Mobile Crowd Sensing Task Allocation Method Based on Non-Explicit Manipulation Mechanism
By constructing a sequential sub-module combinatorial optimization network with non-obvious manipulation mechanisms, the problems of user heterogeneity utilization and budget feasibility in mobile swarm intelligence sensing task allocation are solved. This achieves efficient task allocation and user selection, explicitly models the value changes of sequential sub-modules, suppresses policy manipulation, and improves the overall value of the system.
Patent Information
- Application Number
- CN202611268794.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-10
AI Technical Summary
Existing mobile crowd sensing task allocation methods struggle to fully leverage user heterogeneity in scenarios with large-scale users and limited resources. Traditional methods are inefficient, deep learning methods struggle to explicitly model changes in the value of sequential sub-modules, and strict incentive compatibility mechanisms fail to meet budget feasibility and individual rationality requirements in real-world environments.
A sequential sub-module combinatorial optimization network based on a non-obvious manipulation mechanism is adopted. Through a user rating module, a stopping decision unit, a relaxation network, a resource cost calculation module, and a global budget pruning module, a differentiable sequential selection network is constructed. The marginal value diminishing characteristic of sequential dependence is explicitly modeled, and the stability of the mechanism is guaranteed by extreme adversary configuration verification.
It significantly increases the total value of the system under different user and task scales, effectively suppresses user strategy manipulation behavior, achieves a balance between budget feasibility and user rationality, and improves allocation efficiency.
Smart Images

Figure CN123070580A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mobile crowd sensing technology, and more specifically, relates to a mobile crowd sensing task allocation method based on a non-obvious manipulation mechanism. Background Technology
[0002] Mobile Crowdsensing (MCS) systems collect data on urban environment, traffic conditions, communication quality, and public safety through user-carried mobile terminals, vehicle-mounted terminals, wearable devices, or other networked sensing devices. Task publishers typically set up several Points of Interest (POIs) or task areas on the task publishing platform, specifying the task's time, spatial scope, and budget constraints. Figure 1 This is a schematic diagram of a mobile crowd sensing process based on points of interest. For example... Figure 1 As shown, the platform first releases multiple perception tasks with spatiotemporal constraints. Mobile users submit quotes and route information based on their own trajectories and execution costs. The platform then selects suitable users to execute tasks under budget constraints and pays users based on task completion.
[0003] In real-world engineering scenarios, such systems face the following characteristics: 1. User trajectories exhibit spatial and temporal heterogeneity, with different users contributing different values to the same task; 2. Task data overlaps, resulting in different marginal contributions from users who execute earlier and later; 3. The platform has a limited budget, requiring the completion of user selection, user execution order ranking, and payment generation within the total budget; 4. User bids are strategic, as users may influence their selection and payment by adjusting their bids; 5. Existing mechanisms primarily employ a dominant incentive compatibility mechanism, which requires that a user's gain under any strategic bid does not exceed the gain under an honest bid, potentially leading to a narrow system selection space and reduced overall perceived value.
[0004] Existing methods for assigning and incentivizing mobile crowdsourcing sensing tasks typically include greedy selection, threshold payment, key payment, online recruitment, and deep learning-based assignment-payment models. These methods suffer from the following problems when implemented in engineering:
[0005] 1. Traditional manual design methods have limited allocation efficiency and optimization capabilities. Traditional methods typically rely on manually designed greedy rules, threshold rules, or heuristic strategies. In scenarios with a large user base, complex tasks, and limited resources, it is difficult to fully utilize the heterogeneity and combinatorial relationships among users, thus making it difficult to obtain efficient allocation results.
[0006] 2. The integration of deep learning-based mechanisms with the perceived value of sequential sub-modules remains challenging. Deep learning can improve decision-making efficiency in large-scale scenarios by learning complex allocation strategies through data-driven approaches. However, existing learning-based allocation methods typically output continuous selection scores or allocation probabilities and do not explicitly generate user execution order. Therefore, they struggle to characterize the changes in marginal perceived value during progressive selection and to directly integrate sequential sub-module structures into trainable allocation networks. Currently, there is a lack of methods for jointly learning user selection and execution order and explicitly modeling the perceived value of sequential sub-modules.
[0007] 3. Deep learning mechanisms struggle to structurally guarantee economic properties. Existing methods typically use budget constraints, individual rationality, or incentive constraints as penalty terms in the loss function, allowing the model to approximately satisfy these properties during training. However, such soft constraints cannot guarantee that the model consistently meets requirements for budget feasibility, individual rationality, and reasonable payment under any real-world input, potentially leading to over-budget or unreasonable payment outcomes in out-of-distribution scenarios.
[0008] 4. Traditional mechanism design is mainly based on the dominant incentive compatible mechanism (DSIC), which is too demanding and may sacrifice actual efficiency. Traditional mechanisms require users to be able to identify honest reporting as the optimal strategy through complete reasoning in any competitive environment. This places excessive demands on the cognitive abilities of actual MCS users. At the same time, excessively strong incentive constraints may also compress the feasible allocation space and limit the platform's utilization of user heterogeneity and trajectory information. Summary of the Invention
[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a mobile crowd-sensing task allocation method based on a non-obvious manipulation mechanism. For the mobile crowd-sensing reverse auction scenario, a task allocation system based on a non-obvious manipulable (NOM) mechanism is proposed. This system retains the stability constraints on user strategy behavior, limits users from obtaining better benefits through obvious manipulation behavior by verifying extreme adversary configurations, and achieves an engineering balance between budget feasibility, user participation rationality, and allocation efficiency by adopting sequential task allocation and budget pruning.
[0010] To achieve the above-mentioned objectives, the mobile crowd sensing task allocation method based on a non-obvious manipulation mechanism of the present invention includes the following steps:
[0011] S1: Construct a sequential sub-module combinatorial optimization network, including a user rating module, a stopping decision unit, a relaxation network, a resource cost calculation module, a candidate assignment set generation module, and a global budget pruning module, wherein:
[0012] The user rating module is used to rate users based on iteration rounds. User state matrix for each active point of interest Generate user priority score vector And send it to the candidate allocation set generation module, , Indicates the maximum number of iterations. , Indicates the number of users. , Indicates the number of points of interest. Indicates the iteration round A collection of active points of interest ; Indicates the iteration round Each user For each point of interest The state characteristics are expressed as follows:
[0013] ,
[0014] in, This represents vector concatenation. Indicates user Statement of interest Resource overhead of the task Indicates user Points of interest The value of the task Indicate points of interest The number of remaining users with demand, Indicate points of interest Number of users selected for the task This indicates that the selected user has selected points of interest. Average resource cost of the task This indicates that the selected user has selected points of interest. Average contribution value of the task Indicates the remaining resource budget;
[0015] Indicates user For each point of interest Priority score;
[0016] The stopping decision unit is used to determine the iteration round. The assigned state vector of each active interest point Generate stopping score And send it to the candidate allocation set generation module; The expression is as follows:
[0017] ;
[0018] Relaxed networks are used to adjust the user state matrix. Generate resource cost score vector And send it to the resource overhead calculation module;
[0019] The resource cost calculation module is used to calculate costs based on points of interest. Corresponding user-declared resource overhead vector and resource cost score vector Generate resource cost vector And send it to the candidate allocation set generation module;
[0020] The candidate allocation set generation module is used to generate user priority score vectors based on active points of interest. Stop fractions and resource overhead vector Generate iteration rounds Candidate allocation set The data is then sent to the global budget pruning module. The specific method for generating the candidate allocation set is as follows:
[0021] Based on resource overhead vector Get each active point of interest candidate user set :
[0022] ;
[0023] User priority score vector After correction, the corrected user priority score vector is obtained. :
[0024] ;
[0025] The corrected user priority score vector and stopping fractions splicing together to obtain motion features Then, the Gumbel-Softmax function is used to analyze the action features. The process is performed to obtain the action score vector. , , Represented as points of interest Select User The score, This indicates that the user's score has been stopped from being selected;
[0026] If the score Points of interest Stop selecting users, otherwise select action score. Users and points of interest corresponding to the median maximum value Task composition - user interaction and add to the candidate allocation set. ;
[0027] The global budget pruning module is used to prune resources based on the budget. For candidate allocation set Perform trimming to obtain the final allocation set. The cutting method is as follows:
[0028] First, based on the candidate allocation set Each candidate task - user pair Priority score Sort the candidate task-user pairs from largest to smallest. Then, use a greedy selection strategy to select the highest-scoring candidate task-user pairs and accumulate resource costs until the resource budget constraint can no longer be satisfied, and return the final allocation set. ;
[0029] The state update module is used to update the allocation set. To perform a state update, the specific method is as follows: for the allocation set Each task - user , assign identifier resource consumption , will users Add to Points of Interest Execution sequence Simultaneously update the allocation set. Each point of interest The number of remaining users with demand Remaining resource budget When there is interest In iteration rounds Stop selecting users or the number of remaining users with demand. Remove it from the set of active points of interest Delete the set of active points of interest When the set of active points of interest or The sequential sub-modulus combination optimization network stops operating;
[0030] S2: Collect several mobile crowd sensing task scenario samples to train the sequential sub-mode combination optimization network and obtain the trained sequential sub-mode combination optimization network.
[0031] S3: Obtain each point of interest in the current mobile crowd sensing task scenario. Number of users in demand Each user Points of interest The contribution value of the task ,user Statement of interest Task resource overhead And the resource budget for all sensing task information. ;
[0032] S4: Determine if there are any users in the current mobile crowd intelligence sensing task scenario who meet the extreme adversary configuration. If there are users who meet the extreme adversary configuration, proceed to step S5; otherwise, proceed to step S6.
[0033] S5: Invoke the closed allocation rule to execute the task - user pairing and allocation;
[0034] S6: Call the sequential submodulus combination optimization network execution task - user allocation, where interest points are initialized. The number of remaining users with demand Number of selected users Selected users' points of interest Average resource cost of the task Average contribution value Remaining resource budget .
[0035] This invention presents a mobile crowd sensing task allocation method based on a non-obvious manipulation mechanism. It constructs a sequential sub-modulus combinatorial optimization network comprising a user scoring module, a stopping decision unit, a relaxation network, a resource cost calculation module, a candidate allocation set generation module, and a budget pruning module. Several mobile crowd sensing task scenario samples are collected to train the sequential sub-modulus combinatorial optimization network, resulting in a trained network. The network then acquires interest point task information and user information for the current mobile crowd sensing task scenario. If a user meets the extreme opponent configuration, a closed allocation rule is invoked to perform task-user pairing allocation; otherwise, the sequential sub-modulus combinatorial optimization network is invoked to perform task-user allocation.
[0036] The present invention has the following beneficial effects:
[0037] 1) This invention proposes a framework for a non-obvious manipulation mechanism for mobile crowd-aware reverse auctions. For the first time, the concept of weak incentive stability (NOM) is systematically introduced into the budget-constrained mobile crowd-aware reverse auction scenario, providing a new trade-off between mechanism stability and system efficiency in complex environments without relying on strict incentive compatibility constraints.
[0038] 2) This invention proposes a differentiable sequential sub-modular combinatorial optimization network oriented towards the sequential sub-modular value structure. Unlike most existing learning mechanism design methods that use continuous allocation variables or probabilistic allocation methods to approximate discrete allocation processes, the sequential sub-modular combinatorial optimization network is a differentiable sequential choice network that can directly generate the user's discrete 0 / 1 choice results and execution order through a step-by-step decision-making approach. This explicitly models the diminishing marginal value characteristics of sequence dependence at the network level and achieves an end-to-end representation of the discrete sequential combinatorial optimization process. Simultaneously, this network directly guarantees individual rationality and budget feasibility at the network level through a structured resource cost calculation module and a budget pruning module.
[0039] 3) Based on the differentiable sequential selection network mentioned above, this invention further introduces a structured NOM mechanism module based on extreme adversary configuration, which realizes a learning-based reverse auction mechanism that satisfies the non-obvious manipulation (NOM) property. Unlike existing methods that mainly rely on regret optimization or incentive constraint penalty terms, this invention directly guarantees the stability of weak incentives through the mechanism structure, while the deep learning module is only responsible for optimizing the system allocation efficiency.
[0040] 4) Experimental results show that the present invention can achieve higher total system value under different user scales, task scales and budget conditions, and is significantly better than the traditional greedy mechanism and the deep learning mechanism based on strict DSIC. Further user behavior analysis shows that users are unlikely to gain a significant advantage in the best and worst benefit levels by deviating from the actual resource expenditure, which verifies the effective suppression of explicit policy manipulation behavior of the present invention. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of a mobile crowd sensing process based on points of interest.
[0042] Figure 2 This is a flowchart illustrating a specific implementation of the mobile swarm intelligence sensing task allocation method based on a non-obvious manipulation mechanism according to the present invention.
[0043] Figure 3 This is a structural diagram of the sequential sub-modulus combination optimization network in this invention;
[0044] Figure 4 This is a comparison chart of the total value of the present invention and the baseline method in different scenarios in this embodiment;
[0045] Figure 5 This is a comparison chart of the total resource overhead of the present invention and the baseline method in different scenarios in this embodiment;
[0046] Figure 6 This is a comparison chart of the number of selected users in different scenarios for the present invention and the baseline method in this embodiment;
[0047] Figure 7 This is a comparison chart of the total system value and total resource cost of the present invention and the baseline method under different resource budgets in this embodiment;
[0048] Figure 8 This is a typical allocation result diagram of the present invention under different budgets in this embodiment;
[0049] Figure 9 This is a graph showing the changes in resource overhead and utility of the present invention under extreme adversary configurations in this embodiment. Detailed Implementation
[0050] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.
[0051] To better illustrate the technical solution of the present invention, the principle of the present invention will be briefly explained first.
[0052] This invention assumes that the system is running at Within the timeframe, considering that data requesters need to collect data in multiple locations, each data collection area is defined as a Point of Interest (POI), therefore there are... , Represents a set of points of interest, each point of interest The perceived task information includes task location. Task release time Space-sensitive parameters Time-sensitive parameters And the number of users required for the task , , This indicates the number of POIs.
[0053] The set of users participating in this perception task data collection is denoted as . Each user's user trajectory can be recorded as follows: , , Indicates user At any moment Location, , This represents the length of the user's trajectory. To facilitate task allocation decisions, users also submit the resource costs for completing different tasks. , Indicates user interest points The resource overhead of perception is also the basis for paying users in practical applications. Therefore, user report information can be recorded as... In mobile crowd sensing, the information for each user can be calculated based on task requirements and user information. The collected data is for points of interest The value of contribution .
[0054] Since later-arriving users may provide information that overlaps with that of earlier users, earlier-arriving users bring higher value to the system. Therefore, even if the selected users are the same, the order in which users join affects the total value of the final data. Thus, for points of interest... If the execution sequence of the selected user set is , Indicates the execution of points of interest The first perception task User serial number , Indicates the execution of points of interest If the number of users sensing the task indicates that the execution sequence is at the point of interest... The value created Satisfying the order modulo property, it is expressed as follows:
[0055] .
[0056] Each point of interest The winning user set can be selected independently, and the same user can be selected by multiple points of interest simultaneously. Selected, therefore the total system value This can be represented as points of interest. The sum of values is expressed as follows:
[0057] ,
[0058] in, .
[0059] Therefore, in the mobile crowd sensing scenario considered in this invention, the system value not only satisfies the classic monotonous submodularity, but also further satisfies the more general ordered submodularity. This property characterizes the impact of the order in which users join on the system value and is an important theoretical basis for the subsequent design and optimization analysis of task allocation mechanisms.
[0060] All participants in the task will receive a report. Since the platform calculates user payments based on the determined resource costs, if the platform determines that the user... For the task Resource overhead is less than At this point, the user's reward is insufficient to cover their resource costs, so the user will abandon the task, and their reward will be zero. The true report information is recorded as The user-declared report type is denoted as In order to increase their profits, users may misreport their resource consumption on various Points of Interest (POIs). Therefore, it is crucial to design an effective mechanism to prevent users from manipulating the system.
[0061] When users The report information is The collection of report information from other users is denoted as At that time, the mechanism determines the user Allocation results on each POI With resource expenditure Record the allocation identifier , indicating user Whether or not they were assigned to a point of interest The perception task, if assigned and successfully executed, involves points of interest. The corresponding task and its corresponding resource overhead are denoted as follows: Then the user Points of interest The utility function is When the user The report information is (The user falsely claimed to be in various...) The price quoted at the location is Other users' report types are At that time, the user Points of interest utility function From the perspective of individual rationality, users At each point of interest The utility above must satisfy .
[0062] Previous research has often required strict Dominant-Strategy Incentive Compatibility (DSIC) in mechanism design, meaning that users can obtain the maximum benefit when they honestly report their costs or values under any circumstances, and any deviation from the strategy will not lead to greater utility. While such strict incentive-compatible mechanisms theoretically guarantee the stability of participant behavior, they have significant limitations in practice: on the one hand, they assume that users possess sufficient rationality and complete reasoning ability to accurately analyze the differences in benefits of different strategies under various circumstances, an assumption that is often difficult to hold in real-world mobile crowdsourcing environments; on the other hand, strict DSIC imposes strong constraints on the mechanism structure, making it difficult to achieve high system value under budget constraints. Therefore, in such scenarios, it is necessary to introduce a weaker concept of incentive stability. Non-obvious manipulation (NOM), initially proposed by Troyan and Morrill, has the core idea that not all profitable manipulations are cognitively "obvious." In complex mechanisms, users often only focus on the extreme consequences that a strategy might bring, rather than precise comparisons of conditional utility. For points of interest... The above provides accurate report information. ,user The maximum and minimum utility obtainable through honest reporting are defined as follows:
[0063] .
[0064] Similarly, when the user's statement report information is ,and The maximum and minimum utility of lying are similarly defined as follows: and In short, users can identify the best and worst outcomes of honest and lying scenarios, but they are not clear about the intermediate scenarios.
[0065] The mechanism satisfies the definition of Not Obvious Manipulation (NOM): if for any user Points of interest Any real report information And any false reporting information It must simultaneously meet the following two conditions:
[0066] (1) In the best-case scenario, it cannot be obviously manipulated (BNOM):
[0067] ,
[0068] (2) In the worst case, it cannot be obviously manipulated (WNOM):
[0069] .
[0070] Unlike DSIC, NOM only requires that any strategy of lying is no better than honest reporting in either the best or worst case, thus providing a more relaxed incentive constraint. In mobile crowdsourcing sensing systems, users face a highly uncertain competitive environment: the trajectories, bids, and selection probabilities of other users are often difficult to predict accurately. In this context, users typically struggle to make precise conditional utility comparisons when evaluating strategies, preferring instead to rely on simple and robust decision criteria, such as whether a strategy might lead to a higher best-case gain or a worse worst-case outcome. Non-obvious Manipulation (NOM) is based on this decision-making characteristic, imposing constraints on user-perceptible extreme outcomes at the mechanism level, thereby excluding manipulative behaviors that are cognitively obvious and of low risk. By ensuring that any strategy deviating from truthful reporting is no better than honest reporting in either the best or worst case, NOM achieves a weaker but stable incentive constraint in practice.
[0071] Within the above framework, the task allocation mechanism design problem of this invention can be formulated as follows: Under the premise of satisfying task feasibility, system budget constraints, individual rationality, and non-obvious manipulation constraints, for each point of interest... Find a winning user sequence This allows the data requester to extract the maximum value from the data. The specific definition is as follows.
[0072] .
[0073] Unlike existing NOM mechanisms that primarily focus on forward auctions or do not explicitly consider budget constraints, the mechanism designed in this invention is geared towards reverse auction scenarios and explicitly introduces system budget feasibility constraints at the mechanism level. This ensures that while satisfying non-manipulative properties, the overall resource expenditure does not exceed a given resource budget, thus better meeting the resource-constrained requirements of practical applications such as mobile crowdsourcing sensing. Furthermore, considering that the data value contributed by users in mobile crowdsourcing sensing scenarios exhibits significant diminishing marginal returns and depends on the order in which users are selected, this invention further models the system value as a submodular structure with order dependence (order submodularity), thereby more realistically depicting the dynamic characteristics of value accumulation during multi-user collaborative sensing.
[0074] Based on the above analysis, this invention proposes a mobile swarm intelligence sensing task allocation method based on a non-obvious manipulation mechanism. Figure 2 This is a flowchart illustrating a specific implementation of the mobile crowd sensing task allocation method based on a non-obvious manipulation mechanism according to the present invention. Figure 2 As shown, the specific steps of the mobile swarm intelligence sensing task allocation method based on a non-obvious manipulation mechanism of the present invention include:
[0075] S201: Constructing a sequential sub-modulus combinatorial optimization network:
[0076] In this invention, to achieve sequential submodular value modeling, a differentiable sequential submodular combinatorial optimization network is designed. This network is responsible for learning the discrete choices and execution order of users under budget constraints and serves as the basic allocation module in the NOM mechanism. Unlike existing learning mechanism design methods based on continuous allocation, the sequential submodular combinatorial optimization network proposed in this invention can directly generate user choice sequences through stepwise decision-making, thereby explicitly characterizing the diminishing marginal value characteristics of sequence dependence.
[0077] Figure 3 This is a structural diagram of the sequential sub-modulus combination optimization network in this invention. For example... Figure 3 As shown, the sequential sub-module combinatorial optimization network in this invention includes a user scoring module (ScoreNet), a stopping decision unit (StopNet), a relaxation network (SlackNet), a resource cost calculation module, a candidate assignment set generation module, and a global budget pruning module. Each module will be described in detail below:
[0078] The ScoreNet user rating module is based on iteration rounds. User state matrix for each active point of interest Generate user priority score vector And send it to the candidate allocation set generation module, , Indicates the maximum number of iterations. , Indicates the number of users. , Indicates the number of points of interest. Indicates the iteration round A collection of active points of interest . Indicates the iteration round Each user For each point of interest The state characteristics are expressed as follows:
[0079] ,
[0080] in, This represents vector concatenation. Indicates user Statement of interest Resource overhead of the task Indicates user Points of interest The value of the task Indicate points of interest The number of remaining users with demand, Indicate points of interest Number of users selected for the task This indicates that the selected user has selected points of interest. Average resource cost of the task This indicates that the selected user has selected points of interest. Average contribution value of the task This indicates the remaining resource budget.
[0081] Indicates the iteration round Time user For each point of interest The priority score is calculated using a multilayer perceptron. In this embodiment, the user scoring module ScoreNet is implemented using a multilayer perceptron. Unlike the fixed scoring function in traditional greedy methods, the user priority scoring function in this invention depends on the system state, thus possessing stronger expressive power.
[0082] The StopNet decision unit is used to determine the number of iterations. The assigned state vector of each active interest point Generate stopping score And send it to the candidate allocation set generation module. The expression is as follows:
[0083] .
[0084] Stop scores are used to determine points of interest. Whether the task continues to select users can be used to avoid redundant allocation when marginal benefits are low.
[0085] SlackNet, a relaxed network, is used to adjust the user state matrix. Generate resource cost score vector And send it to the resource overhead calculation module.
[0086] The resource cost calculation module is used to calculate costs based on points of interest. Corresponding user-declared resource overhead vector and resource cost score vector Generate resource cost vector And send it to the candidate allocation set generation module.
[0087] The candidate allocation set generation module is used to generate user priority score vectors based on active points of interest. Stop fractions and resource overhead vector Generate iteration rounds Candidate allocation set The data is then sent to the global budget pruning module. The specific method for generating the candidate allocation set is as follows:
[0088] Based on resource overhead vector Get each active point of interest candidate user set :
[0089] .
[0090] According to the above formula, this set can guarantee that: (i) the user is not selected repeatedly; and (ii) the cost of a single candidate resource does not exceed the remaining resource budget.
[0091] To filter out infeasible candidate users, the user priority score vector After correction, the corrected user priority score vector is obtained. :
[0092] .
[0093] The corrected user priority score vector and stopping fractions splicing together to obtain motion features Then, the Gumbel-Softmax function is used to analyze the action features. The process is performed to obtain the action score vector. , , Represented as points of interest Select User The score, This indicates where to stop selecting a user's score. Gumbel-Softmax provides a differentiable discrete selection approximation, allowing the model to be trained end-to-end via gradient descent. During training, continuous relaxation of the output is used to ensure gradient flow; during inference, the selection degenerates into a one-hot form, i.e., for each interest point... In the current round, select at most one user or stop.
[0094] If the score Points of interest Stop selecting users, otherwise select action score. Users and points of interest corresponding to the median maximum value Task composition - user interaction and add to the candidate allocation set. .
[0095] The global budget pruning module is used to prune resources based on the budget. For candidate allocation set Perform trimming to obtain the final allocation set. Since the budget constraint is a global constraint, relying solely on local candidate screening cannot guarantee that the overall resource expenditure will not exceed the budget. Therefore, this invention designs a global budget pruning module, the pruning method of which is as follows:
[0096] First, based on the candidate allocation set Each candidate task - user pair Priority score Sort the candidate task-user pairs from largest to smallest. Then, use a greedy selection strategy to select the highest-scoring candidate task-user pairs and accumulate resource costs until the resource budget constraint can no longer be satisfied, and return the final allocation set. .
[0097] Through this structured budget feasible projection process, the candidate assignment results generated by the neural network can be directly mapped to the feasible domain that satisfies the global budget constraints, thereby ensuring budget feasibility at the mechanism structure level.
[0098] It's important to note that the global budget pruning module primarily serves as a structured budget feasible projection module and does not perform mechanism learning itself. The learning process within the model is mainly accomplished by ScoreNet, SlackNet, and a differentiable action selection process based on Gumbel-Softmax. Meanwhile, because StopNet dynamically decides whether to continue assigning users to corresponding POIs based on the remaining budget, some POIs will prematurely stop generating candidate actions when the budget is tight. This reduces the generation of invalid candidates due to subsequent budget pruning and improves overall training stability.
[0099] The state update module is used to update the allocation set. To perform a state update, the specific method is as follows: for the allocation set Each task - user , assign identifier resource consumption , will users Add to Points of Interest Execution sequence Simultaneously update the allocation set. Each point of interest The number of remaining users with demand Remaining resource budget When there is interest In iteration rounds Stop selecting users or the number of remaining users with demand. Remove it from the set of active points of interest Delete the set of active points of interest When the set of active points of interest or The sequential modular combination optimization network stops operating.
[0100] S202: Training the sequential sub-mode combination optimization network:
[0101] Several mobile crowd sensing task scenario samples were collected to train the sequential sub-module combination optimization network, resulting in a well-trained sequential sub-module combination optimization network.
[0102] The loss function during training in this embodiment , Each point of interest Execution sequence The value of the system. Considering that the system value satisfies the order-dependent property, that is, the marginal contribution of users decreases with the order of selection, this embodiment explicitly introduces an order-dependent discounting mechanism into the computational system, that is, for each execution sequence... , record the first The number of iterations when a user is selected is: Therefore, its contribution value is revised as follows:
[0103] ,
[0104] in, , Execution sequences before and after the correction are respectively. The Middle individual users Points of interest The value of the task , Indicates the execution sequence The length.
[0105] Then, the following formula is used to calculate the value at each point of interest. Execution sequence value :
[0106] .
[0107] This design characterizes the diminishing marginal returns at the model level, enabling the network to learn allocation strategies that conform to the sequential sub-module structure. Backpropagation is performed based on the loss function, simultaneously updating the user rating module, stopping the decision unit, and relaxing the network parameters. The training process employs a Gumbel temperature annealing strategy to gradually tighten the relaxation.
[0108] S203: Obtain user and task information:
[0109] Acquire each point of interest in the current mobile crowd sensing task scenario. Number of users in demand Each user Points of interest The contribution value of the task ,user Statement of interest Task resource overhead And the resource budget for all sensing task information. .
[0110] In this embodiment, to better evaluate the contribution value of users to point-of-interest tasks, contribution value is assessed based on user trajectories, thereby better representing the spatiotemporal matching degree of user trajectories to tasks. For example, if a user's trajectory passes near the task location within a short period after the task is posted, the user's contribution value to the task is high; if the user's trajectory is far from the task location, or the arrival time is significantly delayed, the contribution value is low. Through this conversion, user trajectories of different formats and lengths can be uniformly mapped to comparable numerical features for task allocation and payment price calculation. Therefore, the specific calculation method for contribution value in this embodiment is as follows:
[0111] Get each point of interest The task location of the perception task Task release time Space-sensitive parameters Time-sensitive parameters Space-sensitive parameters This indicates the task's tolerance for spatial deviations; a smaller value indicates a greater emphasis on precise location, while a larger value indicates a wider acceptable spatial coverage. (Time-sensitive parameter) This indicates the task's tolerance for spatiotemporal deviations. Similarly, the smaller the value, the more the task emphasizes precise time, while the larger the value, the wider the acceptable time coverage.
[0112] Obtain user trajectory , Indicates user At any moment Location, , Indicates the length of the user's trajectory. Location This can be latitude and longitude coordinates, map grid numbers, or two-dimensional location coordinates used internally by the platform; time The time can be the actual sampling time, the estimated arrival time, or the time calculated from the navigation route. Trajectory data can come from user terminal location records, user-submitted planned routes, historical travel pattern predictions, or driving paths uploaded by in-vehicle devices. In this embodiment, the user trajectory is not required to cover the entire task area; instead, the contribution value is calculated based on the degree of matching between trajectory points and task location and task release time.
[0113] User location For points of interest Spatial quality of the perception task It can be represented using Gaussian kernel form:
[0114] .
[0115] The user's location is calculated using the following formula. For points of interest Perceive the time quality of tasks :
[0116] .
[0117] Therefore, the present invention uses the following formula to calculate the user For the task The value of contribution :
[0118] ,
[0119] in, Indicates user Quality parameters of the extracted sensory data.
[0120] S204: Determine whether there are users in the current mobile crowd intelligence sensing task scenario who meet the extreme adversary configuration. If there are users who meet the extreme adversary configuration, otherwise proceed to step S205.
[0121] In this embodiment, the extreme adversary configuration includes the best configuration (Golden Ticket, GT) or the worst configuration (Wooden Spoon, WS), and the determination method is as follows:
[0122] Calculate each user Total resource overhead for all tasks :
[0123] .
[0124] Calculate each user Total value :
[0125] .
[0126] Then The users are sorted in descending order of total value to obtain the user sequence. .
[0127] Gold certificate configuration represents the user The most favorable competitive environment. If the user sequence All rankings above users All users participated in the bidding with a total resource cost of 0, while all other users reported a total resource cost that was their resource budget. In other words, if excluding users Total resource overhead combination for non-users satisfy:
[0128] ,
[0129] Then the user Meets the optimal configuration.
[0130] The wooden spoon voucher configuration corresponds to the "worst competitive environment" faced by the user, if excluding the user Total resource overhead combination for other users satisfy:
[0131] ,
[0132] Then the user Satisfy the worst-case scenario.
[0133] S206: Invoke the closed allocation rule to perform allocation:
[0134] Invoke the closed allocation rule to execute the task - user pairing and allocation.
[0135] The specific allocation method of the closed allocation rule in this embodiment is as follows:
[0136] If there are users To meet the best configuration, and to provide users with the best performance Assigned to all points of interest Let the corresponding allocation identifier resource consumption At the same time, the user sequence Chinese users Previous users Assigned to all points of interest Let the corresponding allocation identifier resource consumption And in the user Subsequent users will not be selected, therefore the user sequence will not be entered. Chinese users Subsequent users Assigned to any point of interest Let the corresponding allocation identifier resource consumption This construction allows the target user to obtain the maximum possible benefit within the limits allowed by the mechanism under the GT configuration, thus characterizing the best utility boundary under real-world reporting conditions.
[0137] If there are users If the worst-case configuration is met, then the user will not be... Assigned to any point of interest Make all assigned identifiers resource consumption If the user Then, user 1 will be assigned to all points of interest. Let the corresponding allocation identifier resource consumption If the user Then all other users will be assigned to all points of interest. Let the corresponding allocation identifier resource consumption , and Therefore, at this point, it only applies to the user. The system allocates resources to users other than those mentioned above, thereby constructing the worst-case benefit scenario that users may face, which is used to characterize the worst-case utility frontier under the real report.
[0138] For each point of interest Get its allocation set Then allocate the set Users in the community are categorized by their points of interest. The contribution value of the task Arrange from largest to smallest to obtain points of interest. Execution sequence This explicitly characterizes the diminishing marginal returns under ordered submodule conditions.
[0139] S206: Use the sequential submodulus combination optimization network for allocation:
[0140] The sequential submodulus combination optimization network execution task is user allocation, which includes initializing interest points. The number of remaining users with demand Number of selected users Selected users' points of interest Average resource cost of the task Average contribution value Remaining resource budget .
[0141] Example
[0142] To better illustrate the technical effects of this invention, specific examples are used to experimentally verify it. To evaluate the performance of this invention in different mobile crowdsourcing sensing scenarios, this embodiment normalizes the user's contribution value to points of interest (POIs) calculated based on the user's movement trajectory to the range [0,1], used to characterize the user's overall contribution in the corresponding sensing task. For each POI, the required number of users... The coverage requirements for adaptively generating different Points of Interest (POIs) may vary depending on the current user base, with a maximum of [number missing]. Regarding resource overhead declarations, this embodiment assumes a correlation between the user's declared resource overhead and their contribution capability. For each user-task pair, the user's declared resource overhead... The system budget is generated by a combination of contribution values and independent random terms. This means that high-value users typically have higher resource expenditures, but a degree of randomness is retained to simulate the uncertainty of user cost and benefit expectations in a real-world environment. The budget in the experiment is determined by the user base. When no fixed budget is specified, the system budget is set to... ,in In this experiment, it was set to 0.8. This represents the average bid level. This setting allows the budget to grow linearly with the number of users while maintaining a tight budget constraint across experiments of different sizes.
[0143] To fully verify the effectiveness of this invention (denoted as NOMNet), a comparative baseline is constructed from three levels: heuristic mechanism, traditional realism mechanism, and deep learning realism mechanism.
[0144] Baseline Method 1: NOMGreedy (Traditional Greedy NOM Mechanism)
[0145] This method is built upon the overall framework of this invention, employing a classic greedy strategy as a heuristic baseline during the user selection phase. Specifically, in each iteration, the mechanism prioritizes users with the highest unit cost benefit, i.e., ranking them according to marginal value density, and then sequentially selecting users in descending order until the budget constraint can no longer be satisfied. This method has low computational complexity and can reflect the performance of traditional heuristic mechanisms in budget-constrained mobile crowdsourcing perception scenarios.
[0146] Baseline Method 2: OSPM (Traditional DSIC Mechanism)
[0147] The OSPM algorithm is a validity mechanism based on classical auction theory. It employs a proportional share strategy for user allocation and combines this with a critical payment rule to determine the winning user's payment. Because its allocation process satisfies monotonicity and its payment process follows a critical value pricing rule, it guarantees individual rationality (IR) and dominant strategy incentive compatibility (DSIC). This method is one of the representative budget feasibility validity mechanisms in the field of mobile crowdsourcing perception.
[0148] Baseline Method 3: RegretNet (Deep Learning DSIC Mechanism)
[0149] To further compare with current deep learning mechanism design methods, this paper constructs a neural network-based realism mechanism as a strong baseline. This method references the core idea of RegretNet, utilizing a neural network to simultaneously learn the task allocation function and the payoff function, and achieving realism optimization through regret minimization constraints. Unlike the original RegretNet, which is geared towards standard auction scenarios, this embodiment extends it to budget-constrained mobile crowdsourcing perception environments and adapts it to a POI-perceived value model. This method represents the mainstream research paradigm in current deep learning-driven mechanism design.
[0150] Figure 4 This is a comparison chart of the total value of the present invention and the baseline method in different scenarios in this embodiment. For example... Figure 4As shown, in a scenario with a fixed number of POIs (POIs=5), the total system value of each mechanism generally increases with the number of users, indicating that a larger user pool can provide the platform with more potential high-value users. Among them, NOMNet consistently achieves the highest total system value across all user scales, and its leading advantage further expands with the increase in the number of users. For example, when the user scale increases to 8 and 10, the system value of NOMNet is significantly higher than other mechanisms, indicating that it can more effectively utilize the heterogeneity among users and the marginal contribution information in the sequential submodular value structure, thereby learning a better budget allocation and user selection strategy. In contrast, although the traditional heuristic mechanism NOMGreedy can also continuously improve system revenue with the growth of the user scale, its overall performance is consistently lower than that of NOMNet, indicating that greedy strategies based on local marginal revenue are difficult to achieve globally optimal resource allocation. Furthermore, the total system value of RegretNet fluctuates significantly with changes in the user scale; for example, when the number of users increases from 4 to 8, its total system value actually decreases. This indicates that strict DSIC constraints compress the mechanism learning space under complex, budget-constrained environments, easily leading to limited value optimization capabilities and decreased training stability. In contrast, NOMNet employs the NOM weak incentive stability constraint, restricting only explicit manipulation behavior, thus preserving a larger mechanism optimization space. This allows the model to focus more on improving system efficiency, which is a key reason why it maintains stable performance growth even as the user scale expands. Meanwhile, OSPM's total system value is significantly lower than other mechanisms across all user scales, indicating that traditional rule-based mechanisms have limited resource utilization efficiency in complex dynamic scenarios.
[0151] Under a fixed user scale (Users=5) and increasing POI count, the overall system value of each mechanism also increased, indicating that more tasks can bring higher potential benefits to the platform. NOM-Net consistently maintained the best performance, and its advantage further expanded with the increase in the number of POIs. When the number of POIs reached 10, the overall system value of NOMNet was significantly higher than NOMGreedy, RegretNet, and OSPM, demonstrating that the proposed deep learning mechanism can effectively learn the dynamic competitive relationships between different tasks and budget resource allocation strategies, thereby making fuller use of the limited budget to improve overall system benefits. In contrast, while RegretNet achieved some benefit improvement in medium-sized POI scenarios, its growth slowed down under large-scale task conditions, and even showed performance fluctuations, further illustrating that the strict DSIC constraint limited the model's optimization ability in complex high-dimensional environments. Overall, regardless of the fixed user scale or fixed POI scale, deep learning-based mechanisms significantly outperformed traditional heuristic methods, and the NOM-Net-based mechanism could explore a larger feasible space compared to the strict DSIC mechanism. Experimental results show that combining deep learning mechanism design with NOM weak stimulus stability can effectively improve system efficiency while maintaining mechanism stability, and is more suitable for complex mobile crowd sensing scenarios.
[0152] Figure 5 This is a comparison chart of the total resource overhead of the present invention and the baseline method in different scenarios in this embodiment. For example... Figure 5As shown, in a scenario with a fixed number of users (Users=5), the total resource overhead of each mechanism generally increases with the increase in the number of POIs, indicating that larger task demands require the platform to invest more budget. Specifically, the total resource overhead of NOMNet is consistently slightly higher than NOMGreedy and OSPM, and close to RegretNet. This indicates that NOM-Net can more actively incentivize high-value users to participate in task execution under budget constraints. Combined with the previous results on the total system value, it can be found that NOMNet can obtain higher data value gains through more effective budget allocation, demonstrating a superior value-cost trade-off capability. In a scenario with a fixed number of POIs (POIs=5), the total resource overhead of each mechanism also increases with the increase in the user base. This is because a larger user pool can provide more potential high-value users. Compared to the traditional heuristic mechanism NOMGreedy, NOMNet exhibits higher budget utilization and a more stable growth trend across most user scales, indicating that the deep learning-based sequential allocation mechanism can more effectively mine user heterogeneity and marginal contribution information in the sequential submodulus value structure, thereby achieving a more reasonable resource overhead allocation strategy. Furthermore, OSPM's total resource overhead was significantly lower than other mechanisms across all experimental settings, but its total system value was also lower, indicating that the strict DSIC mechanism struggles to fully utilize budget to improve system efficiency in budget-constrained scenarios. In contrast, the present invention, NOMNet, achieves more efficient budget utilization and task allocation while maintaining budget feasibility and NOM stability.
[0153] Figure 6 This is a comparison chart of the number of selected users in different scenarios for the present invention and the baseline method in this embodiment. For example... Figure 6As shown, in a scenario with a fixed number of POIs (POIs=5), the number of users selected by each mechanism generally increases with the increase in user scale. This is because a larger user pool provides the platform with more candidate users, enabling the mechanism to select more participants to complete the task within the budget. Among them, NOMGreedy consistently selects the most users, and its growth is the most significant with the increase in the number of users. However, combined with the previous results on the total system value, it can be found that its system benefits do not achieve optimal performance in tandem, indicating that traditional greedy mechanisms tend to increase coverage by increasing the number of users, but lack effective modeling of user marginal contributions and global budget efficiency. In contrast, NOM-Net maintains a moderate number of users selected throughout the user scale growth process, with significantly fewer selected users than NOMGreedy, but a higher total system value. This shows that NOM-Net does not simply rely on increasing the number of participating users to improve system benefits, but is able to more effectively identify high-value users based on the sequential submodular value structure and learn more efficient budget allocation and user selection strategies, thereby obtaining higher data value benefits with fewer users. Furthermore, RegretNet's growth in the number of selected users in large-scale scenarios is relatively limited, indicating that deep learning mechanisms based on strict DSIC constraints are easily constrained by the mechanism under budget constraints, making it difficult to fully explore a better allocation space. In scenarios with a fixed number of users (Users=5), the number of selected users for each mechanism also generally increases with the increase in the number of POIs. This is because more task areas require more users to participate in data collection. Among them, NOM-Net's growth in the number of selected users is relatively stable and always remains within a reasonable range, while NOMGreedy still selects significantly more users. Experimental results further show that NOM-Net can achieve higher system value without significantly increasing the number of users, indicating that the proposed structure-aware order allocation mechanism can more effectively utilize limited budgets to mine user value, thereby achieving a better balance between system efficiency and resource utilization.
[0154] Figure 7 This is a comparison chart of the total system value and total resource cost of the present invention and the baseline method under different resource budgets in this embodiment. Figure 7As shown, when the budget is 6 and 8, the system is under relatively strict budget constraints; however, when the budget increases to 12 and 14, the budget constraints gradually loosen, and the platform has more room for resource allocation. The total value results show that the total system value of each mechanism generally increases with the budget. This indicates that a higher budget can support the platform in selecting more high-value users to participate in task execution, thereby improving overall data returns. NOM-Net achieved the highest total system value under most budget settings and showed a more stable growth trend as the budget gradually transitioned from tight to loose constraints. Especially after the budget reached 10, NOM-Net's advantage over NOMGreedy and OSPM further expanded, indicating that the proposed structure-aware order allocation mechanism can more effectively utilize additional budgets to mine high-value users and fully utilize the marginal contribution information in the order submodular value structure. In contrast, NOMGreedy performs relatively similarly to NOM-Net in low-budget scenarios, but its total system value growth gradually leveled off as the budget further increased. This indicates that traditional greedy mechanisms primarily rely on local marginal gains for user selection, making it difficult to effectively improve overall resource allocation efficiency when budgets are relatively ample. RegretNet's total system value gradually approaches that of NOM-Net in high-budget scenarios, but remains slightly lower overall, suggesting that deep learning mechanisms based on strict DSIC constraints are still somewhat limited in terms of budget utilization efficiency. OSPM exhibits the lowest system value under all budget conditions, further illustrating that traditional strict realism mechanisms struggle to balance mechanism stability and system efficiency in budget-constrained mobile crowdsourcing sensing scenarios.
[0155] The total resource overhead results show that the total resource overhead of each mechanism gradually increases with budget. NOM-Net's total resource overhead consistently remains within the budget and is generally lower than RegretNet's, while its total system value is higher. This indicates that NOM-Net does not simply rely on higher resource overhead to increase system returns, but rather achieves higher data value gains at a lower cost through more effective user selection and budget allocation strategies. Particularly in scenarios with more flexible budgets, RegretNet's total resource overhead grows faster, but the corresponding value increase does not significantly exceed that of NOM-Net, further demonstrating NOM-Net's advantage in budget utilization efficiency and value-cost trade-offs. Overall, regardless of whether the budget is tight or loose, NOM-Net consistently achieves better total system value while maintaining a reasonable level of resource overhead, demonstrating that the proposed structure-aware learning-based NOM mechanism can achieve more efficient resource allocation and budget utilization in budget-constrained mobile swarm intelligence sensing environments.
[0156] Figure 8This is a typical allocation result diagram of the present invention under different budgets in this embodiment. For example... Figure 8 As shown, with an increased budget, this invention can select more high-value users to participate in task execution, thereby increasing the overall value of the system. When the budget is low, the system only prioritizes some high-value task-user pairs, such as p3 and p4, which are mainly completed by user u3. With an increased budget, some POIs can be further allocated to additional users; for example, p1, p2, and p5 all have newly added users, thereby increasing the overall revenue. It can also be observed that an increased budget does not mean that existing users are always retained. For example, low-value user u4, selected in p2 when Budget=6, is replaced by higher-value users under a higher budget, indicating that NOM-Net dynamically reconstructs the allocation results based on the current budget, rather than simply accumulating users. Furthermore, high-value users typically maintain a higher execution order, while subsequently added users are added in a lower order, demonstrating a clear sequential submodeling.
[0157] To verify the non-obvious manipulation nature of NOMNet in this invention, this embodiment analyzes the mechanism behavior when users deviate from their true bids in the Golden Ticket, Wooden Spoon, and normal competitive environments under 4 user and 5 POI scenarios. Figure 9 This is a graph showing the changes in resource overhead and utility of the present invention under extreme adversary configurations in this embodiment. For example... Figure 9 As shown, in the Golden Ticket configuration, the user's utility remains constant even when continuously deviating from the actual resource cost. This demonstrates that the false reporting strategy cannot outperform the true report in the sense of "best outcome," experimentally verifying that the present invention satisfies the BNOM condition. In the Wooden Spoon configuration, regardless of how the user adjusts the declared resource cost, both the resource cost and utility remain zero, indicating that the user cannot improve the worst-case outcome by manipulating the bid. Therefore, the present invention satisfies the WNOM condition.
[0158] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.
Claims
1. A mobile swarm intelligence sensing task allocation method based on a non-obvious manipulation mechanism, characterized in that, Includes the following steps: S1: Construct a sequential sub-module combinatorial optimization network, including a user rating module, a stopping decision unit, a relaxation network, a resource cost calculation module, a candidate assignment set generation module, and a global budget pruning module, wherein: The user rating module is used to rate users based on iteration rounds. User state matrix for each active point of interest Generate user priority score vector And send it to the candidate allocation set generation module, , Indicates the maximum number of iterations. , Indicates the number of users. , Indicates the number of points of interest. Indicates the iteration round A collection of active points of interest ; Indicates the iteration round Each user For each point of interest The state characteristics are expressed as follows: , in, This represents vector concatenation. Indicates user Statement of interest Resource overhead of the task Indicates user Points of interest The value of the task Indicate points of interest The number of remaining users with demand, Indicate points of interest Number of users selected for the task This indicates that the selected user has selected points of interest. Average resource cost of the task This indicates that the selected user has selected points of interest. Average contribution value of the task Indicates the remaining resource budget; Indicates user For each point of interest Priority score; The stopping decision unit is used to determine the iteration round. The assigned state vector of each active interest point Generate stopping score And send it to the candidate allocation set generation module; The expression is as follows: ; Relaxed networks are used to adjust the user state matrix. Generate resource cost score vector And send it to the resource overhead calculation module; The resource cost calculation module is used to calculate costs based on points of interest. Corresponding user-declared resource overhead vector and resource cost score vector Generate resource cost vector And send it to the candidate allocation set generation module; The candidate allocation set generation module is used to generate user priority score vectors based on active points of interest. Stop fractions and resource overhead vector Generate iteration rounds Candidate allocation set The data is then sent to the global budget pruning module. The specific method for generating the candidate allocation set is as follows: Based on resource overhead vector Get each active point of interest candidate user set : ; User priority score vector After correction, the corrected user priority score vector is obtained. : ; The corrected user priority score vector and stopping fractions splicing together to obtain motion features Then, the Gumbel-Softmax function is used to analyze the action features. The process is performed to obtain the action score vector. , , Represented as points of interest Select User The score, This indicates that the user's score has been stopped from being selected; If the score Points of interest Stop selecting users, otherwise select action score. Users and points of interest corresponding to the median maximum value Task composition - user interaction and add to the candidate allocation set. ; The global budget pruning module is used to prune resources based on the budget. For candidate allocation set Perform trimming to obtain the final allocation set. The cutting method is as follows: First, based on the candidate allocation set Each candidate task - user pair Priority score Sort the candidate task-user pairs from largest to smallest. Then, use a greedy selection strategy to select the highest-scoring candidate task-user pairs and accumulate resource costs until the resource budget constraint can no longer be satisfied, and return the final allocation set. ; The state update module is used to update the allocation set. To perform a state update, the specific method is as follows: for the allocation set Each task - user , assign identifier resource consumption , will users Add to Points of Interest Execution sequence Simultaneously update the allocation set. Each point of interest Number of remaining users Remaining resource budget When there is interest In iteration rounds Stop selecting users or the number of remaining users with demand. Remove it from the set of active points of interest Delete the set of active points of interest When the set of active points of interest or The sequential sub-modulus combination optimization network stops operating; S2: Collect several mobile crowd sensing task scenario samples to train the sequential sub-mode combination optimization network and obtain the trained sequential sub-mode combination optimization network. S3: Obtain each point of interest in the current mobile crowd sensing task scenario. Number of users in demand Each user Points of interest The contribution value of the task ,user Statement of interest Task resource overhead And the resource budget for all sensing task information. ; S4: Determine if there are any users in the current mobile crowd intelligence sensing task scenario who meet the extreme adversary configuration. If there are users who meet the extreme adversary configuration, proceed to step S5; otherwise, proceed to step S6. S5: Invoke the closed allocation rule to execute the task - user pairing and allocation; S6: Call the sequential submodulus combination optimization network execution task - user allocation, where interest points are initialized. Number of remaining users Number of selected users Selected users' points of interest Average resource cost of the task Average contribution value Remaining resource budget .
2. The mobile crowd sensing task allocation method according to claim 1, characterized in that, The value of the contribution It is calculated based on user trajectory, and the specific method is as follows: Get each point of interest The task location of the perception task Task release time Space-sensitive parameters Time-sensitive parameters ; Obtain user trajectory , Indicates user At any moment Location, , Indicates the length of the user's trajectory; The user's location is calculated using the following formula. For points of interest Spatial quality of the perception task : ; The user's location is calculated using the following formula. For points of interest Perceive the time quality of tasks : ; The user is calculated using the following formula. For points of interest The contribution value of perception tasks : , in, Indicates user Quality parameters of the extracted sensory data.
3. The mobile crowd sensing task allocation method according to claim 1, characterized in that, The loss function during the training of the sequential sub-modulus combination optimization network The calculation formula is as follows: , in, Indicated at each point of interest Execution sequence The value of [item] is calculated using the following formula: , in, Indicates the execution sequence The Middle individual users Points of interest The value of the task , Indicates the execution sequence Length, Indicates the execution sequence The Middle individual users The number of iterations when selected.
4. The mobile crowd sensing task allocation method according to claim 1, characterized in that, The method for determining the configuration of the extreme adversary is as follows: Calculate each user Total resource overhead for all tasks : ; Calculate each user Total value : ; Then The users are sorted in descending order of total value to obtain the user sequence. ; If user sequence Excluding users Total resource overhead combination for non-users satisfy: , Then the user Meets the optimal configuration; If user sequence Excluding users Total resource overhead combination for other users satisfy: , Then the user Satisfy the worst-case scenario.
5. The mobile crowd sensing task allocation method according to claim 4, characterized in that, The specific allocation method of the closed allocation rule is as follows: If there are users To meet the best configuration, and to provide users with the best performance Assigned to all points of interest Let the corresponding allocation identifier resource consumption Simultaneously, the user sequence Chinese users Previous users Assigned to all points of interest Let the corresponding allocation identifier resource consumption Do not include user sequences Chinese users Subsequent users Assigned to any point of interest Let the corresponding allocation identifier resource consumption ; If there are users If the worst-case configuration is met, then the user will not be... Assigned to any point of interest Make all assigned identifiers resource consumption If the user Then, user 1 will be assigned to all points of interest. Let the corresponding allocation identifier resource consumption If the user Then all other users will be assigned to all points of interest. Let the corresponding allocation identifier resource consumption , and ; For each point of interest Get its allocation set Then allocate the set Users in the community are categorized by their points of interest. The contribution value of the task Arrange from largest to smallest to obtain points of interest. Execution sequence .
