076] ,

[0077] in, is the reputation information of the worker, is the maximum working distance of the workers, 、 and Represent the weight coefficients of different items respectively, is the Jaccard similarity, which is used to evaluate the similarity between the worker's equipment situation and the equipment requirements for task completion, where Expressed as:

[0078] ,

[0079] in, Represents the set of equipment requirements needed to complete the task, which includes the equipment required to complete a specific task The required characteristics of various equipment, performance requirements and other factors, Indicates workers The equipment status collection covers the various characteristics of the equipment actually equipped by the worker.

[0080] Step 10: After the workers have calculated this information locally, they upload it to the crowdsourcing platform for the next task assignment.

[0081] Task-worker matching phase:

[0082] Step 11: After receiving the attribute information from the workers, the crowdsourcing platform generates an initial set of workers for each task.

[0083] Step 12: The platform puts this set of workers into the EH-STLS algorithm, which considers the type of task, budget, and other relevant constraints to select the workers most suitable for performing each task.

[0084] The worker recruitment problem is modeled as a multi-objective optimization problem. The enhanced heuristic search algorithm based on taboo and local search is used to process the worker information and task information to obtain the task Best executive worker The specific steps are:

[0085] Step 12.1: Categorize all information and organize it by task. Each task corresponds to a separate list of worker candidates, and the system generates an initial solution set based on these candidate lists.

[0086] Step 12.2: In the initial solution set, each task has M candidate workers. The algorithm will generate a preliminary global optimal solution set for each task based on this, where each task is matched with only one worker. This preliminary solution set is then added to the taboo list, which is used to record the solutions that have been searched to prevent the algorithm from repeatedly selecting the same solution in a short period of time.

[0087] Step 12.3: During the evolutionary process, crossover and mutation operations are combined to increase the solution exploration capability through diverse operation mechanisms. This embodiment designs a probability strategy based on dynamic adjustment of the number of iterations, so that the crossover and mutation probabilities can change dynamically as the iterations proceed. The specific formula is as follows:

[0088] ,

[0089] ,

[0090] Among them, g is the current iteration number, N is the total number of iterations, is the crossover probability, is the mutation probability.

[0091] Step 12.4: After a certain number of iterations, a local search strategy with a variable neighborhood is introduced. Different neighborhood structures are used to deeply optimize the solution to improve the convergence quality of the optimal solution and ultimately generate the optimal task allocation list. The specific formula is as follows:

[0092] ,

[0093] ,

[0094]

[0095] in, is the maximum neighborhood value, is the maximum mutation probability, is the minimum mutation probability, Pm(g) is the neighborhood coefficient of generation g, N(g) is the neighborhood size of generation g, T is the taboo table, It is the desire criterion. When the objective function value of a taboo solution is better than the currently known optimal solution, it is allowed to be selected as the new solution even if it is in the taboo table.

[0096] Step 13: Based on the screening results of the EH-STLS algorithm, the platform generates an optimal task allocation information table, which lists the correspondence between each task and its best worker.

[0097] Step 14: The platform inputs the optimal task allocation information table into the pricing model and proceeds to the next step of dynamic pricing operation.

[0098] The following example details the worker recruitment problem. Suppose a task publisher submits 10 tasks. The platform then broadcasts the locations of these 10 tasks. To account for worker distance efficiency, the platform sets the broadcast range to 50 kilometers from the task point. After receiving the task information, workers at each location calculate and upload relevant information for tasks that meet the requirements. After receiving the information uploaded by the workers, the platform generates a list of 10 candidate workers and feeds this list into a multi-objective optimization algorithm. After multiple iterations, while ensuring task completion quality and meeting task requirements while also satisfying constraints, an optimal worker matching list is obtained, containing 10 workers and 10 tasks, with a one-to-one correspondence.

[0099] To verify the effectiveness of the method in this embodiment, an enhanced heuristic search algorithm based on taboo and local search was experimented with a real dataset. The real dataset was derived from existing public data and obtained the consent of the information owner.

[0100] (1) Task quality,

[0101] Among them, EH-STLS represents the multi-objective optimization algorithm proposed in this paper, and the remaining algorithms are comparative algorithms, namely: SPEA2: It evaluates individual fitness by combining dominance relations and density estimation, and uses external archiving to save the optimal solution, thereby ensuring the quality and diversity of solutions. NSGA-II: It retains high-quality individuals by using an elitist strategy, generates new solutions using crossover and mutation operations, and ultimately finds a set of balanced Pareto front solutions. MOEA-D: It solves the multi-objective optimization problem by decomposing it into several single-objective subproblems. Each subproblem is represented by a different weighted combination and leverages neighborhood information to collaborate with other subproblems to improve search efficiency and solution diversity. MOGNDO: It preserves non-dominated solutions and guides the optimization process by introducing external archiving and a leader selection mechanism, ensuring solution diversity and convergence.

[0102] like Figure 3 As shown in the figure, with a fixed number of workers, the task quality of each algorithm decreases as the number of tasks increases. This is because as the number of tasks increases, the system prioritizes ensuring that each task is assigned to a worker, which partially compromises task quality. The EH-STLS algorithm outperforms the other compared algorithms in task quality when the number of tasks is 50, 100, and 150. In particular, compared to the MOEA-D algorithm, task quality improves by 6.6%, 11.1%, and 11.9%, respectively. This demonstrates that the EH-STLS algorithm is more effective in maintaining task execution quality as the task size increases.

[0103] like Figure 4 As shown in the figure, when the number of tasks is fixed, the task quality of each algorithm shows an upward trend as the number of workers increases. This is because the increase in the number of workers provides the system with more options when allocating tasks, thereby improving task quality. The EH-STLS algorithm outperforms the other compared algorithms in task quality when the number of tasks is 50, 100, and 150, with improvements of 21.7%, 13.3%, and 12.0%, respectively, compared to the MOEA-D algorithm. This demonstrates that EH-STLS is more effective in improving task allocation efficiency when the number of workers is small. Although the improvement decreases as the number of workers increases, EH-STLS still maintains a significant advantage.

[0104] When the number of workers is fixed, the task efficiency of each algorithm decreases as the number of tasks increases. This is because as the number of tasks increases, the system prioritizes ensuring that every task has a worker assigned to it, thus compromising task efficiency. However, the EH-STLS algorithm achieves higher task efficiency than the comparison algorithm for 50, 100, and 150 tasks, with improvements of 2.8%, 5.4%, and 9.7%, respectively, compared to the MOEA-D algorithm. This demonstrates that EH-STLS's advantage in task efficiency becomes increasingly significant as the task size increases.

[0105] When the number of tasks is fixed, the task efficiency of each algorithm increases with the number of workers. This is because an increase in the number of workers provides the system with more options, thereby improving task efficiency. The EH-STLS algorithm achieves higher task efficiency than the other compared algorithms when the number of workers is 200, 300, and 400, respectively, improving by 4.3%, 5.6%, and 3.6% compared to the MOEA-D algorithm. This demonstrates that when there are a sufficient number of workers, EH-STLS has a stronger task allocation capability and can significantly improve task efficiency.

[0106] By comparing the task quality and task efficiency of different algorithms under different numbers of workers and tasks, the EH-STLS algorithm outperforms all other algorithms. This is because the EH-STLS algorithm uses a higher mutation probability in the early stages of the search to increase the diversity of solutions and enhance the algorithm's ability to escape local optima. As the number of iterations increases, the mutation probability is gradually reduced to enhance the ability to deeply search the current solution and ensure convergence to a better global optimal solution. This dynamic adjustment mechanism achieves a balance between exploration and exploitation between different stages, effectively improving the overall quality and efficiency of task allocation. Secondly, to further improve the quality of the solution, a variable neighborhood local search strategy is introduced after a certain number of iterations. Different neighborhood structures are used to deeply optimize the solution to improve the convergence quality of the optimal solution.

[0107] Extensive experiments are conducted on three real datasets to verify the performance of the proposed enhanced heuristic search algorithm based on taboo and local search. The results show that the enhanced heuristic search algorithm based on taboo and local search outperforms other baselines.

[0108] In this paper, the Kolmogorov-Arnold network (KAN) replaces the MLP layer in the DQN. Unlike the MLP, the KAN does not use fixed activation functions (such as ReLU and Sigmoid). Instead, it uses learnable activation functions on the "weights," parameterized by splines, enabling it to learn more complex function mappings during training. While the MLP learns features through an alternating combination of linear weight matrices (fully connected layers) and nonlinear activation functions, the KAN does not have such a linear weight matrix. Instead, each weight parameter is replaced by a single-variable learnable function (typically a spline function). This enables the KAN to outperform the MLP in fitting low-dimensional data and solving partial differential equations (PDEs). In a KAN, nodes simply sum the incoming signals without applying any nonlinearity; all nonlinearity is applied to the edges.

[0109] like Figure 2 As shown in Figure 1, KAN is introduced to replace MLP. KAN uses a learnable activation function, which makes it more flexible in approximating complex functional relationships, thereby better fitting the Q value. At this time, the Q function is changed to:

[0110] ,

[0111] in, is a learnable activation function, are the parameters of KAN.

[0112] In addition, the structure of KAN enables it to adjust the form of the activation function according to the complexity of the input data during training, thereby improving the convergence speed of the model. The convergence speed is improved to:

[0113] ,

[0114] in, It represents the convergence rate index related to the number of model parameters N. Expressed as the convergence speed of the KAN network, In KAN, the activation function itself is also part of the network and is continuously learned and adjusted through the optimization process. This allows the activation function of each layer to be adaptively adjusted according to the characteristics of the data, thereby achieving more flexible nonlinear mapping. The output function of KAN is expressed as:

[0115] ,

[0116] in, It is based on the input state s and action a and model parameters To calculate an estimate, is a learnable function of a single variable, is a global nonlinear combination function.

[0117] This means the model can fine-tune each input feature Compared with MLP, KAN provides richer derivative forms and can therefore more accurately explain how input features affect the output.

[0118] To verify the effectiveness of the method in this embodiment, an experiment was conducted on a dynamic pricing model based on reinforcement learning using a Stackelberg game using a real dataset. The dataset was obtained from existing public data and with the consent of the information owner.

[0119] (1) Pricing benefits,

[0120] Among them, KDQN represents the reinforcement learning dynamic pricing model based on Stackelberg game proposed in this invention, and the other algorithms are comparative algorithms, namely: DQN: an algorithm that combines deep learning and reinforcement learning, using neural networks to approximate the Q-value function, and is used to solve decision-making problems in high-dimensional state spaces. RBFNN: a feedforward neural network, commonly used for tasks such as function approximation, classification, and time series prediction. It uses radial basis function as the activation function and constructs a nonlinear mapping by calculating the distance between the input and the center point. TAPRIM: a task allocation and pricing incentive mechanism for edge-assisted crowd intelligence perception scenarios, which coordinates and optimizes the benefits of each participant in the system through a three-stage Stackelberg game model, while ensuring the rationality of task allocation and data quality.

[0121] like Figure 5 As shown in the figure, as the number of tasks increases, the pricing gains of all models (KDQN, DQN, RBFNN, and TAPRIM) show an upward trend. However, the KDQN model achieves significantly higher gains than the other models, especially when the number of tasks is large, where it performs best. In contrast, the TAPRIM model performs slightly worse than KDQN, but still outperforms DQN and RBFNN, demonstrating strong competitiveness.

[0122] Specifically, the KDQN model better approximates the profit maximization strategy through deep reinforcement learning, so its pricing profit is always ahead under different numbers of tasks.

[0123] We conduct extensive experiments on three real-world datasets to validate the performance of the proposed Stackelberg game-based reinforcement learning dynamic pricing model. The results show that the Stackelberg game-based reinforcement learning dynamic pricing model outperforms other baselines.

[0124] Example 2

[0125] This embodiment provides a mobile crowdsourcing strategy optimization system based on multi-objective optimization, including:

[0126] The data acquisition module is configured to: utilize the crowdsourcing platform to publish task information, including task type, maximum budget, and equipment requirements;

[0127] The push module is configured to: push information according to the publishing location of the task information, including extracting the geographical location where the task information is published and pushing the task according to a preset broadcast range;

[0128] An attribute module is configured to: calculate worker information attributes based on the published task information and upload the information to the crowdsourcing platform, wherein the calculated worker information includes calculating the distance attribute, reputation attribute, minimum income and ability value attribute of the task;

[0129] The feature extraction module is configured to: perform task-worker matching based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table;

[0130] The model module is configured to: input the optimal task allocation information table into the KDQN dynamic pricing model, perform dynamic pricing operations through the Stackelberg game, and obtain an initial pricing model;

[0131] The transformation module is configured to use the initial pricing model to perform real-time dynamic pricing based on the current task and the actual situation of the worker, including combining the number of tasks completed by the worker on the day, the current reputation value, and the encryption agency's current reputation and current score.

[0132] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a mobile crowdsourcing strategy optimization method based on multi-objective optimization.

[0133] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement a mobile crowdsourcing strategy optimization method based on multi-objective optimization.

[0134] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A mobile crowdsourcing strategy optimization method based on multi-objective optimization, characterized in that: include: Utilize crowdsourcing platforms to publish task information, including task type, maximum budget, and equipment requirements; Pushing information based on the location where the task information is published, including extracting the geographical location where the task information is published and pushing the task based on the preset broadcast range; Calculate worker information attributes based on the published task information and upload them to the crowdsourcing platform. Calculating worker information includes calculating the distance attribute, reputation attribute, minimum income, and capability value attribute of the task. Calculating worker information attributes based on the published task information and uploading them to the crowdsourcing platform includes calculating the distance attribute based on the geographic information in the task information and the worker's current geographic information, and normalizing the calculated distance attribute. The distance attribute calculation formula is: ， in, is the radius of the Earth, is the dimensional difference between the worker's location and the task's posting location, is the dimension of the worker’s current geographical information, is the dimension of the task release location, is the longitude difference between the worker's location and the location where the task is posted; It also includes workers calculating their own ability values ​​for the tasks they are interested in based on the task requirements and their own equipment. The calculation is as follows: ， in, is the reputation information of the worker, is the maximum working distance of the workers, 、 and Represent the weight coefficients of different items respectively, is the Jaccard similarity, which is used to evaluate the similarity between the worker's equipment situation and the equipment requirements for task completion, where Expressed as: ， in, Represents the set of equipment requirements needed to complete the task, which includes the equipment required to complete a specific task The characteristics and performance requirements of various equipment required, Indicates workers The equipment status collection covers the various characteristics of the equipment actually equipped by the worker; Calculate your own reputation attribute based on the completion of the last N tasks. Reputation attribute The calculation is as follows: ， in, Is a reputation base value, is the average reputation value of all workers, is the sum of the reputation scores of the worker in the last N tasks; Task-worker matching is performed based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table. The task-worker matching using the EH-STLS algorithm includes preliminary screening of worker information attributes of all uploaded information based on task information and worker information attributes to obtain a list of worker candidates, and generating an initial solution set based on the list of worker candidates. The EH-STLS algorithm is used to perform a comprehensive evaluation in the initial solution set, so that each task is matched with only one worker at this stage, thereby generating a preliminary global optimal solution set. It also includes performing dynamic crossover and mutation iterations based on the initial solution set and under the conditions of meeting requirements and constraints, increasing the solution exploration capability through a diverse operation mechanism, and dynamically adjusting the probability strategy based on the number of iterations so that the crossover and mutation probabilities can change dynamically as the iterations proceed. The specific formula is shown below: ， ， Among them, g is the current iteration number, N is the total number of iterations, is the crossover probability, is the mutation probability. After a certain number of iterations, the optimal Pareto solution is further sought through local search in the variable domain to generate the optimal task allocation information table. Inputting the optimal task allocation information table into the KDQN dynamic pricing model, performing dynamic pricing operations through the Stackelberg game, and obtaining an initial pricing model, including using the backward induction method to solve the optimal strategy of followers, solving the optimal work effort and pricing strategy of workers and encryption agencies respectively, and then solving the optimal pricing strategy of the platform, and then determining the Stackelberg equilibrium solution, constructing the initial pricing model based on the equilibrium solution, and obtaining the model function by fitting the relationship between the pricing strategy and the task requirements and budget; inputting the optimal task allocation information table into the KDQN dynamic pricing model, including parsing and preprocessing the data in the task allocation information table, and constructing a task-worker allocation matrix, then extracting task features and detailed attributes of workers from the task allocation information table, and normalizing the extracted numerical data; The KDQN dynamic pricing model introduces KAN to replace the MLP layer in DQN. KAN uses a learnable activation function, making it more flexible in approximating complex functional relationships, thereby better fitting the Q value. In this case, the Q function is changed to: ， in, is a learnable activation function, are the parameters of KAN; In addition, the structure of KAN enables it to adjust the form of the activation function according to the complexity of the input data during training, thereby improving the convergence speed of the model. The convergence speed is improved to: ， Among them, α represents the convergence rate index related to the number of model parameters N, Expressed as the convergence speed of the KAN network, Represents complexity; The initial pricing model is used to perform real-time dynamic pricing based on the current task and the actual situation of the worker. This includes combining the number of tasks completed by the worker on that day, the current reputation value, and the encryption agency's current reputation and current score. For workers, the number of tasks completed on that day and the current reputation value are obtained in real time and integrated with the worker's ability information in the initial pricing model to form a real-time state vector for the worker. For tasks, the task type, maximum budget, and equipment requirements in the initial pricing model are combined to form a real-time state vector for the task, and finally summarized into a complete real-time state vector for pricing calculation.

2. A mobile crowdsourcing strategy optimization system based on multi-objective optimization, executing the method according to claim 1, characterized in that: include: The data acquisition module is configured to: utilize the crowdsourcing platform to publish task information, including task type, maximum budget, and equipment requirements; The push module is configured to: push information according to the publishing location of the task information, including extracting the geographical location where the task information is published and pushing the task according to a preset broadcast range; An attribute module is configured to: calculate worker information attributes based on the published task information and upload the information to the crowdsourcing platform, wherein the calculated worker information includes calculating the distance attribute, reputation attribute, minimum income and ability value attribute of the task; The feature extraction module is configured to: perform task-worker matching based on worker information attributes and task information, including using the EH-STLS algorithm to perform task-worker matching and generate an optimal task allocation information table; The model module is configured to: input the optimal task allocation information table into the KDQN dynamic pricing model, perform dynamic pricing operations through the Stackelberg game, and obtain an initial pricing model; The transformation module is configured to use the initial pricing model to perform real-time dynamic pricing based on the current task and the actual situation of the worker, including combining the number of tasks completed by the worker on the day, the current reputation value, and the encryption agency's current reputation and current score.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the mobile crowdsourcing strategy optimization method based on multi-objective optimization as claimed in claim 1.

4. A terminal device comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement various instructions; and the computer-readable storage medium is configured to store a plurality of instructions, wherein: The instructions are suitable for being loaded by a processor and executing the mobile crowdsourcing strategy optimization method based on multi-objective optimization as described in claim 1.

Images