Method and device for task allocation based on ccv-mab, equipment and storage medium

Abstract

The application is suitable for the technical field of crowd sensing, and provides a crowd sensing task allocation method and device based on CCV-MAB, equipment and storage medium, the method comprises the following steps: obtaining user characteristics of all users under a current crowd sensing problem and task characteristics of a current task, processing the user characteristics and the task characteristics through a decision model to obtain a selected probability of each user, sampling all users according to the selected probability to obtain a target user group, processing user characteristics and task characteristics of each user in the target user group through an evaluation model to obtain an estimated reward, obtaining a real reward through a crowd sensing task allocation model based on CCV-MAB according to the target user group, adjusting model parameters of the decision model and the evaluation model according to the estimated reward and the real reward, and then performing task allocation in the next round, and thus iteratively updating the model, so that the efficiency of the crowd sensing task allocation and the quality of task completion are improved, and the user experience is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of crowd sensing, and particularly relates to a crowd sensing task allocation method and device based on CCV-MAB, equipment and a storage medium. BACKGROUND

[0002] Crowd sensing computing is a typical application based on swarm intelligence, in which the executors of tasks mainly participate in the tasks through mobile devices. The advantage of this way is its extensive coverage and real-time performance, however, it also brings a series of complexities, especially in the aspect of task allocation. The essence of swarm intelligence is to utilize the wisdom, skills and resources of a large number of people to achieve common goals, however, when this mode is applied to the mobile sensing environment, the allocation and coordination of tasks become particularly important. In fact, finding the best task allocation strategy is one of the core challenges of crowd sensing problems.

[0003] In crowd sensing, the difficulty of task allocation mainly comes from the following aspects:

[0004] (1) Suitable crowd selection: Each hired person needs to spend a certain cost, how to allocate tasks to the most suitable crowd becomes a great challenge. This needs to consider the skills, interests, geographical location and other factors of the participants;

[0005] (2) Cooperation among users: For complex tasks such as open source project development, the cooperation of front-end, back-end and core business is usually needed. If there are too many participants in a certain field, it may affect the overall progress of the project. Therefore, how to balance the participants in different fields and promote cooperation among them is an important problem;

[0006] (3) Dynamic adjustment: As the task proceeds, it may be necessary to dynamically adjust the number of hired users. For example, the task difficulty may exceed the expectation, or some participants may not be able to complete the task on time. Therefore, how to dynamically adjust the task allocation according to the task difficulty and user ability is another key challenge.

[0007] In general, the task allocation problem in crowdsourcing requires consideration of multiple factors and a balance between real-time performance and effectiveness. Several studies have attempted to address this issue. For example, a neural network-based combined multi-armed bandit (Neural-MAB) algorithm is used to solve the task allocation problem in crowdsourcing. This algorithm uses one neural network to predict the reward for each individual user (i.e., a single arm), and another neural network to select a group of users (i.e., a super arm) to complete a task. This algorithm integrates gradient ascent and confidence upper bound strategies for task allocation and updates the neural network parameters and confidence upper bounds based on user feedback. However, this method has a significant drawback: it ignores the relationships between users. In other words, a user's individual reward may be influenced by the behavior of other users. Therefore, we cannot accurately evaluate a user's individual reward without considering other users. For instance, when recommending local restaurants to users, the platform does not want to recommend all users to the same highly-rated restaurant to maximize the diversity of reviews and user experience. This illustrates that the relationships between users are an important factor in the task allocation process. Another algorithm for solving the context-dependent combinatorial volatile multi-armed bandit with adaptive discretization (CCV-MAB) problem is as follows: In each round, the platform needs to select a set of tasks to assign to the user based on the user's contextual information (such as location, device type, etc.) to maximize the overall perceived effect. This algorithm considers two characteristics: first, the availability of tasks may change over time, i.e., volatile arm; second, the total reward of the tasks may not be a simple sum of the rewards of individual tasks, but a submodular function with diminishing returns, i.e., submodular reward. However, sub-modal rewards also have some drawbacks, especially in certain crowd perception problems. On the one hand, sub-modal rewards may not fully reflect the complex relationship between users and tasks because they ignore factors such as users' personalized preferences and the difficulty of tasks. On the other hand, sub-modal rewards may lead to some unreasonable or unfair results because they tend to select tasks that are low in relevance to the selected tasks or have high redundancy, while ignoring tasks that are high in relevance to the selected tasks or have low redundancy. For example, in a crowd perception system, if a user has already completed some tasks on a certain topic, the sub-modal rewards may assign him some tasks that are unrelated to or opposite to this topic, which may reduce his interest and motivation; while another user may only receive tasks that are related to or the same as this topic, which may increase his burden and stress. Summary of the Invention

[0008] The purpose of this invention is to provide a method, apparatus, device, and storage medium for crowd sensing task allocation based on CCV-MAB, aiming to solve the problems of low efficiency and poor task completion quality caused by the inability of existing technologies to provide an effective crowd sensing task allocation method.

[0009] On one hand, the present invention provides a crowd-sensing task allocation method based on CCV-MAB, the method comprising the following steps:

[0010] Obtain the user characteristics of all users and the task characteristics of the current task in the current crowd intelligence perception problem;

[0011] The user characteristics and task characteristics are processed by a pre-created decision model to obtain the selection probability of each user;

[0012] Based on the selection probability, all users are sampled to obtain the target user group;

[0013] The estimated reward is obtained by processing the user characteristics and task characteristics of each user in the target user group using a pre-created evaluation model.

[0014] Based on the target user group, real rewards are obtained through a pre-built CCV-MAB-based crowd intelligence perception task allocation model;

[0015] Based on the estimated reward and the actual reward, the model parameters of the decision model and the evaluation model are adjusted, and the process jumps to the step of obtaining the user characteristics of all users and the task characteristics of the current task under the current crowd intelligence perception problem, in order to carry out the next round of task allocation.

[0016] Preferably, the step of processing the user features and the task features using a pre-created decision model includes:

[0017] The first submodule in the decision model performs feature interaction between the user features and the task features to obtain the feature vector of each user.

[0018] The feature vector is processed by the second submodule in the decision model to obtain the corresponding selection probability.

[0019] Preferably, the first submodule consists of an encoder of a converter model, and the second submodule consists of a multilayer perceptron.

[0020] Preferably, the evaluation model consists of an encoder of a converter model and multiple multilayer perceptrons.

[0021] On the other hand, the present invention provides a crowd sensing task allocation device based on CCV-MAB, the device comprising:

[0022] The feature acquisition unit is used to acquire the user features of all users and the task features of the current task under the current crowd intelligence perception problem.

[0023] The probability acquisition unit is used to process the user features and the task features through a pre-created decision model to obtain the selection probability of each user;

[0024] The user sampling unit is used to sample all the users according to the selection probability to obtain the target user group;

[0025] The feature processing unit is used to process the user features and task features of each user in the target user group through a pre-created evaluation model to obtain an estimated reward;

[0026] The reward-obtaining unit is used to obtain real rewards based on the target user group using a pre-built CCV-MAB-based crowd-aware task allocation model; and

[0027] The parameter adjustment unit is used to adjust the model parameters of the decision model and the evaluation model according to the estimated reward and the actual reward, and to trigger the feature acquisition unit to carry out the next round of task allocation.

[0028] Preferably, the probability acquisition unit includes:

[0029] A feature vector acquisition unit is configured to perform feature interaction between the user features and the task features through the first submodule in the decision model to obtain a feature vector for each user; and

[0030] The probability acquisition subunit is used to process the feature vector through the second submodule in the decision model to obtain the corresponding selection probability.

[0031] Preferably, the first submodule consists of an encoder of a converter model, and the second submodule consists of a multilayer perceptron.

[0032] Preferably, the evaluation model consists of an encoder of a converter model and multiple multilayer perceptrons.

[0033] On the other hand, the present invention also provides a mobile device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps described in the above-described CCV-MAB-based crowd sensing task allocation method.

[0034] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the above-described CCV-MAB-based crowd sensing task allocation method.

[0035] This invention acquires the user characteristics of all users and the task characteristics of the current task in a current crowdsourcing sensing problem. It processes these user and task characteristics using a decision model to obtain the selection probability for each user. Based on the selection probability, it samples all users to obtain a target user group. An evaluation model processes the user and task characteristics of each user within the target user group to obtain an estimated reward. Based on the target user group, it obtains the actual reward using a CCV-MAB-based crowdsourcing sensing task allocation model. Based on the estimated and actual rewards, it adjusts the model parameters of both the decision and evaluation models before proceeding to the next round of task allocation. This iterative process updates the model, thereby improving the efficiency of crowdsourcing sensing task allocation and the quality of task completion, ultimately enhancing the user experience. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the implementation of the CCV-MAB-based crowd sensing task allocation method provided in Embodiment 1 of the present invention.

[0037] Figure 2 This is a schematic diagram of the data flow for model updates of the decision model and evaluation model in the CCV-MAB-based crowd sensing task allocation method provided in Embodiment 1 of the present invention.

[0038] Figure 3 This is a schematic diagram of the structure of the CCV-MAB-based crowd sensing task allocation device provided in Embodiment 2 of the present invention;

[0039] Figure 4 This is a schematic diagram of a preferred structure of the CCV-MAB-based crowd sensing task allocation device provided in Embodiment 2 of the present invention;

[0040] Figure 5 This is a schematic diagram of the structure of the mobile device provided in Embodiment 3 of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] The specific implementation of the present invention will be described in detail below with reference to specific embodiments:

[0043] Example One:

[0044] Figure 1 The implementation flow of the crowd sensing task allocation method based on CCV-MAB provided in Embodiment 1 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0045] In step S101, the user characteristics of all users and the task characteristics of the current task under the current crowd intelligence perception problem are obtained.

[0046] This invention is applicable to mobile devices, such as smartphones, smartwatches, and tablets. In embodiments of this invention, user features and task features refer to the attribute characteristics of the user and the task, respectively. For example, user features for intelligent recommendation problems on shopping platforms could be user age, gender, and historical shopping records, while the corresponding task features could be the characteristics of a product, such as category, price, and sales volume.

[0047] In step S102, user features and task features are processed using a pre-created decision model to obtain the selection probability for each user.

[0048] In this embodiment of the invention, the decision model (Actor model) is a deep learning model used to decide which users to assign the task to in the current task. The decision model receives the user features of all current users and the task features of the current task, processes them accordingly, and then outputs the selection probability of each user. The selection probability represents the probability of selecting that user to perform the current task.

[0049] Preferably, the processing of user characteristics and task characteristics is achieved through the following steps:

[0050] (1) The first sub-module in the decision model performs feature interaction between user features and task features to obtain the feature vector of each user;

[0051] In this embodiment of the invention, the decision model includes a first sub-module, which is used to interactively input the user features and task features of all users. Through the first sub-module, the user features and task features are interactively processed to extract the relationships between users and between users and tasks, thereby obtaining the feature vector of each user.

[0052] More preferably, the first submodule consists of the encoder of the Transformer model, namely the Transformer Encoder module. Transformer is a neural network structure based on the attention mechanism. The self-attention mechanism can capture local and global dependencies in the sequence. In the decision model, local dependencies refer to the relationship between the task and the user, while global dependencies refer to the relationship between users. Thus, the Transformer Encoder module learns the relationship between users and between the task and the user, and more reasonably learns the impact of different user combinations on the quality of task completion.

[0053] (2) The feature vector is processed by the second sub-module in the decision model to obtain the corresponding selection probability.

[0054] In this embodiment of the invention, the decision model further includes a second sub-module, which processes the feature vector of each user output by the Transformer Encoder module to obtain the probability of each user being selected and the corresponding selection probability.

[0055] More preferably, the second submodule consists of a multi-layer perceptron (MLP), thereby improving the processing speed of feature vectors and the accuracy of the probability of each user being selected.

[0056] By processing user features and task features through the above steps (1)-(2), the probability of each user being selected can be obtained directly by inputting the context information and task features of all users into the model, thereby improving the processing efficiency of the model.

[0057] In step S103, all users are sampled according to the selection probability to obtain the target user group.

[0058] In this embodiment of the invention, by sampling the selection probabilities of all users output by the Actormodel, the users corresponding to the sampled selection probabilities constitute a target user group. This target user group is the task allocation scheme used to complete the current task. The target user group consists of one or more users, and the task is completed through the target user group. This allows for dynamic adjustment of the number of users allocated to the task based on the task difficulty and user capabilities, thereby improving the efficiency of task allocation. As an example, when the decision model is not well trained initially, the uncertainty of the output probability of the decision model is relatively large. For example, for 5 users, the probability of each user being selected may be 0.5. However, after the model has undergone multiple rounds of training, for example, for complex tasks, the probability of all users being selected output by the model is close to 1, which means that the model needs multiple users to complete the task to obtain a large reward. For simple tasks, the model may only select the most suitable user to complete the task, that is, the probability of this user being selected output by the model is close to 1, and the probability of other users being selected is close to 0.

[0059] In addition, in this embodiment of the invention, when sampling all users, various sampling methods can be used. For example, the probability sampling method means that if a user's selection probability is 0.7, then he or she has a 70% probability of being selected and a 30% probability of not being selected. Other methods include roulette wheel sampling, etc. The sampling method is not specifically limited here.

[0060] In step S104, the user characteristics and task characteristics of each user in the target user group are processed by a pre-created evaluation model to obtain an estimated reward.

[0061] In this embodiment of the invention, the evaluation model (Critic model) is a deep learning model used to estimate the quality of a group of users' performance on a certain task. The user characteristics of each user in the selected target user group and the task characteristics of the current task are input into the evaluation model. After the evaluation model processes the data, the estimated reward is obtained, which is the estimated quality of task completion.

[0062] Preferably, the evaluation model consists of a Transformer Encoder and multiple Multilayer Perceptrons (MLPs), thereby improving the rationality and accuracy of the estimated reward prediction. Specifically, the Transformer Encoder module in the evaluation model interacts with the task features of each user within the target user group, extracting the relationships between users and between users and tasks within the target user group to obtain the feature vectors of the selected users. The multiple MLP modules process the feature vectors of each selected user output by the Transformer Encoder module to obtain the estimated task completion quality. Furthermore, in the evaluation model, the MLP is a three-layer fully connected network, and the number of MLPs can be dynamically adjusted according to the scale of the specific crowdsourcing problem, without limitation here.

[0063] In step S105, based on the target user group, a real reward is obtained through a pre-built CCV-MAB-based crowd intelligence perception task allocation model.

[0064] In this embodiment of the invention, the specific crowd intelligence perception problem is pre-modeled as a crowd intelligence perception task allocation model based on CCV-MAB. This model is a CCV-MAB without a single reward, referred to as CCV-MAB env. In the CCV-MAB env model, the feedback part of the single reward for each user is removed. That is, after the decision-maker selects a group of users, the framework no longer feeds back the individual reward of each user, but only feeds back the contribution of the user group to the quality of task completion. The contribution to the quality of task completion is also the real reward for the target user group to complete the current task. As an example, for the problem of shopping websites recommending products to users, the CCV-MAB env model is considered as the shopping website. Here, the task corresponds to a shopper on the shopping website, and the user corresponds to various products. The task assignment is to recommend multiple products to a shopper. The shopping website (i.e., the CCV-MAB env model) publishes shopper and product features. First, the decision model selects a series of products that the shopper may buy based on these features. Next, the shopping website actually recommends these selected products to the shopper. Finally, the shopper receives a reward, i.e., a real reward, based on whether the shopper buys (clicks) these products.

[0065] In step S106, the model parameters of the decision-making model and the evaluation model are adjusted based on the estimated reward and the actual reward.

[0066] In this embodiment of the invention, the model parameters of the decision-making model and the evaluation model are adjusted based on the estimated reward and the actual reward to update the decision-making model and the evaluation model. Then, the next round of task allocation is carried out. The model is updated iteratively in this way so that the target user group obtained according to this application can complete a certain task with high quality.

[0067] When adjusting the model parameters of the decision-making model and the evaluation model, specifically, based on the estimated reward and the actual reward, the losses of the evaluation model and the decision-making model are calculated. The decision-making model and the evaluation model are then updated based on these two loss functions. The loss function of the evaluation model is... The goal is to make the estimated reward as close as possible to the actual reward; the loss function of the decision model is... The goal is to maximize the quality of task completion, i.e., the reward. Maximizing the reward is equivalent to minimizing the negative reward, i.e., minimizing the loss function. g(s) represents the evaluation model's estimate of the quality of task completion for the target user group s in the current state (i.e., the estimated reward), r represents the actually observed quality of task completion (i.e., the actual reward), and p... a Let represent the probability that user a is selected (i.e., the probability of being chosen), and s represent the group of users selected (i.e., the target user group).

[0068] Figure 2 The structure of the decision model and evaluation model, as well as the data flow for model updates, are shown.

[0069] In this embodiment of the invention, the user characteristics of all users and the task characteristics of the current task under the current crowd sensing problem are obtained. The user characteristics and task characteristics are processed by a decision model to obtain the selection probability of each user. Based on the selection probability, all users are sampled to obtain the target user group. The user characteristics and task characteristics of each user in the target user group are processed by an evaluation model to obtain the estimated reward. Based on the target user group, the actual reward is obtained by a crowd sensing task allocation model based on CCV-MAB. Based on the estimated reward and the actual reward, the model parameters of the decision model and the evaluation model are adjusted, and then the next round of task allocation is carried out. This process is repeated to update the model, thereby improving the efficiency of crowd sensing task allocation and the quality of task completion, and thus improving the user experience.

[0070] Example Two:

[0071] Figure 3 The structure of the CCV-MAB-based crowd sensing task allocation device provided in Embodiment 2 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, including:

[0072] The feature acquisition unit 31 is used to acquire the user features of all users and the task features of the current task under the current crowd intelligence perception problem.

[0073] In this embodiment of the invention, user features and task features refer to the attribute features of users and tasks, respectively. For example, user features for the intelligent recommendation problem of shopping platforms may be user age, gender, historical shopping records, etc., and the corresponding task features may be the features of a certain product, such as category, price, sales volume, etc.

[0074] The probability acquisition unit 32 is used to process user features and task features through a pre-created decision model to obtain the selection probability of each user.

[0075] In this embodiment of the invention, the decision model (Actormodel) is a deep learning model used to decide which users to assign the task to in the current task. The decision model receives the user features of all current users and the task features of the current task, processes them accordingly, and then outputs the selection probability of each user. The selection probability represents the probability of selecting that user to perform the current task.

[0076] User sampling unit 33 is used to sample all users based on the selection probability to obtain the target user group.

[0077] In this embodiment of the invention, by sampling the selection probabilities of all users output by the Actormodel, the users corresponding to the sampled selection probabilities constitute a target user group. This target user group is the task allocation scheme used to complete the current task. The target user group consists of one or more users, and the task is completed through the target user group. This allows for dynamic adjustment of the number of users allocated to the task based on the task difficulty and user capabilities, thereby improving the efficiency of task allocation. As an example, when the decision model is not well trained initially, the uncertainty of the output probability of the decision model is relatively large. For example, for 5 users, the probability of each user being selected may be 0.5. However, after the model has undergone multiple rounds of training, for example, for complex tasks, the probability of all users being selected output by the model is close to 1, which means that the model needs multiple users to complete the task to obtain a large reward. For simple tasks, the model may only select the most suitable user to complete the task, that is, the probability of this user being selected output by the model is close to 1, and the probability of other users being selected is close to 0.

[0078] In addition, in this embodiment of the invention, when sampling all users, various sampling methods can be used. For example, the probability sampling method means that if a user's selection probability is 0.7, then he or she has a 70% probability of being selected and a 30% probability of not being selected. Other methods include roulette wheel sampling, etc. The sampling method is not specifically limited here.

[0079] The feature processing unit 34 is used to process the user features and task features of each user in the target user group through a pre-created evaluation model to obtain an estimated reward.

[0080] In this embodiment of the invention, the evaluation model (Critic model) is a deep learning model used to estimate the quality of a group of users' performance on a certain task. The user characteristics of each user in the selected target user group and the task characteristics of the current task are input into the evaluation model. After the evaluation model processes the data, the estimated reward is obtained, which is the estimated quality of task completion.

[0081] Preferably, the evaluation model consists of a Transformer Encoder and multiple Multilayer Perceptrons (MLPs), thereby improving the rationality and accuracy of the estimated reward prediction. Specifically, the Transformer Encoder module in the evaluation model interacts with the task features of each user within the target user group, extracting the relationships between users and between users and tasks within the target user group to obtain the feature vectors of the selected users. The multiple MLP modules process the feature vectors of each selected user output by the Transformer Encoder module to obtain the estimated task completion quality. Furthermore, in the evaluation model, the MLP is a three-layer fully connected network, and the number of MLPs can be dynamically adjusted according to the scale of the specific crowdsourcing problem, without limitation here.

[0082] The reward-obtaining unit 35 is used to obtain real rewards based on the target user group through a pre-built CCV-MAB-based crowd intelligence perception task allocation model.

[0083] In this embodiment of the invention, the specific crowd intelligence perception problem is pre-modeled as a crowd intelligence perception task allocation model based on CCV-MAB. This model is a CCV-MAB without a single reward, referred to as CCV-MAB env. In the CCV-MAB env model, the feedback part of the single reward for each user is removed. That is, after the decision-maker selects a group of users, the framework no longer feeds back the individual reward of each user, but only feeds back the contribution of the user group to the quality of task completion. The contribution to the quality of task completion is also the real reward for the target user group to complete the current task. As an example, for the problem of shopping websites recommending products to users, the CCV-MAB env model is considered as the shopping website. Here, the task corresponds to a shopper on the shopping website, and the user corresponds to various products. The task assignment is to recommend multiple products to a shopper. The shopping website (i.e., the CCV-MAB env model) publishes shopper and product features. First, the decision model selects a series of products that the shopper may buy based on these features. Next, the shopping website actually recommends these selected products to the shopper. Finally, the shopper receives a reward, i.e., a real reward, based on whether the shopper buys (clicks) these products.

[0084] The parameter adjustment unit 36 ​​is used to adjust the model parameters of the decision-making model and the evaluation model based on the estimated reward and the actual reward.

[0085] In this embodiment of the invention, the model parameters of the decision-making model and the evaluation model are adjusted based on the estimated reward and the actual reward to update the decision-making model and the evaluation model. Then, the next round of task allocation is carried out. The model is updated iteratively in this way so that the selected target user group can complete a certain task with high quality.

[0086] When adjusting the model parameters of the decision-making model and the evaluation model, specifically, based on the estimated reward and the actual reward, the losses of the evaluation model and the decision-making model are calculated. The decision-making model and the evaluation model are then updated based on these two loss functions. The loss function of the evaluation model is... The goal is to make the estimated reward as close as possible to the actual reward; the loss function of the decision model is... The goal is to maximize the quality of task completion, i.e., the reward. Maximizing the reward is equivalent to minimizing the negative reward, i.e., minimizing the loss function. g(s) represents the evaluation model's estimate of the quality of task completion for the target user group s in the current state (i.e., the estimated reward), r represents the actually observed quality of task completion (i.e., the actual reward), and p... a Let represent the probability that user a is selected (i.e., the probability of being chosen), and s represent the group of users selected (i.e., the target user group).

[0087] like Figure 4As shown, preferably, the probability acquisition unit 32 includes:

[0088] The feature vector acquisition unit 321 is used to perform feature interaction between user features and task features through the first sub-module in the decision model to obtain the feature vector of each user.

[0089] In this embodiment of the invention, the decision model includes a first sub-module, which is used to interactively input the user features and task features of all users. Through the first sub-module, the user features and task features are interactively processed to extract the relationships between users and between users and tasks, thereby obtaining the feature vector of each user.

[0090] More preferably, the first submodule consists of the encoder of the Transformer model, namely the Transformer Encoder module. Transformer is a neural network structure based on the attention mechanism. The self-attention mechanism can capture local and global dependencies in the sequence. In the decision model, local dependencies refer to the relationship between the task and the user, while global dependencies refer to the relationship between users. Thus, the Transformer Encoder module learns the relationship between users and between the task and the user, and more reasonably learns the impact of different user combinations on the quality of task completion.

[0091] The probability acquisition subunit 322 is used to process the feature vector through the second submodule in the decision model to obtain the corresponding selection probability.

[0092] In this embodiment of the invention, the decision model further includes a second sub-module, which processes the feature vector of each user output by the Transformer Encoder module to obtain the probability of each user being selected and the corresponding selection probability.

[0093] More preferably, the second submodule consists of a multi-layer perceptron (MLP), thereby improving the processing speed of feature vectors and the accuracy of the probability of each user being selected.

[0094] In this embodiment of the invention, each unit of the CCV-MAB-based crowd sensing task allocation device can be implemented by corresponding hardware or software units. Each unit can be an independent hardware or software unit, or it can be integrated into a single hardware or software unit. This is not intended to limit the invention.

[0095] Example Three:

[0096] Figure 5The structure of the mobile device provided in Embodiment 3 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown.

[0097] The mobile device 5 of this embodiment includes a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and executable on the processor 50. When the processor 50 executes the computer program 52, it implements the steps described in the embodiment of the CCV-MAB-based crowd sensing task allocation method, for example... Figure 1 The steps S101 to S106 are shown. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each unit in the above-described device embodiments, for example... Figure 3 The functions of units 31 to 36 are shown.

[0098] In this embodiment of the invention, the user characteristics of all users and the task characteristics of the current task under the current crowd sensing problem are obtained. The user characteristics and task characteristics are processed by a decision model to obtain the selection probability of each user. Based on the selection probability, all users are sampled to obtain the target user group. The user characteristics and task characteristics of each user in the target user group are processed by an evaluation model to obtain the estimated reward. Based on the target user group, the actual reward is obtained by a crowd sensing task allocation model based on CCV-MAB. Based on the estimated reward and the actual reward, the model parameters of the decision model and the evaluation model are adjusted, and then the next round of task allocation is carried out. This process is repeated to update the model, thereby improving the efficiency of crowd sensing task allocation and the quality of task completion, and thus improving the user experience.

[0099] The mobile device in this embodiment of the invention can be a smartphone. The steps implemented by the processor 50 in the mobile device 5 when executing the computer program 52 to implement the CCV-MAB-based crowd sensing task allocation method can be referred to the description of the foregoing method embodiments, and will not be repeated here.

[0100] Example Four:

[0101] In this embodiment of the invention, a computer-readable storage medium is provided, which stores a computer program. When executed by a processor, the computer program implements the steps described in the embodiment of the CCV-MAB-based crowd sensing task allocation method. For example... Figure 1 The steps S101 to S106 are shown. Alternatively, when the computer program is executed by the processor, it implements the functions of each unit in the above-described device embodiments, for example... Figure 3 The functions of units 31 to 36 are shown.

[0102] In this embodiment of the invention, the user characteristics of all users and the task characteristics of the current task under the current crowd sensing problem are obtained. The user characteristics and task characteristics are processed by a decision model to obtain the selection probability of each user. Based on the selection probability, all users are sampled to obtain the target user group. The user characteristics and task characteristics of each user in the target user group are processed by an evaluation model to obtain the estimated reward. Based on the target user group, the actual reward is obtained by a crowd sensing task allocation model based on CCV-MAB. Based on the estimated reward and the actual reward, the model parameters of the decision model and the evaluation model are adjusted, and then the next round of task allocation is carried out. This process is repeated to update the model, thereby improving the efficiency of crowd sensing task allocation and the quality of task completion, and thus improving the user experience.

[0103] The computer-readable storage medium in embodiments of the present invention may include any entity or device capable of carrying computer program code, a recording medium, such as ROM / RAM, disk, optical disk, flash memory, etc.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A crowd-sensing task allocation method based on CCV-MAB, characterized in that, The method includes the following steps: Obtain the user characteristics of all users and the task characteristics of the current task in the current crowd intelligence perception problem; The user characteristics and task characteristics are processed by a pre-created decision model to obtain the selection probability of each user; Based on the selection probability, all users are sampled to obtain the target user group; The user characteristics and task characteristics of each user in the target user group are processed by a pre-created evaluation model to obtain an estimated reward. The evaluation model consists of an encoder of a converter model and multiple multilayer perceptrons. Based on the target user group, real rewards are obtained through a pre-built CCV-MAB-based crowd intelligence perception task allocation model; Based on the estimated reward and the actual reward, the model parameters of the decision model and the evaluation model are adjusted, and the process jumps to the step of obtaining the user characteristics of all users and the task characteristics of the current task under the current crowd intelligence perception problem, so as to carry out the next round of task allocation; The step of processing the user features and task features using a pre-created decision model to obtain the selection probability for each user includes: The first submodule in the decision model performs feature interaction between the user features and the task features to obtain the feature vector of each user. The feature vector is processed by the second submodule in the decision model to obtain the corresponding selection probability; The step of processing the user characteristics and task characteristics of each user in the target user group using a pre-created evaluation model to obtain an estimated reward includes: The encoder of the converter model interacts with the features of each user and task within the target user group to extract the relationships between users and between users and tasks within the target user group, thereby obtaining the feature vector of the selected user. The multiple multilayer perceptrons process the feature vector of each selected user to obtain the estimated task completion quality. The step of adjusting the model parameters of the decision model and the evaluation model based on the estimated reward and the actual reward includes: Based on the estimated reward and the actual reward, calculate the loss functions of the evaluation model and the decision model, and update the decision model and evaluation model according to the calculated losses. The loss function of the evaluation model is... The loss function of the decision model is g(s) represents the evaluation model's estimate of the task completion quality of the target user group s under the current state, and r represents the actual observed task completion quality. This represents the probability that user a is selected.

2. The method as described in claim 1, characterized in that, The first submodule consists of an encoder of a converter model, and the second submodule consists of a multilayer perceptron.

3. A crowd-sensing task allocation device based on CCV-MAB, characterized in that, The device includes: The feature acquisition unit is used to acquire the user features of all users and the task features of the current task under the current crowd intelligence perception problem. The probability acquisition unit is used to process the user features and the task features through a pre-created decision model to obtain the selection probability of each user; The user sampling unit is used to sample all the users according to the selection probability to obtain the target user group; The feature processing unit is used to process the user features and task features of each user in the target user group through a pre-created evaluation model to obtain an estimated reward. The evaluation model consists of an encoder of a converter model and multiple multilayer perceptrons. The reward-obtaining unit is used to obtain real rewards based on the target user group using a pre-built CCV-MAB-based crowd-aware task allocation model; and The parameter adjustment unit is used to adjust the model parameters of the decision model and the evaluation model according to the estimated reward and the actual reward, and to trigger the feature acquisition unit to carry out the next round of task allocation; When the probability acquisition unit processes the user features and the task features through a pre-created decision model to obtain the selection probability of each user, it includes: The first submodule in the decision model performs feature interaction between the user features and the task features to obtain the feature vector of each user. The feature vector is processed by the second submodule in the decision model to obtain the corresponding selection probability; When the feature processing unit processes the user features and task features of each user in the target user group using a pre-created evaluation model to obtain an estimated reward, it includes: The encoder of the converter model interacts with the features of each user and task within the target user group to extract the relationships between users and between users and tasks within the target user group, thereby obtaining the feature vector of the selected user. The multiple multilayer perceptrons process the feature vector of each selected user to obtain the estimated task completion quality. When the parameter adjustment unit adjusts the model parameters of the decision model and the evaluation model based on the estimated reward and the actual reward, it includes: Based on the estimated reward and the actual reward, calculate the loss functions of the evaluation model and the decision model, and update the decision model and evaluation model according to the calculated losses. The loss function of the evaluation model is... The loss function of the decision model is g(s) represents the evaluation model's estimate of the task completion quality of the target user group s under the current state, and r represents the actual observed task completion quality. This represents the probability that user a is selected.

4. The apparatus as described in claim 3, characterized in that, The first submodule consists of an encoder of a converter model, and the second submodule consists of a multilayer perceptron.

5. A mobile device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 2.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 2.

Images