A stackelberg dynamic heterogeneous task allocation method and system for crowd sensing

Abstract

The present application relates to the technical field of task allocation, in particular to a Stackelberg dynamic heterogeneous task allocation method and system for crowd-sensing. The method comprises obtaining task data of crowd-sourcing tasks, requester historical data and worker historical data; calculating task allocation strategy indicators according to the task data, the requester historical data and the task historical data; balancing the task allocation strategy indicators based on an improved fireworks algorithm; predicting user task response allocation according to the task allocation strategy indicators; and obtaining platform operation benefits based on the user task response allocation prediction. Under the Stackelberg game model, the super multi-objective optimization of worker profit, requester cost, platform profit and QoS is realized, the worker profit is guaranteed on the premise of taking the platform benefit as the first starting point, and the requester participation rate in the platform is also ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of task allocation, in particular to a Stackelberg dynamic heterogeneous task allocation method and system for crowd sensing. BACKGROUND

[0002] With the wide popularity of mobile devices and wireless communication technologies, Mobile Crowd Sensing (MCS) as a data collection and processing mode based on public smart devices has rapidly developed in many fields such as traffic management, environmental monitoring, disaster response, and health care in recent years. MCS achieves efficient use of resources and intelligent deployment of services by encouraging individual users to participate in sensing tasks, which has significant economic and social value. However, although MCS has made some progress in theory and application, current technologies still have many shortcomings in dynamic heterogeneous task allocation, which restricts its further development and popularization.

[0003] Firstly, existing task allocation strategies generally fail to fully consider the heterogeneity of tasks and the dynamic changes of the environment. In actual MCS applications, task types differ significantly, with complex and diverse attributes, and different tasks have different requirements for time, precision, cost, and other dimensions. However, traditional methods often assume uniform task attributes and ignore the differences in scheduling of diversified tasks, resulting in allocation results that are difficult to meet actual needs. In addition, the MCS environment is highly dynamic, with continuous changes in task and worker states, making it difficult for existing static or semi-dynamic strategies to adapt to changes in real time, leading to problems such as resource waste and reduced efficiency.

[0004] Secondly, existing optimization models are difficult to coordinate the multi-party interest conflicts between the platform, workers, and requesters. In the MCS system, the platform focuses on profit maximization, the requester hopes to complete the task at the lowest cost, and the worker expects to obtain the highest return. Current optimization algorithms, such as genetic algorithm, particle swarm optimization, etc., often focus on the goals of one party, and fail to build an effective game model to balance the multi-party game relationship, resulting in a decrease in task participation rate.

[0005] In addition, existing multi-objective optimization algorithms generally have slow convergence speed, fall into local optimum, and poor adaptability. Especially in dealing with high-dimensional, multi-objective, and dynamic changes of MCS, traditional algorithms show obvious limitations. Their structure is complex and parameter sensitive, making it difficult to obtain a global optimal solution within a limited time, and unable to meet the efficiency requirements of real-time task scheduling. SUMMARY

[0006] To solve the above-mentioned problems, the present application provides a Stackelberg dynamic heterogeneous task allocation method and system for crowd sensing.

[0007] The first aspect of the present application provides a Stackelberg dynamic heterogeneous task allocation method for crowd intelligence perception, which adopts the following technical scheme:

[0008] A Stackelberg dynamic heterogeneous task allocation method for crowd intelligence perception, comprising:

[0009] Obtaining task data, requester historical data and worker historical data of crowdsourcing tasks;

[0010] Calculating task allocation strategy indicators according to the task data, requester historical data and task historical data;

[0011] Balancing the task allocation strategy indicators based on the improved firework algorithm;

[0012] Performing user task response allocation prediction according to the task allocation strategy indicators;

[0013] Obtaining a task response allocation result.

[0014] Further, the calculating of the task allocation strategy indicators according to the task data, requester historical data and worker historical data comprises calculating the matching degree of the task and the worker according to the task data and the worker historical data, wherein the project demand coverage rate is calculated according to the worker skill set and the project demand set, the matching degree of the worker and the task geographic location is calculated by using the Manhattan distance, the Manhattan distance between the worker and the task is represented as d(Lw, Lt)=∣xw-xt∣+∣yw-yt∣, the worker historical data is evaluated, the worker historical performance is taken as an important basis for the platform to allocate tasks, and the worker historical performance is represented as: The worker reputation value is calculated, wherein Q represents the task completion quality, g represents the worker praise rate, i.e., the number of positive evaluations, and the worker reputation value is represented as:

[0015] .

[0016] Further, the calculating of the task allocation strategy indicators according to the task data, requester historical data and worker historical data further comprises calculating the worker task reward coefficient according to the task data and the worker historical data, wherein the worker reputation value is calculated according to the task completion quality and the worker historical praise data; the worker behavior response rate is calculated according to the task data and the worker historical performance data; the worker reward coefficient is determined based on the worker reputation value and the worker behavior response rate, and the reward coefficient is represented as:

[0017] ,

[0018] wherein Δ is a constant.​

[0019] Further, the task allocation strategy index calculated according to the task data, the requester historical data and the worker historical data further comprises a requester reputation value calculated according to the task data and the requester historical data, wherein the requester experience value is calculated according to the difficulty division and different coefficients of the tasks published and completed by the requester according to the task data; the requester reputation value is calculated according to the historical data of the requester on the platform, wherein PO represents the number of the requester who does not pay the task reward in time after the worker normally completes the task, PO1 represents the number of the requester who exceeds the time limit for payment, PO0 represents the number of the requester who never pays, and the requester reputation value is represented as:

[0020] 。

[0021] Further, the task allocation strategy index balanced based on the improved firework algorithm comprises the following steps: taking the maximization of the worker skill and task matching degree as the target, taking the value of the worker skill and task matching degree as the fitness index, calculating the average fitness after screening by using the greedy algorithm, and obtaining the initial firework population and the optimal weight through iteration; the formula for calculating the amplitude and the explosion spark number of each firework is:

[0022] A = A_min + (A_max - A_min) * (M - fmin) / (fmax - fmin)

[0023] N_sparks = N_sparks_min + (N_sparks_max - N_sparks_min) * (M - fmin) / (fmax - fmin)

[0024] The explosion generated spark is represented as Fij, wherein i represents the i-th firework, j represents the j-th spark generated by the i-th firework explosion, j∈[1,N_sparks]; the spark mutation is set to be subject to Gaussian distribution, the data in the spark candidate pool is updated to the mutated spark, and the sparks in the candidate pool are selected and iterated, and the final Fire obtained after the iteration termination is a set of optimal solutions of the worker skill and task matching degree.

[0025] Further, the user task response allocation prediction according to the task allocation strategy index comprises the following steps: calculating the platform reputation value based on the worker historical data and the requester historical data, wherein the platform reputation value is calculated based on the scores given to the platform by the workers and the scores given to the platform by the requesters; calculating the requester risk perception coefficient based on the platform reputation value and the worker reputation value, and represented as:

[0026] ,

[0027] wherein Rrisk is the risk perception coefficient of the requester, k1, k2 and k3 are adjustment parameters; and further calculating the worker risk perception coefficient through the platform reputation value and the requester reputation value, expressed as:

[0028] ,

[0029] Further, the user task response allocation prediction according to the task allocation strategy index further comprises calculating a requester behavior response function according to the task allocation strategy index, expressed as:

[0030] ,

[0031] wherein represents the task urgency, represents the requester budget.

[0032] Further, the user task response allocation prediction according to the task allocation strategy index further comprises calculating a worker behavior response function according to the task allocation strategy index, expressed as:

[0033] ,

[0034] wherein represents the worker skill and task adaptation ratio obtained by strategy one, represents the worker reward coefficient obtained by strategy two, similar to the requester behavior response function, represents the number of cross-grids of the worker performing the task.

[0035] Further, the user task response allocation prediction according to the task allocation strategy index further comprises predicting the optimal solution of the task allocation strategy based on the worker behavior response function and the requester behavior response function, and verifying the effect of the task allocation strategy by using the requester risk perception coefficient and the worker risk perception coefficient.

[0036] In a second aspect, a Stackelberg dynamic heterogeneous task allocation system for crowd sensing comprises:

[0037] A data acquisition module configured to acquire task data of a crowdsourcing task, requester historical data and worker historical data;

[0038] A strategy index module configured to calculate a task allocation strategy index according to the task data, the requester historical data and the task historical data;

[0039] A balancing module configured to balance the task allocation strategy index based on an improved firework algorithm;

[0040] The prediction module is configured to perform user task response allocation prediction according to the task allocation strategy index.

[0041] The allocation module is configured to obtain a task response allocation result.

[0042] In a third aspect, the present application provides a computer readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and to execute the method.

[0043] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer readable storage medium, the processor being used to implement instructions, and the computer readable storage medium being used to store a plurality of instructions, the instructions being adapted to be loaded by the processor and to execute the method.

[0044] To sum up, the present application has the following beneficial technical effects:

[0045] 1. Realize multi-party benefit maximization: under the Stackelberg game model, the worker profit, the requester cost, the platform profit and the QoS are super multi-objective optimized, the worker profit is guaranteed under the premise of taking the platform benefit as the first starting point, and the requester participation rate in the use task of the platform is also ensured.

[0046] 2. Improve the rationality of platform strategy: the demand of the worker and the requester is predicted in advance, and the improved firework algorithm is used to optimize the conflicting objectives in the platform strategy, so that the platform strategy optimal solution is obtained under the premise of meeting the demand of the three parties. The improved firework algorithm of the present application has excellent stability and optimization performance in the dynamic environment, especially in the scene with more optimization objectives.

[0047] 3. Effectively ensure user usage rate: three rounds of simulation experiments are carried out by using three data sets, and the experimental results show that the average probability of the worker selecting to accept the task is 89.14%, the average probability of the requester selecting to publish the task is 73.76%, and the probability of the worker selecting to cross the grid is 39.5%. It can be seen that the dynamic reward mechanism of the present application has a positive effect on the behavior response, and the task allocation mechanism can effectively allocate according to the user preference, so as to ensure the user participation rate and improve the platform usage rate.

[0048] 4. Effectively predict user behavior characteristics: the present application integrates three user individual demand theories into the game model, deeply analyzes the influence of user historical behavior on decision-making, and minimizes the negative influence. From the average use probability of the user, it can be seen that integrating the behavior preference constraint model into the game theory framework can effectively improve the behavior response rate and the task participation rate of the user. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure Figure 1 is the MCS system operation architecture diagram of the present application.

[0050] Figure 2 is the Stackelberg game model flowchart of the present application.

[0051] Figure 3 is the grid partitioning diagram of the present application for workers to perform tasks across grids.

[0052] Figure 4 is the comparison diagram of the optimized platform strategy of the improved fireworks algorithm of the present application and the optimization effect of other algorithms. Among them, Figure 4 (a) in the figure represents the comparison diagram of the experimental effect of the worker reputation value, Figure 4 (b) in the figure represents the comparison diagram of the experimental effect of the worker skill and task matching degree.

[0053] Figure 5 is the columnar analysis diagram of the influence of the dynamic reward mechanism and the behavior preference constraint mechanism on the task participation rate. Among them, Figure 5 (a) in the figure represents the positive selection probability of the requester under different conditions, Figure 5 (b) in the figure represents the positive selection probability of the worker under different conditions.

[0054] Figure 6 is the comparison and analysis of the task market pricing and the market basic price obtained by the present application.

[0055] Figure 7 is the algorithm optimization effect comparison analysis of the three-round experimental simulation dynamic environment. Among them, Figure 7 (a) in the figure represents the optimization effect comparison of the platform evaluation of the worker, Figure 7 (b) in the figure represents the optimization effect comparison of the task pricing, Figure 7 (c) in the figure represents the optimization effect comparison of the worker initiative.

[0056] Figure 8 is the cross-grid number calculation schematic diagram of the present application. DETAILED DESCRIPTION

[0057] The present application will be further described in detail below in combination with the drawings.

[0058] Embodiment 1

[0059] Referring to Figure 1 , the Stackelberg dynamic heterogeneous task allocation method for crowd-sensing of the present embodiment comprises:

[0060] obtaining task data, requester historical data and worker historical data of the crowdsourcing task;

[0061] Calculate a task allocation strategy index according to task data, requester historical data and task historical data;

[0062] Balance the task allocation strategy index based on the improved firework algorithm;

[0063] Perform user task response allocation prediction according to the task allocation strategy index;

[0064] Obtain a task response allocation result.

[0065] Specifically,

[0066] The application uses five real data sets. Respectively: ① Uber New York City ride data set, containing about 18.8 million ride records and rental company data in 2014-2015; ② Nice Ride shared bicycle data set, containing usage times, riding times and other attributes; ③ Credit card fraud data set, containing 284807 European credit card transactions, 492 fraudulent transactions, and PCA transformation; ④ Shenzhen main board stock daily line data set, containing three kinds of price data of 466 stocks; ⑤ American health insurance market data set, covering insurance scope, type, rate and other information.

[0067] The first aspect of the application proposes a Stackelberg two-stage game model for the platform, the worker and the requester, so that the interests of the three parties are maximized.

[0068] The first stage of the game:

[0069] The platform formulates relevant strategies for the needs of workers and requesters. The platform first formulates corresponding rules in combination with the needs of the three parties, and preliminarily obtains the optimal solution of the platform strategy that maximizes the interests of the platform.

[0070] S1. The relevant indexes required for the platform to formulate strategies for workers are as follows:

[0071] S1.1. Relevant indexes when workers are matched with tasks. When selecting suitable workers, the platform mainly needs to consider the skills, geographical location, historical performance and reputation value of the workers. The higher the quality of these index data, the more suitable the workers are for the task. The index calculation process is as follows:

[0072] S1.1.1. Calculate the matching degree of worker skills and tasks. Let the worker skill set S={s1,s2…sm}, the project demand set N={n1,n2…nk}, where m represents the number of workers and k represents the number of tasks. Therefore, the project demand coverage rate is represented as:

[0073] (1)

[0074] S1.1.2. Calculate the matching degree of worker and task geographical location using Manhattan distance. Let the worker geographical location task geographic location The Manhattan distance between the worker and the task is represented as:

[0075] (2)

[0076] S1.1.3 Evaluate worker historical data. In order to improve the quality of the worker group's perception, the worker's historical performance needs to be considered as an important basis for the platform to assign tasks. Let Therefore, the historical performance is:

[0077] (3)

[0078] At this point, the worker's skill matching degree, geographic location matching, and historical data have been obtained. Weight analysis is performed on these three data, and the improved fireworks algorithm (which will be explained later) is used for optimization to obtain the optimal worker and task matching result, i.e., several ternary optimal solutions.

[0079] S1.2 Calculate the worker's reputation value. Let Q represent the task completion quality (Q ∈ [1, 10]), and g represent the worker's positive evaluation rate, i.e., the number of positive evaluations. The worker's reputation value is represented as:

[0080] (4)

[0081] S1.3 The indicators required for the platform's incentive strategy for workers. Including:

[0082] S1.3.1 To ensure the selection of a high-quality worker group, the platform evaluates the worker's comprehensive performance. The platform's evaluation formula for the worker is:

[0083] (5)

[0084] One of the key requirements is must be strictly less than or equal to This means that tasks that are not completed on time will not be included in this evaluation item to ensure the accuracy and fairness of the incentive mechanism.

[0085] S1.3.2 Considering the worker's motivation to perform tasks, the worker's behavior response rate is quantified based on task pricing and historical performance data. The worker's behavior response rate is represented as:

[0086] (6)

[0087] ​​​​where P represents the pricing strategy of the platform, and its specific decision-making process will be elaborated in subsequent chapters.

[0088] Based on the above evaluation and quantification, the worker reward coefficient is determined to represent the additional reward index that the worker can obtain for the next task. The reward coefficient can be expressed as:

[0089] (7)

[0090] where Δ is a constant that can obtain more rewards when the number of worker tasks is greater than the average level.

[0091] After weight analysis and algorithm optimization of the above two indicators, the optimal worker reward coefficient is obtained. It is used to calculate the additional reward that the worker can obtain in addition to the task remuneration.

[0092] S2. The relevant indicators required by the platform to develop strategies for requesters are as follows:

[0093] S2.1 The pricing coefficient of the platform for the task. Including:

[0094] S2.1.1 Calculate the task complexity according to the task difficulty and task type. Let Th: task difficulty level [1, 10], Tl: task technical level requirement type [1, 10], then the task complexity is expressed as:

[0095] (8)

[0096] S2.1.2 Quantify the urgency of the task. Let : task required time, : task remaining time, : task published time, I: task importance (I ∈ [1, 10]). Therefore, the urgency is expressed as:

[0097] (9)

[0098] where when ≠ , represents the proportion of task remaining time to total time. As the remaining time decreases, increases, and values will increase, indicating that the urgency of the task is increasing. When = , that is, the task is published on the same day, the urgency of the task is mainly affected by and I. The greater the I is, the lower the task urgency is, the greater the task importance is, and the higher the task urgency is.

[0099] The two indexes are analyzed by weight, and the optimal pricing coefficient is obtained by using the algorithm of the application. The task pricing in different scenarios is calculated.

[0100] S2.2 Platform required indexes for incentive strategy of requester. Including:

[0101] S2.2.1 According to the difficulty division and different coefficients of the tasks published and completed by the requester, the requester experience value is calculated. The number of tasks published and completed by the requester is PF. In order to avoid the requester from frequently publishing easy-to-complete tasks to improve the experience value, the application divides the tasks into two categories according to the task difficulty, and calculates the requester experience value. The number of tasks with task difficulty in [1, 5] is , the number of tasks with task difficulty in [6, 10] is , and the application sets the influence index to 0.32 and 0.68 respectively. Then the requester experience value is:

[0102] (10)

[0103] S2.2.2 According to the historical data of the requester on the platform, the requester reputation value is calculated (the requester with negative reputation value cannot participate in the incentive mechanism). Let PO represent the number of tasks that the requester does not pay the task reward in time after the worker completes the task, is the number of overdue payments, is the number of tasks that the requester has never paid. Then the requester reputation value is represented as:

[0104] (11)

[0105] Based on the above obtained data, the requester reward coefficient is determined, which represents the discount index that the requester can obtain when publishing the next task.

[0106] All the parts related to algorithm optimization are optimized by using the firework algorithm. The improved algorithm is described in the following text. The optimization target is to maximize the platform benefit, and the optimal solution that maximizes the platform benefit is obtained.

[0107] The second stage of the game:

[0108] Based on the optimal solution of the platform related indexes obtained above, the second stage of the game, i.e. the worker and requester decision-making stage, is entered.

[0109] To better reflect how to give reasonable pricing of tasks under different conditions, the present application uses a payment matrix to embody the comparison of task pricing under various conditions. Taking one of the task types as an example, assume that its market base pricing is 10. Combining the pricing coefficient obtained in the first stage, the payment matrix C is obtained:

[0110] ,

[0111] The second aspect of the present application constructs three groups of user historical behavior feature parameters, establishes a user behavior preference constraint model, and minimizes the influence of user negative decision factors. First, a risk quantification evaluation model is used to analyze the influence of potential risks of the platform on the response of user behavior. The successful execution of a task cannot be separated from the comprehensive quality of workers and the reasonable operation of the platform, so the present application sets a platform reputation value evaluation mechanism to more accurately reflect the ability of the platform to protect the interests of users. The platform reputation value calculation formula is:

[0112] (12)

[0113] wherein represents the score given by the worker to the platform, represents the score given by the requester to the platform, represents the number of workers, represents the number of requesters.

[0114] The task allocation mechanism model, worker reputation value and platform reputation value are combined to construct a risk perception model. The requester risk perception coefficient formula is expressed as:

[0115] (13)

[0116] wherein is the requester risk perception coefficient, whose value is between 0 and 1, indicating the degree of risk perception of the requester, i.e. the probability of the requester abandoning the uploading of the task this time. 、 and are adjustment parameters for controlling the influence degree of each factor on the risk perception coefficient.

[0117] The formula uses a Sigmoid function ( ), which can map the input value to between 0 and 1, thereby realizing a smooth threshold effect. The Sigmoid function is used here to convert the original score of each factor into a part of the risk perception coefficient. Therefore, the formula can be expressed as:

[0118] (14)

[0119] In this formula, each Sigmoid function independently processes a factor and multiplies the result by its corresponding weight coefficient. In this way, each factor jointly influences the risk perception coefficient through the Sigmoid function and the weight coefficient.

[0120] The average risk perception coefficient is obtained , which is used to predict how many of every 100 requesters will abandon the uploading task.

[0121] The worker risk perception coefficient is evaluated according to the platform reputation value and the requester reputation value, and is expressed as:

[0122] (15)

[0123] Similarly, the requester risk perception coefficient is evaluated according to the platform reputation value and the worker reputation value, and is expressed as:

[0124] ,

[0125] The risk perception coefficient is used to verify whether the decision results of the workers and the requesters after the platform makes a benefit maximization decision in the first stage meet the expectations of the platform, thereby verifying the feasibility of the model of the present application.

[0126] wherein represents the requester reputation value.

[0127] The user behavior response function is constructed according to the indicators obtained in the first stage. The requester behavior response function is affected not only by the risk perception coefficient and the task urgency, but also by the expected cost. That is, the present application introduces a second user individualized demand theory, the expected utility theory. If the task pricing exceeds the requester's budget, the requester will evaluate the value of the excess part, and if increasing the budget can bring satisfactory utility, the requester may choose to increase the budget. Therefore, the requester behavior response function is expressed as:

[0128] (16)

[0129] wherein represents the task urgency, represents the requester budget. The budget limit of the requester is expressed as represents the requester budget. When the task pricing exceeds the requester's budget, the requester will abandon the task, at which time takes the value of 0, otherwise 1.

[0130] The worker behavior response function is constructed according to the above obtained indicators related to the worker. The function is expressed as:

[0131] (17)

[0132] wherein represents the worker skill and task adaptation ratio obtained by strategy one, represents the reward coefficient of the worker obtained by strategy two, Similar to the requester behavior response function, if the reward is lower than the worker's expectation, then is 0, otherwise, it is 1. represents the number of grids the worker crosses when performing the task, which is obtained based on the Manhattan distance in strategy one.

[0133] The reward weight obtained by the worker crossing the grid can be obtained according to the Manhattan distance The formula is:

[0134] (18)

[0135] wherein represents the reward weight obtained by the worker crossing the grid. The more the number of grids crossed, the more the additional reward obtained.

[0136] After obtaining the behavior response functions of the requester and the worker, the decision results of the requester and the worker can be predicted, that is, a set of optimal solutions of the requester and the worker strategies is obtained. The behavior response functions calculated in the foregoing are used to verify the final effect of the model of the application - that is, the maximum benefit of the platform is prioritized, and the decision of the worker and the requester is not affected too much.

[0137] At this time, the Nash equilibrium solution of the platform, the requester and the worker is obtained, so the three-party benefit function can be constructed to obtain the platform profit, the worker profit and the requester cost finally output by the model. The three-party benefit function is represented as:

[0138] (19)

[0139] wherein represents the additional reward obtained by the worker as a group leader. The application proposes to assign a group leader to coordinate the project work. The selection of the group leader is based on the worker reputation value ranking, and the worker with the highest reputation value is preferentially invited to serve as the group leader and obtains an additional reward; is the additional salary paid by the requester to the worker ∈[0.1,10] in dollars, that is, the reward mechanism set by the application, that is, the third user individual demand theory introduced by the application - dynamic reward feedback mechanism. The dynamic reward feedback mechanism is designed to compensate the coordination of the requester and the worker behavior through additional salary, and promote stable cooperation. It clearly defines the work quality and reward expectation, reduces misunderstanding, and improves efficiency. The application embodies this theory through additional salary, helps both parties to flexibly respond to challenges, and realizes win-win in the mobile crowdsourcing environment. The extra reward for the worker who is the group leader is set to 10% of the task price in the present application, and the proportion of the total payment value that the worker can obtain is set to 0.6.

[0140] The three-party benefit function is used to dataize the model result of the present application, that is, to intuitively show the final calculation result of the profits of the platform and the workers and the cost of the requester under real data set simulation.

[0141] In the third aspect, the present application provides an improved fireworks algorithm, that is, the greedy slot machine fireworks algorithm.

[0142] First step:

[0143] The present application aims to maximize the matching degree of worker skills and tasks. First, 500 groups of candidate data are randomly selected, each group of data being a three-tuple (cov, mat, H), denoted as F. The value of the matching degree of worker skills and tasks is used as the fitness index, and three weights are randomly generated to calculate the fitness value of the 500 groups of data. Subsequently, the greedy algorithm is used to select the 200 groups of data with the highest fitness, and the average fitness of these data is calculated. Through 1000 iterations, the iteration with the largest average fitness is found, and the 200 groups of data in this iteration are included in the initial fireworks population. In order to prevent the algorithm from converging to a local optimal solution and ensure the diversity of the solution, the present application further randomly selects 100 groups of data from the 500 groups of data and adds them to the initial fireworks population to broaden the search range of the optimal solution.

[0144] Second step:

[0145] After the initial population is determined, a set of optimal weight distribution is found through the random lottery mechanism of the slot machine. Similar to the first stage, the optimal weight distribution is determined by comparing the average fitness of the initial population obtained by each group of weight distribution, and the maximum fitness fmax and the minimum fitness fmin of the initial population under the optimal weight are calculated.

[0146] Third step:

[0147] According to the first two stages, 300 initial fireworks and three optimal weights ω1, ω2, ω3 have been obtained, and this stage enters the core link of GGMFA. First, set the maximum and minimum amplitude of the explosion of the fireworks, and the maximum and minimum number of sparks generated by the explosion. Set the explosion amplitude as A, and the number of sparks generated by the explosion as N_sparks, so the maximum and minimum amplitudes are A_max and A_min, and the maximum and minimum explosion spark numbers are N_sparks_max and N_sparks_min. The initial fireworks population is represented as Fire, Fire={F1, F2……Fi}. Set the fitness of the fireworks and the sparks generated by the explosion to be the value of M. The calculation formula of the amplitude and the explosion spark number of each fireworks is:

[0148] A = A_min + (A_max - A_min) * (M - fmin) / (fmax - fmin)

[0149] N_sparks = N_sparks_min + (N_sparks_max - N_sparks_min) * (M - fmin) / (fmax - fmin)

[0150] The sparks generated by the explosion are represented as Fij, where i represents the i th fireworks, and j represents the j th spark generated by the i th fireworks explosion, j∈[1, N_sparks].

[0151] The invention sets the explosion mode of the fireworks as the coordinate axis explosion, and the coordinates of each fireworks are set as (cov, mat, H). The mutation of the sparks is set to follow the Gaussian distribution, the mean is 0.5, the standard deviation is 0.1, and the value range of each attribute of the mutated sparks is limited to [0, 1], so the mutated sparks are represented as . Update all the data in the candidate pool to the mutated sparks, and select the sparks in the candidate pool. The invention adopts the elite reservation strategy, that is, select 20 sparks each round, compare the fitness of the 20 sparks with the fireworks in Fire, and replace the fireworks in Fire with the sparks with higher fitness. Repeat the explosion, mutation and selection process, and the final Fire obtained after the iteration is terminated is a set of optimal solutions of the worker skill and task matching degree.

[0152] The above describes the experimental process of a round of data set. Considering the dynamic characteristics of platform data, the present application adopts three different data sets to simulate the real-time dynamics of platform data. Through three rounds of experiments, three sets of optimal solutions are obtained, and these optimal solutions are fused and summarized to form a comprehensive optimal solution of worker skill and task matching degree. In addition, the weight coefficients of some formulas with determined numerical values are calculated by another optimization algorithm, bandit algorithm. In order to demonstrate the weight calculation process, the present application takes the requester experience value as an example.

[0153] The requester experience value is represented as , which is simplified as . First, the initial weight is 0.5, then a number of rounds of experiments are performed, and a set of AB data is introduced in each round to observe the change of J value. The goal of the present application is to maximize the requester experience value, that is, to maximize the value of J. Therefore, if the change of A is higher than the recent average level, the weight is increased (for example, +0.01), and if the change of A is lower than the recent average level, the weight is reduced (for example, -0.01). Repeat for several rounds until the weight changes little (for example, the adjustment amplitude is less than 0.01), then stop. The weight distribution obtained at this time is the optimal distribution.

[0154] In the fourth aspect, the present application introduces the Quality of Service (QoS) index to analyze the optimal data output by the model, ensuring that the high-quality worker group is selected to provide services for the requester, guaranteeing the task execution efficiency and completion quality, and helping to ensure the long-term stability of platform operation from the root.

[0155] As shown in Figure 1 , the system running process is as follows: the requester first publishes the project requirements to the platform, the platform evaluates and classifies each task in the project and assigns it to the appropriate worker, and the worker chooses whether to execute the task. In order to ensure the overall completion efficiency of the project, the applicable scenario of the present application is multi-worker and multi-task, that is, each project published by the platform contains multiple tasks, each task corresponds to a worker, and under the leadership of the project leader, the workers cooperate to complete the project, during which they may interact with the requester. After the task is completed, the data is returned to the platform, and the platform feeds back the results to the requester. In MCS, the needs of participants are different, leading to conflicts of interest. Therefore, the present application adopts the Stackelberg game model, positions the platform as the leader, predicts the needs of workers and requesters, and formulates operation strategies. Workers and requesters as followers adjust their responses according to the platform strategy. The strategy formulation is divided into two stages, as shown in Figure 2The present application focuses on the heterogeneous multi-project multi-task assignment problem (HMPMTA) in the MCS, which is NP-hard and involves diverse attributes of projects, tasks, teams and workers. Although existing researches complete tasks through team collaboration, they limit workers to participate in only one task in a project, ignore the cases of insufficient worker recruitment and multi-skilled workers, and do not consider worker priority when selecting team leaders, which may lead to the selection of suboptimal team leaders. Therefore, the present application proposes an assignment process that satisfies the following conditions:

[0156] 1) Each project consists of multiple heterogeneous tasks, each with specific skill requirements and can only be completed by workers with corresponding skills.

[0157] 2) The platform performs two rounds of task assignment. The first round aims to maximize task completion efficiency, with each worker being assigned only one task in each project, but can participate in multiple projects. The second round removes this restriction, allowing workers to participate in multiple tasks within the same project.

[0158] 3) To coordinate project work, a team leader is assigned. The selection of team leaders is based on the ranking of worker reputation values, with the worker with the highest reputation value being invited to serve as the team leader first and receiving an additional reward.

[0159] Based on the above description, the present application will describe the game process of the constructed model, as shown in Figure 2 The starting stage of the game is the platform decision-making link, i.e., the leader formulates the rules. In view of the diverse needs of the participants, the present application customizes corresponding platform strategies for workers and requesters, and optimizes each strategy using the fireworks algorithm to obtain a set of optimal decision solutions. At the end of this stage, a set of optimal solutions that balance the interests of the three parties and prioritize the maximization of platform interests is obtained. The second stage is the decision-making stage of the followers (i.e., workers and requesters). First, a payment matrix is generated to clearly display the payment amounts obtained from various decision-making of each task type. Then, three user individualized demand theories are introduced into the follower decision-making process to obtain the behavior response function of the followers and the optimal decision prediction solution of the followers. At this time, the interests of the platform, workers and requesters have reached Nash equilibrium. According to the optimal decisions of the three parties, the expected profits of the platform and workers and the cost situation of the requesters can be calculated according to the interest calculation formula. The cross-grid reward calculation method in the worker profit is as follows Figure 3 The present application uses a (10*10) grid division method to more intuitively represent the location of workers and tasks, and measures the distance between workers and tasks using Manhattan distance. In addition, if a worker and a task are in the same grid, the distance between them is ignored. If a worker needs to cross a grid to perform a task, an additional reward is given according to the number of grids crossed. The incentive mechanism strategy of the present application selects 200 worker data for cross-grid calculation, as shown in Figure 8The horizontal axis represents the number of workers crossing the grid, i.e., d (x) is rounded down. , The results show that most of the 200 workers cross the grid 0-2 times. The probability of workers choosing to cross the grid is 39.5%, and the average cross-grid reward is 0.81$. This shows that workers have a preference for geographical location, and the platform can maintain balance in task allocation. The diversity of the number of workers crossing the grid shows that the incentive mechanism improves participation and task completion efficiency, proving the feasibility and effectiveness of the incentive mechanism strategy in promoting resource optimization in the MCS environment.

[0160] Figure 4 is a comparison chart of the experimental results of the GGMFA algorithm proposed in the application and other algorithms, wherein Figure 4 (a) in the application represents a comparison chart of the experimental results of the worker reputation value, Figure 4 (b) in the application represents a comparison chart of the experimental results of the worker skill and task matching degree. From Figure 4 (a) and Figure 4 (b) in the application, it can be seen from the two charts that the worker reputation value and the worker skill and task matching degree have good performance under the optimization of the algorithm in the application. Through final calculation, the average value of the worker reputation value is 7.54 (the maximum value is 10), which shows that the overall reputation level of workers on the platform is high. Such a high-reputation worker group improves the user's task participation rate and loyalty through their high-quality service. Secondly, the average value of the worker skill and task matching degree is 0.87 (the maximum value is 1), which shows that the skills of workers are highly matched with the task requirements. This high matching not only improves the task allocation speed and completion efficiency, but also indirectly guarantees the quality of task completion. Figure 4 shows the change trend of the platform strategies under the influence of different indicators, which generally shows an upward trend, but may show a downward trend between adjacent points. The appearance of this local downward phenomenon may be because the selected indicator in the horizontal coordinate is the most influential, but under the influence of other indicators, there may still be a decrease between adjacent points, but this does not affect the overall upward trend.

[0161] Figure 5 shows the comparison of the probability of positive selection of whether the conditions meet the user's demand (i.e., the probability of successful task allocation and execution). The experimental results show that after the optimization of the platform strategy by Stackelberg two-stage game and GGMFA algorithm, and the prediction of different reactions of workers and requesters to different needs and user historical behavior characteristics in the second stage, the average probability of task execution by workers is 89.14%, and the average probability of task execution by workers is 89.14%. Figure 5 (a) andFigure 5 (b) in FIG. 5 shows that task pricing has little effect on the decisions of workers and requesters, indicating that the platform's pricing adjustment strategy is flexible and can be adjusted in real time to meet the needs of both parties. The selection probability of requesters and the task urgency change little, and they pay more attention to the completion time than the price. The selection probability of workers increases slightly with the increase of the additional reward, but it is stable overall, and the salary expectation fluctuates within a certain range. Long-term work enables workers to have an accurate understanding of their market value, and stable employment opportunities and good environment also affect their decisions.

[0162] Figure 6 The left y-axis is the comparison between the task base price and the actual task pricing under different task types. The right y-axis is the percentage of the actual task pricing exceeding the base price. By taking the base price as the cost, it can be seen that the actual task pricing is usually higher than the base price, because the platform will adjust the price according to factors such as task difficulty, urgency, and worker skills. This price adjustment mechanism not only ensures the platform's profit and encourages workers to participate in tasks, but also ensures that workers receive reasonable remuneration.

[0163] The present application is carried out in a dynamic environment, and three experimental data sets are constructed from five real data sets to verify the stability of the GGMFA algorithm. As shown in FIG. 6, Figure 7 three key indicators are selected to verify the stability and adaptability of the GGMFA algorithm on the three experimental data sets. Among them, Figure 7 (a) in FIG. 7 shows the comparison of the optimization effect of the platform's evaluation of workers, Figure 7 (b) in FIG. 7 shows the comparison of the optimization effect of task pricing, Figure 7 (c) in FIG. 7 shows the comparison of the optimization effect of worker enthusiasm. Although there are differences in the experimental data sets, the fluctuation range of the three indicators in the three experiments is small, indicating that the GGMFA algorithm can maintain stable strategies when facing the diversity and dynamics of data sets.

[0164] Embodiment 2

[0165] The present embodiment provides a Stackelberg dynamic heterogeneous task allocation system for crowd sensing, comprising:

[0166] The data acquisition module is configured to acquire task data, requester historical data and worker historical data of the crowdsourcing task;

[0167] The strategy index module is configured to calculate the task allocation strategy index according to the task data, the requester historical data and the task historical data;

[0168] The balance module is configured to balance the task allocation strategy index based on the improved firework algorithm;

[0169] The prediction module is configured to make user task response allocation prediction according to the task allocation strategy index;

[0170] The distribution module is configured to obtain a task response distribution result.

[0171] A computer readable storage medium, wherein a plurality of instructions are stored, the instructions are adapted to be loaded by a processor of a terminal device and execute the method.

[0172] A terminal device, comprising a processor and a computer readable storage medium, the processor is used to implement instructions, the computer readable storage medium is used to store a plurality of instructions, the instructions are adapted to be loaded by the processor and execute the method.

[0173] The above are preferred embodiments of the present application, not limited by the protection scope of the present application, therefore: all equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1.A Stackelberg dynamic heterogeneous task allocation method for crowd sensing, characterized in that, The method comprises the following steps: obtaining task data, requester historical data and worker historical data of a crowdsourcing task; calculating a task allocation strategy index according to the task data, the requester historical data and the worker historical data; balancing the task allocation strategy index based on an improved firework algorithm; performing user task response allocation prediction according to the task allocation strategy index; obtaining a task response allocation result; The task allocation strategy index is calculated according to the task data, the requester historical data and the worker historical data, and the matching degree between the task and the worker is calculated according to the task data and the worker historical data, wherein the project demand coverage is calculated according to the worker skill set and the project demand set, the matching degree between the worker and the geographic position of the task is calculated by using the Manhattan distance, the geographic position of the worker is L w =( x w , y w ), the geographic position of the task is L t =( x t , y t ), the Manhattan distance between the worker and the task is represented as: d(L w , L t )=∣x w -x t ∣+∣y w -y t ∣, and the worker historical data is evaluated, the worker historical performance is taken as an important basis for the platform to allocate tasks, and it is assumed that , the worker historical performance is: , the worker reputation value is calculated, Q represents the task completion quality, g represents the worker praise rate, i.e. the number of positive evaluations, and the worker reputation value is represented as: ； The method further comprises the following steps of calculating a worker task reward coefficient according to the task data and the worker historical data, wherein a worker reputation value is calculated according to a task completion quality and worker historical good review data; a worker behavior response rate is calculated according to the task data and worker historical performance data; and the worker reward coefficient is determined based on the worker reputation value and the worker behavior response rate, and the reward coefficient is expressed as: ， wherein Δ is a constant, eva represents an evaluation of the platform on the worker, and pos represents the worker behavior response rate; The method further comprises the following steps of calculating a requester reputation value according to the task data and the requester historical data, wherein a requester experience value is calculated according to a difficulty division and different coefficients of tasks published and completed by the requester; and the requester reputation value is calculated according to historical data of the requester on the platform, wherein PO represents a number of tasks for which the requester fails to pay a task reward in time after the worker completes the task, PO1 represents a number of tasks for which the requester fails to pay the task reward within a deadline, and PO0 represents a number of tasks for which the requester fails to pay the task reward at all, and the requester reputation value is expressed as: ； The method of balancing the task allocation strategy index based on the improved firework algorithm comprises the following steps of taking maximization of a worker skill and task matching degree as a target, taking a value of the worker skill and task matching degree as a fitness index, calculating an average fitness after screening by using a greedy algorithm, and obtaining an initial firework population and an optimal weight through iteration, and setting a maximum and a minimum amplitude of firework explosion, and a maximum and a minimum number of sparks generated by the firework explosion, and the formula for calculating an amplitude and a number of sparks of each firework is: A = A_min + (A_max - A_min) * (M - fmin) / (fmax - fmin) N_sparks = N_sparks_min + (N_sparks_max - N_sparks_min) * (M - fmin) / (fmax - fmin) The sparks generated by the explosion are represented as Fij, where i represents the i th firework, j represents the j th spark generated by the explosion of the i th firework, j∈[1,N_sparks], A_max and A_min represent the maximum and minimum amplitudes of the firework respectively, N_sparks_max and N_sparks_min represent the maximum and minimum numbers of sparks generated by the explosion of the firework respectively, fmax and fmin are the maximum fitness and the minimum fitness in the initial population under the optimal weight respectively; the mutation of the sparks is set to follow a Gaussian distribution, all the data in the candidate pool of sparks are updated to the mutated sparks, the sparks in the candidate pool are selected and iterated, and the final Fire obtained after the iteration is terminated is a set of optimal solutions of the matching degree of the worker skills and the task; The user task response allocation prediction according to the task allocation strategy index comprises calculating a platform reputation value based on worker historical data and requester historical data, wherein the platform reputation value is calculated based on a score given to the platform by the worker and a score given to the platform by the requester, calculating a requester risk perception coefficient based on the platform reputation value and the worker reputation value, and is represented as: wherein is the requester risk perception coefficient, , and is an adjustment parameter, M represents a task adaptation ratio of a skill of the worker obtained by the strategy, represents the platform reputation value; and calculating a worker risk perception coefficient based on the platform reputation value and the requester reputation value, and is represented as: ； The user task response allocation prediction according to the task allocation strategy index further comprises calculating a requester behavior response function according to the task allocation strategy index, represented as: ， wherein represents the urgency of the task, represents the requester budget; The user task response allocation prediction according to the task allocation strategy index further comprises calculating a worker behavior response function according to the task allocation strategy index, represented as: ， wherein represents the ratio of the worker's skill to the task in the strategy one, represents the reward coefficient of the worker in the strategy two, and the requester behavior response function is the same, represents the number of grids the worker crosses when performing the task. 2.The Stackelberg dynamic heterogeneous task allocation method for crowd sensing, according to claim 1, wherein, The user task response allocation prediction according to the task allocation strategy index further comprises predicting the optimal solution of the task allocation strategy based on the worker behavior response function and the requester behavior response function, and verifying the effect of the task allocation strategy by using the requester risk perception coefficient and the worker risk perception coefficient. 3.A Stackelberg dynamic heterogeneous task allocation system for crowd sensing, which executes the Stackelberg dynamic heterogeneous task allocation method for crowd sensing according to claim 1. It comprises: A data acquisition module configured to acquire task data, requester historical data and worker historical data of a crowdsourcing task; A strategy index module configured to calculate a task allocation strategy index according to the task data, the requester historical data and the worker historical data; A balance module configured to balance the task allocation strategy index based on the improved firework algorithm; A prediction module configured to predict user task response allocation according to the task allocation strategy index; An allocation module configured to obtain a task response allocation result.

Images