Space crowdsourcing task allocation regulation and control method for crowd-sourcing cooperative service

Abstract

The invention discloses a space crowdsourcing task allocation regulation and control method for crowd-sourcing cooperative service, and provides a model ACPtrNet based on deep reinforcement learning. The model not only can consider the benefits of a task requester, a crowdsourcing worker and a crowdsourcing platform at the same time, but also combines the advantages of the powerful function fitting capability of deep learning and the decision-making capability of reinforcement learning. The ACPtrNet generates a worker-task matching degree by using an Actor network based on a pointer network, and calculates a real reward value; using a Critic network based on RNN and an attention mechanism to predict an award value which can be obtained by the Actor network; actor and Critic are mutually guided, and strategy gradient descent and stochastic gradient descent are respectively used for parameter updating. Finally, the Actor network can generate a reliable worker-task matching degree, and the task with the highest matching degree with the worker is allocated to the worker. The space crowdsourcing task allocation method has the advantages of being general and suitable for task allocation of multiple scenes in space crowdsourcing.

Description

Technical Field

[0001] The present invention relates to the fields of complex systems and computer technology, and more specifically, to a spatial crowdsourcing task allocation and control method for group intelligence collaborative services. Background Art

[0002] With the rapid development of mobile Internet and geographic information technology, a new location-based crowdsourcing model, spatial crowdsourcing, has emerged. This location-based crowdsourcing service publicly recruits non-specific workers to go to a designated location to complete tasks through an Internet platform. Due to its advantages such as high flexibility, universal applicability and low cost, spatial crowdsourcing has aroused a wave of research in the academic community. At the same time, in the context of the sharing economy era, spatial crowdsourcing has also been widely used in the industrial sector, creating social and economic benefits and service value.

[0003] How to efficiently allocate tasks is a basic and core research problem in spatial crowdsourcing, because other research problems in spatial crowdsourcing are related to task allocation. The extent to which the results of task allocation can meet the expectations of task requesters and crowdsourcing workers is an important factor affecting the potential of users to continue to choose crowdsourcing applications, and then affecting the development of crowdsourcing applications. Therefore, designing a reasonable task allocation algorithm is particularly important to promote the development of spatial crowdsourcing applications. Researchers at home and abroad have done a lot of work, which is of great value in solving the problem of spatial crowdsourcing task allocation. However, there are also some shortcomings:

[0004] (1) Spatial crowdsourcing task allocation is often a multi-objective optimization problem. However, from the current research status, it can be seen that most research works optimize single objectives such as task completion rate, number of assigned tasks, platform revenue, travel cost, response time, and cost within a fixed allocation time interval.

[0005] (2) Metaheuristic algorithms usually require a long period of iteration to obtain a better solution and cannot meet the real-time online task allocation requirements with low latency. Deep learning-based methods find it difficult to obtain high-quality labeled data sets, and inaccurate prediction results will bring additional overhead. Pure reinforcement learning methods often require a very large "action-behavior" space.

[0006] Therefore, it is urgent to design an efficient, reliable spatial crowdsourcing task allocation algorithm that can take into account the interests of multiple parties. Summary of the invention

[0007] In order to overcome the shortcomings of the existing methods, the present invention proposes a spatial crowdsourcing task allocation and control method AC_PtrNet for crowd intelligence collaborative services. This model can improve the quality of task allocation. High-quality task allocation not only helps to improve user satisfaction, but also increases the benefits of the platform and workers, thereby promoting the development of spatial crowdsourcing applications and providing new ideas for solving the complex problem of spatial crowdsourcing task allocation.

[0008] To achieve the above purpose, the technical solution of the present invention is a spatial crowdsourcing task allocation and control method for group intelligence collaborative services: the model defines the task allocation problem as a multi-objective optimization problem, while taking into account the interests of all parties involved in the crowdsourcing platform. In addition, the algorithm takes advantage of deep reinforcement learning's ability to solve complex system problems, combining the advantages of deep learning's powerful function fitting ability and reinforcement learning's decision-making ability. The specific steps are as follows:

[0009] 1) Data Processor: The purpose of this part is to convert different types of raw data into a unified model specified in this paper. AC_PtrNet will first pre-process the data to remove outliers and missing values ​​to ensure the integrity and accuracy of the data, and then extract features from the data for better analysis and modeling. In this way, each spatial task can be represented as<loctj,timetj,budgettj> The data is then converted into a vector representation that the model can recognize using the embedding layer. The data is simultaneously input into the "worker-task" matching degree generator and the "worker-task" allocation result judger.

[0010] 2) "Worker-task" matching degree generator: This part aims to solve the task allocation problem in spatial crowdsourcing as a Seq2Seq problem. Specifically, in order to better measure the matching degree between workers and tasks, AC_PtrNet uses the attention mechanism to score all tasks within the distance constraint of a certain worker, and sorts the task matching degree from large to small according to the scoring results. For this reason, the worker selects the task with the highest matching degree for allocation and calculates the reward value of the allocation result.

[0011] 3) "Worker-task" assignment result judge: This part aims to predict the reward value of the assignment result of the "worker-task" matching degree generator and judge the quality of the assignment result. Specifically, the critic network composed of bidirectional GRU and self-attention mechanism aims to output the predicted reward value of the task matching degree sequence and evaluate the quality of the "worker-task" matching degree generator.

[0012] 4) Reinforcement Learning Trainer: This part uses the idea of ​​reinforcement learning to perform unsupervised training on the model. The core is to use Policy Gradient (PG) and Stochastic Gradient Descent (SGD) to guide the Actor network represented by the "worker-task" matching degree generator and the Critic network that constitutes the "worker-task" assignment result judge to update parameters, with the aim of improving the accuracy of the matching degree between the tasks generated by the Actor network and the workers.

[0013] Compared with the prior art, the advantages of the present invention are:

[0014] 1) The allocation algorithm proposed in the present invention not only takes into account the interests of task requesters, crowdsourcing workers and crowdsourcing platforms, but also combines the advantages of deep learning's powerful function fitting ability and reinforcement learning's decision-making ability.

[0015] 2) The technical solution proposed in the present invention is general and can be transferred to other areas for allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings:

[0017] Figure 1 It is a general framework diagram of the AC_PtrNet model provided by an embodiment of the present invention;

[0018] Figure 2 It is a model diagram of the “worker-task” matching degree generator;

[0019] Figure 3 It is a model diagram of the “worker-task” allocation result judge; DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] An embodiment of the present invention provides a spatial crowdsourcing task allocation and control method for group intelligence collaborative services. The specific implementation process includes a data processor, a "worker-task" matching degree generator, a "worker-task" allocation result evaluator, and a reinforcement learning trainer, and further includes:

[0022] Data processor: The first step of the AC_PtrNet model is data processing, the purpose of which is to assign the most suitable task to the worker waiting to be assigned a task on the crowdsourcing platform at a certain moment. Therefore, the data input to the "worker-task" matching degree generator and the "worker-task" assignment result judger in this paper is the set of all candidate tasks within the constraint distance of the crowdsourcing worker w at a certain moment.

[0023] "Worker-task" matching degree generator: This part is designed based on the pointer network. The overall architecture is as follows Figure 2 As shown in Figure 1, it is divided into three parts: encoder, attention-based decoder, and real reward calculator. The encoder obtains the initial vector representation xj of the candidate task. Each candidate task consists of a triple<loctj,timetj,budgettj> The three elements are composed of different elements, and their representations and meanings are different. The value of the information they contain is equally important. Therefore, it is necessary to train three different semantic acquisitions (SA) to extract the important features contained in different elements according to their different characteristics. The detailed structure of SA is as follows: Figure 1 The Encoder part in . First, use the embedding layer to convert each element into a sequence that the neural network can recognize. Then use the corresponding LSTM as the core component to learn the vector representation of the element. Finally, the vector representations of the three elements are directly concatenated and fused to obtain the final vector representation xj of tj. The vector representation xj of a candidate task read by the Encoder at each step is represented as a vector v containing the semantic information of the input sequence, which is finally input into the decoding network by the encoding network. Use the decoder based on the attention mechanism to obtain the matching degree between the final task and the worker, and finally calculate the true reward value of the result.

[0024] "Worker-task" assignment result judger: Figure 3The model diagram of the "worker-task" matching degree generator is described. The specific implementation process is mainly divided into the following four steps: (1) First, the input data passes through the same Semantic Acquisition layer as the Encoder of MDG to obtain the initial vector representation xj of the candidate task. (2) Then, the feature vector representation hj of the candidate task is obtained through a bidirectional GRU. (3) Then, in the self-attention mechanism layer, the value vectors of each feature vector are multiplied by their respective attention scores and then added to obtain the final semantic vector vector. (4) Finally, vector passes through a multilayer perceptron (MLP) to obtain the predicted incentive value. Reinforcement learning trainer: This part uses the idea of ​​reinforcement learning to perform unsupervised training on the model. The core is to use policy gradient (PG) and stochastic gradient descent (SGD) to guide the Actor network represented by the "worker-task" matching degree generator and the Critic network constituting the "worker-task" allocation result judge to update the parameters, with the aim of improving the accuracy of the matching degree between the tasks generated by the Actor network and the workers.

[0025] Reinforcement learning trainer: Use policy gradient to guide the update of the "worker-task" matching degree generator, and use stochastic gradient descent to guide the update of the "worker-task" allocation result judge.

Claims

1. A spatial crowdsourcing task allocation and control method for group intelligence collaborative services, which is characterized by being able to take into account the interests of task requesters, crowdsourcing workers and crowdsourcing platforms at the same time, and combining the advantages of deep learning's powerful function fitting ability and reinforcement learning's decision-making ability. Specifically, it includes the following steps: 1) Data Processor: The purpose of this part is to convert different types of raw data into a unified model specified in this paper. AC_PtrNet will first pre-process the data to remove outliers and missing values ​​to ensure the integrity and accuracy of the data, and then extract features from the data for better analysis and modeling. In this way, each spatial task can be represented as<loctj,timetj,budgettj> The data is then converted into a vector representation that the model can recognize using the embedding layer. The data is simultaneously input into the "worker-task" matching degree generator and the "worker-task" allocation result judger. 2) "Worker-task" matching degree generator: This part aims to solve the task allocation problem in spatial crowdsourcing as a Seq2Seq problem. Specifically, in order to better measure the matching degree between workers and tasks, AC_PtrNet uses the attention mechanism to score all tasks within the distance constraint of a certain worker, and sorts the task matching degree from large to small according to the scoring results. For this reason, the worker selects the task with the highest matching degree for allocation and calculates the reward value of the allocation result. 3) "Worker-task" assignment result judge: This part aims to predict the reward value of the assignment result of the "worker-task" matching degree generator and judge the quality of the assignment result. Specifically, the critic network composed of bidirectional GRU and self-attention mechanism aims to output the predicted reward value of the task matching degree sequence and evaluate the quality of the "worker-task" matching degree generator. 4) Reinforcement Learning Trainer: This part uses the idea of ​​reinforcement learning to perform unsupervised training on the model. The core is to use Policy Gradient (PG) and Stochastic Gradient Descent (SGD) to guide the Actor network represented by the "worker-task" matching degree generator and the Critic network that constitutes the "worker-task" allocation result judge to update parameters, with the aim of improving the accuracy of the matching degree between the tasks generated by the Actor network and the workers.

Images