Multi-target mobile crowd sensing incentive mechanism method based on reverse auction

Abstract

The invention discloses a reverse auction-based multi-target mobile crowd sensing incentive mechanism method, which comprises the following steps of: performing mathematical modeling on a multi-target crowd sensing system, and establishing a social welfare mathematical model containing platform utility and user utility on the basis of reverse auction; according to constraint conditions with limited budget, a social welfare maximization optimization target is constructed; obtaining the geographic position of the user in each time slot and the quality of submitted data; user features are extracted and clustered; and based on deep reinforcement learning, designing a dynamic incentive mechanism algorithm, selecting a proper user to participate in a task, and paying corresponding rewards. According to the method, under the limited budget condition, the task area coverage rate, the platform utility and the user utility are optimized at the same time, and social welfare maximization is achieved; and the users can be dynamically selected and proper remuneration can be provided according to the change of the user moving track and the data quality along with time, so that the adaptability and the overall efficiency of the crowd sensing system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer data processing, and more specifically to a multi-target mobile crowd-sensing incentive mechanism method based on reverse auction. Background Art

[0002] In recent years, with the rapid development of Internet of Things technology, various mobile devices have been embedded with a variety of sensors such as microphones, GPS, cameras and gyroscopes. People can use these devices to perceive the surrounding environment data in their daily lives. Crowd sensing has emerged as an emerging data perception paradigm. Crowd sensing mainly outsources the perception task to a large number of unspecified individual users, making full use of their inherent mobility and the multiple sensing capabilities of mobile devices, so as to complete large-scale data collection at a lower cost. Due to its advantages such as low cost and high flexibility, crowd sensing has been widely used in many fields such as environmental monitoring, traffic data collection and health care.

[0003] Although crowd sensing technology can solve many problems, it is still limited by insufficient user participation and unstable data quality. Therefore, some studies have proposed some incentive models. For example, some people have proposed a user-centric incentive model, which aims to find the optimal user set to maximize the platform utility while compensating for user costs; others consider the correlation between tasks and locations, and choose the user who pays the least as the winner in order to minimize platform costs. These practices mainly emphasize the importance of a single goal and ignore the optimization value of other goals.

[0004] In addition, incentive mechanisms can be divided into offline and online categories according to the degree of prior information the platform has. Offline mechanisms usually assume that the platform has all the information about the user, while online mechanisms only rely on the user's past information and no future information. In location-based services, traditional incentive strategies usually predict the user's movement pattern based on a large number of user movement trajectories, and then use greedy algorithms to select appropriate users to perform tasks. However, this is often not true in actual scenarios. Newly arrived tasks cannot accurately predict the user's movement pattern due to insufficient collected trajectory data. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-objective mobile crowd sensing incentive mechanism method based on reverse auction to solve the shortcomings of existing crowd sensing that mainly emphasizes the importance of a single goal while ignoring other optimization goals, and does not consider the dynamic changes and uncertainty of users in the incentive mechanism.

[0006] The present invention adopts the following technical solution:

[0007] The multi-target mobile crowd intelligence sensing incentive mechanism method based on reverse auction includes the following steps:

[0008] Step 1: Mathematically model the multi-target crowd intelligence perception system and establish a social welfare mathematical model including platform utility and user utility based on reverse auction;

[0009] Step 2: Based on the constraints of limited budget, construct the optimization goal of maximizing social welfare;

[0010] Step 3: Obtain the user’s geographic location and submitted data quality at each time slot;

[0011] Step 4: Extract user features and cluster them;

[0012] Step 5: Based on deep reinforcement learning, design a dynamic incentive mechanism algorithm, select appropriate users to participate in the task and pay corresponding rewards.

[0013] In a preferred embodiment, the multi-target crowd-sensing system in step 1 above consists of a platform and N users. The buyer's budget is B, and he needs to collect noise data in a specified area for a period of time. After receiving the task, the platform first divides the entire sensing area into a set of grids, Z represents the grid set, and each grid j∈Z has a weight w j , indicating the importance of the grid; the entire perception process T is divided into several stages, and the user submits the current location and the quotation b for the task to be performed to the platform at the beginning of each perception stage i ; The platform selects appropriate users to join the winning set W based on the given information, and pays the winning users corresponding rewards p after they perform the task i , use c i It represents the real cost of user i to perform the task. This value is private information known only to the user himself. User utility is defined as the difference between the payment and its real cost. The utility calculation formula of user i is as follows:

[0014]

[0015] Define the user access matrix V t (W) represents whether the winning user has visited a specific grid before the current time slot t, where each row represents a winning user and each column indicates whether the user has visited the grid: v ij ∈{0, 1}, that is, if grid j has been visited by user i, then v ij =1, otherwise v ij = 0; until the current time t, even if user i visits grid j multiple times, v ijStill counted as 1; the platform utility is defined as the platform's revenue minus the reward paid to the winning user. The platform's revenue consists of area coverage revenue and data quality revenue. The area coverage ratio is expressed as the ratio of the number of grids visited by the winner to the number of all grid areas within the trajectory time range: The coverage benefit that the platform obtains from mission area coverage is defined as: The system parameters used to control the gradient of regional coverage benefits use a logarithmic function to describe the law of diminishing marginal returns of the platform benefits; user reputation is defined as the average quality of historical data submitted by users, and the user reputation matrix is: in, represents the average quality of data submitted by user i before the current time t; the data quality benefit obtained by the platform from the data submitted by users is: is the quality of data submitted by user i at time t, is the system parameter that controls the data quality benefit gradient; the platform utility is expressed as:

[0016] The above social welfare is the sum of platform utility and user utility: Use the quote b to perform the task i Instead of the real cost which cannot be accurately obtained i , so the mathematical model of social welfare expression is:

[0017] The formula for constructing the social welfare maximization optimization objective in step 2 above is as follows:

[0018]

[0019]

[0020] The above step 4 uses the adaptive resonance method ARM to extract and cluster user features. The adaptive resonance method ARM integrates the ART network and pruning mechanism. The ART network consists of an input layer, an output layer, and bottom-up and top-down fully connected links; the input layer consists of a user access matrix and a user reputation matrix, and each node of the output layer represents a classified CL, and each participant is assigned to a corresponding node; the input layer and the output layer have two fully directional connections: the bottom-up network weight is represented by W b Indicates that the top-down network weights are represented by W t express.

[0021] In a preferred embodiment, the specific process of step 4 is as follows:

[0022] Step 4.1: Concatenate the user access matrix and the user reputation matrix into a feature matrix C with dimensions |W|×(|Z|+1) t (W);

[0023] Step 4.2: Convert the feature matrix C t (W) Each row is used as the user's feature input into the adaptive resonance network, and the matching degree between each node and the input vector is calculated: net[j, p] represents the matching degree between the jth CL and user p;

[0024] Step 4.3: Select the node j with the highest matching degree * ;

[0025] Step 4.4: Calculate node j * Similarity with user characteristics:

[0026] Step 4.5: Compare the similarity with the set threshold. If it is greater than the threshold, update the network weights; otherwise, traverse all other output nodes to find nodes whose similarity is greater than the threshold; if all nodes in the output layer do not meet the requirements, add a new classification node and assign the user to the category; if the total number of output layer nodes exceeds n, start the pruning mechanism.

[0027] In a preferred embodiment, the dynamic incentive mechanism algorithm of step 5 above includes:

[0028] Step 5.1: Initialize the network Q with random network parameters ω ω (s, a);

[0029] Step 5.2: Copy the same parameters ω - ←ω Initialize the target network

[0030] Step 5.3: Initialize the experience replay pool;

[0031] Step 5.4: Generate experience data and store it in the replay pool;

[0032] Step 5.5: Randomly sample from the replay pool for training.

[0033] Define the following parameters: Get n user state clusters CL from step 4, state: represents the i-th cluster; action: a t Select a state cluster from the user state clusters CL; reward: r t (s t , a t ) is in state s t Take action a tThe social welfare increment obtained minus the penalty for violating the limited budget constraint is calculated as follows: in, is an adjustable parameter; Reward: The reward a user receives at time t is the bid b corresponding to the user selected at time t-1 i ; The above step 5.4 specifically includes the following:

[0034] Step 5.4.1: Use ARM to get the current environment status t ;

[0035] Step 5.4.2: Based on the current network Q ω (s, a) selects action a using the ε-greedy strategy t

[0036]

[0037] Step 5.4.3: Use the ε-greedy strategy to select the user with the largest current reward in the cluster;

[0038] Step 5.4.4: Execute action a t , the environment state becomes s t+1 , get reward r t ;

[0039] Step 5.4.5: t , a t , r t ,s t+1 ) is stored in the replay pool.

[0040] Preferably, the above step 5.5 specifically includes the following:

[0041] Step 5.5.1: Randomly sample data from the playback pool {(s k , a k , r k ,s k )} k＝1，...，K ;

[0042] Step 5.5.2: For each sampled data, use the target network to calculate

[0043] Step 5.5.3: Update the current network Q with the minimum loss target of the loss function ω , the loss function is:

[0044] Step 5.5.4: Update the target network

[0045] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following advantages:

[0046] 1. Under limited budget conditions, the present invention optimizes the task area coverage, platform utility and user utility at the same time to maximize social welfare; and can dynamically select users and provide appropriate rewards based on the changes in user movement trajectories and data quality over time; before using deep reinforcement learning technology, cluster analysis is performed on user location characteristics and data quality characteristics, thereby improving the learning efficiency of the deep reinforcement learning algorithm.

[0047] 2. The present invention extracts features from the state of reinforcement learning and dynamically clusters users based on certain similarities in movement patterns and reputations between users, which can reduce the search space of the algorithm, improve learning efficiency and reduce computing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0049] The specific implementation of the present invention is described below with reference to the accompanying drawings. In order to fully understand the present invention, many details are described below, but for those skilled in the art, the present invention can be implemented without these details. For well-known components, methods and processes, they are not described in detail below.

[0050] Reference Figure 1 This embodiment provides a method for a multi-target mobile crowd intelligence sensing incentive mechanism based on reverse auction, including the following steps:

[0051] Step 1: Mathematically model the multi-target crowd intelligence perception system and establish a social welfare mathematical model including platform utility and user utility based on reverse auction;

[0052] Step 2: Based on the constraints of limited budget, construct the optimization goal of maximizing social welfare;

[0053] Step 3: Obtain the user’s geographic location and submitted data quality at each time slot;

[0054] Step 4: Extract user features and cluster them;

[0055] Step 5: Based on deep reinforcement learning, design a dynamic incentive mechanism algorithm, select appropriate users to participate in the task and pay corresponding rewards.

[0056] The multi-target crowd intelligence perception system of the present invention consists of a platform and N users. Assuming that the buyer's budget is B, he needs to collect noise data in a specified area for a period of time; after receiving the task, the platform first divides the entire perception area into a set of grids, Z represents the grid set, and each grid j∈Z has a weight wj, which represents the importance of the grid; the entire perception process T is divided into several stages.

[0057] When building a mathematical model for the perception data collection process based on reverse auction, the requester is the buyer, the user is the seller, and the platform is the auctioneer. At the beginning of each perception phase, the user submits the current location and the bid for the task to be performed to the platform. i ; The platform selects appropriate users to join the winning set W based on the given information, and pays the winning users corresponding rewards p after they perform the task i , use c i It represents the real cost of user i to perform the task. This value is private information known only to the user himself.

[0058] User utility is defined as the difference between the payment and its true cost. The utility calculation formula for user i is as follows:

[0059]

[0060] The winning user’s geographic location and data quality submitted in each time slot:

[0061] Define the user access matrix V t (W) represents whether the winning user has visited a specific grid before the current time slot t, where each row represents a winning user and each column indicates whether the user has visited the grid: v ij ∈{0, 1}, that is, if grid j has been visited by user i, then v ij =1, otherwise v ij = 0; until the current time t, even if user i visits grid j multiple times, v ij Still counted as 1; the platform utility is defined as the platform's revenue minus the reward paid to the winning user. The platform's revenue consists of area coverage revenue and data quality revenue. The area coverage ratio is expressed as the ratio of the number of grids visited by the winner to the number of all grid areas within the trajectory time range: The coverage benefit that the platform obtains from mission area coverage is defined as: System parameters used to control the gradient of regional coverage benefits, using logarithmic functions to describe the law of diminishing marginal returns of platform benefits;

[0062] Define user reputation as the average quality of historical data submitted by the user, user reputation matrix: in, represents the average quality of data submitted by user i before the current time t; the data quality benefit obtained by the platform from the data submitted by users is: is the quality of data submitted by user i at time t, is the system parameter that controls the data quality benefit gradient; the platform utility is expressed as:

[0063] Social welfare is the sum of platform utility and user utility: Since the platform cannot accurately obtain the user's true cost c i , so use the quotation b of the execution task i Replace the real cost c i , the mathematical model of social welfare expression is changed to:

[0064]

[0065] The platform needs to maximize social welfare under the constraint of a limited budget. The specific formula is as follows:

[0066]

[0067]

[0068] The above step 4 uses the adaptive resonance method ARM to extract and cluster user features. As an unsupervised learning technology, the adaptive resonance method ARM integrates the ART network and pruning mechanism. Among them, the ART network consists of an input layer, an output layer, and fully connected links from bottom to top and from top to bottom; the input layer consists of a user access matrix and a user reputation matrix, and each node of the output layer represents a classified CL, and each participant is assigned to a corresponding node; the input layer and the output layer have two fully directional connections: the bottom-up network weight is represented by W b Indicates that the top-down network weights are represented by W t express.

[0069] The specific process of adaptive resonance method ARM is as follows:

[0070] Step 4.1: Concatenate the user access matrix and the user reputation matrix into a feature matrix C with dimensions |W|×(|Z|+1) t (W);

[0071] Step 4.2: Convert the feature matrix C t (W) Each row is used as the user's feature input into the adaptive resonance network, and the matching degree between each node and the input vector is calculated: represents the matching degree between the jth CL and user p;

[0072] Step 4.3: Select the node j with the highest matching degree * ;

[0073] Step 4.4: Calculate node j * Similarity with user characteristics:

[0074] Step 4.5: Compare the similarity with the set threshold. If it is greater than the threshold, update the network weights; otherwise, traverse all other output nodes to find nodes whose similarity is greater than the threshold; if all nodes in the output layer do not meet the requirements, add a new classification node and assign the user to the category; if the total number of output layer nodes exceeds n, start the pruning mechanism.

[0075] Define the following parameters: Get n user state clusters CL from step 4, state: represents the i-th cluster; action: a t Select a state cluster from the user state clusters CL; reward: r t (s t , a t ) is in state s t Take action a t The social welfare increment obtained minus the penalty for violating the limited budget constraint is calculated as follows: in, is an adjustable parameter; Reward: The reward a user receives at time t is the bid b corresponding to the user selected at time t-1 i .

[0076] The dynamic incentive mechanism algorithm of step 5 above includes:

[0077] Step 5.1: Initialize the network Q with random network parameters ω ω (s, a);

[0078] Step 5.2: Copy the same parameters ω - ←ω Initialize the target network

[0079] Step 5.3: Initialize the experience replay pool;

[0080] Step 5.4: Generate experience data and store it in the replay pool;

[0081] Step 5.5: Randomly sample from the replay pool for training.

[0082] Among them, step 5.4 specifically includes the following:

[0083] Step 5.4.1: Use ARM to get the current environment statust ;

[0084] Step 5.4.2: Based on the current network Q ω (s, a) selects action a using the ε-greedy strategy t :

[0085]

[0086] Step 5.4.3: Use the ε-greedy strategy to select the user with the largest current reward in the cluster;

[0087] Step 5.4.4: Execute action a t , the environment state becomes s t+1 , get reward r t ;

[0088] Step 5.4.5: t , a t , r t ,s t+1 ) is stored in the replay pool.

[0089] Step 5.5 specifically includes the following:

[0090] Step 5.5.1: Randomly sample data from the playback pool {(s k , a k , r k ,s k )} k＝1，...，K ;

[0091] Step 5.5.2: For each sampled data, use the target network to calculate

[0092] Step 5.5.3: Update the current network Q with the minimum loss target of the loss function ω , the loss function is:

[0093]

[0094] Step 5.5.4: Update the target network

[0095] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A multi-target mobile crowd-sensing incentive mechanism method based on reverse auction, characterized in that: The following steps are involved: Step 1: Mathematically model the multi-target crowd intelligence perception system and establish a social welfare mathematical model including platform utility and user utility based on reverse auction; Step 2: Based on the constraints of limited budget, construct the optimization goal of maximizing social welfare; Step 3: Obtain the user’s geographic location and submitted data quality at each time slot; Step 4: Extract user features and cluster them; Step 5: Based on deep reinforcement learning, design a dynamic incentive mechanism algorithm, select appropriate users to participate in the task and pay corresponding rewards.

2. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction according to claim 1, characterized in that: The multi-target crowd intelligence perception system in step 1 is composed of a platform and N users. The buyer's budget is B, and he needs to collect noise data in a specified area for a period of time; After receiving the task, the platform first divides the entire perception area into a set of grids, Z represents the grid set, and each grid j∈Z has a weight w j , indicating the importance of the grid; The entire perception process T is divided into several stages. At the beginning of each perception stage, the user submits the current location and the quotation b for the task to be performed to the platform. i ; The platform selects appropriate users to join the winning set W based on the given information, and pays the winning users corresponding rewards p after they perform the task i , use c i It represents the real cost of user i to perform the task. This value is private information known only to the user himself. User utility is defined as the difference between the payment and its real cost. The utility calculation formula of user i is as follows:

3. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 2, characterized in that: Define the user access matrix V t (W) represents whether the winning user has visited a specific grid before the current time slot t, where each row represents a winning user and each column indicates whether the user has visited the grid: v ij ∈{0, 1}, that is, if grid j has been visited by user i, then v ij =1, otherwise v ij = 0; until the current time t, even if user i visits grid j multiple times, v ij Still counted as 1; the platform utility is defined as the platform's revenue minus the reward paid to the winning user. The platform's revenue consists of area coverage revenue and data quality revenue. The area coverage ratio is expressed as the ratio of the number of grids visited by the winner to the number of all grid areas within the trajectory time range: The coverage benefit that the platform obtains from mission area coverage is defined as: ), The system parameters used to control the gradient of regional coverage benefits use a logarithmic function to describe the law of diminishing marginal returns of the platform benefits; user reputation is defined as the average quality of historical data submitted by users, and the user reputation matrix is: in, represents the average quality of data submitted by user i before the current time t; the data quality benefit obtained by the platform from the data submitted by users is: is the quality of data submitted by user i at time, is the system parameter that controls the data quality benefit gradient; the platform utility is expressed as:

4. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 3 is characterized in that: Social welfare is the sum of platform utility and user utility: Use the quote b to perform the task i Instead of the real cost which cannot be accurately obtained i , so the mathematical model of social welfare expression is:

5. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction according to claim 4 is characterized in that: The formula for constructing the social welfare maximization optimization objective in step 2 is as follows:

6. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction according to claim 1, characterized in that: The step 4 is to extract and cluster user features using the adaptive resonance method ARM, which integrates the ART network and the pruning mechanism. The ART network consists of an input layer, an output layer, and bottom-up and top-down fully connected links; the input layer consists of a user access matrix and a user reputation matrix, and each node of the output layer represents a classified CL, and each participant is assigned to a corresponding node; the input layer and the output layer have two fully directional connections: the bottom-up network weight is represented by W b Indicates that the top-down network weights are represented by W t express.

7. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 6 is characterized in that: The specific process of step 4 is as follows: Step 4.1: Concatenate the user access matrix and the user reputation matrix into a feature matrix C with dimensions |W|×(|Z|+1) t (W); Step 4.2: Convert the feature matrix C t (W) Each row is used as the user's feature input into the adaptive resonance network, and the matching degree between each node and the input vector is calculated: net[j, p] represents the matching degree between the jth CL and user p; Step 4.3: Select the node j with the highest matching degree * ; Step 4.4: Calculate the similarity between node j* and user features: Step 4.5: Compare the similarity with the set threshold. If it is greater than the threshold, update the network weights; otherwise, traverse all other output nodes to find nodes with similarity greater than the threshold; if all nodes in the output layer do not meet the requirements, add a new classification node and assign the user to this category; If the total number of output layer nodes exceeds n, the pruning mechanism is initiated.

8. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 7, characterized in that: The dynamic incentive mechanism algorithm of step 5 includes: Step 5.1: Initialize the network Q with random network parameters ω ω (s, a); Step 5.2: Copy the same parameters ω - ←ω Initialize the target network Step 5.3: Initialize the experience replay pool; Step 5.4: Generate experience data and store it in the replay pool; Step 5.5: Randomly sample from the replay pool for training.

9. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 8, characterized in that: Define the following parameters: Get n user state clusters CL from step 4, state: represents the i-th cluster; Action: a t Select a state cluster from the user state clusters CL; reward: r t (s t , a t ) is in state s t Take action a t The social welfare increment obtained minus the penalty for violating the limited budget constraint is calculated as follows: in, is an adjustable parameter; Reward: The reward a user receives at time t is the bid b corresponding to the user selected at time t-1 i ; The step 5.4 specifically includes the following: Step 5.4.1: Use ARM to get the current environment status t ; Step 5.4.2: Based on the current network Q ω (s, a) selects action a using the ε-greedy strategy t Step 5.4.3: Use the ε-greedy strategy to select the user with the largest current reward in the cluster; Step 5.4.4: Execute action a t , the environment state becomes s t+1 , get reward r t ; Step 5.4.5: t , a t , r t ,s t+1 ) is stored in the replay pool.

10. The method for the incentive mechanism of multi-target mobile crowd intelligence perception based on reverse auction as claimed in claim 9, characterized in that: The step 5.5 specifically includes the following: Step 5.5.1: Randomly sample data from the playback pool {(s k , a k , r k ,s k )} k＝1，...，K ; Step 5.5.2: For each sampled data, use the target network to calculate Step 5.5.3: Update the current network Q with the minimum loss target of the loss function ω , the loss function is: Step 5.5.4: Update the target network

Images