| In recent years,due to the rapid development of the Internet of Things technology,smart devices have gained explosive popularity,such as smartphones,smart wearables,etc.These devices are equipped with powerful sensors,which can collect different data anytime and anywhere,thus creating the birth and development of mobile crowd sensing(MCS)technology.In mobile crowd sensing(MCS),firstly,the task is sent to the sensing platform by the task publisher,and the platform selects the participants to execute the task.Then,participants collect data and upload it to the platform.Finally,the platform organizes and analyzes the data and sends it to the task publisher.The entire process effectively utilizes the mobility of the population and the idea of crowdsourcing,which enables MCS to obtain high-quality sensing data at a lower cost,thus receiving great attention.Compared with the traditional sensing paradigm,mobile crowd sensing is less susceptible to interference from various factors,and has wide application range and flexibility.However,practically,crowd participation in MCS faces some fundamental challenges,such as task allocation,incentive mechanisms,privacy protection,resource constraints,data security,and other issues.Therefore,the reasonable sensing strategy is crucial for the application of mobile crowd sensing in reality.Based on the analysis of existing schemes,this thesis divides sensing strategies into two types:platform based sensing strategies and platform and user interaction based sensing strategies,and conducts research on them respectively.The main contributions of this thesis are as follows:(1)This thesis analyzes,summarizes,and compares the existing sensing strategies,and discusses the future research directions.Currently,sensing strategies are mainly divided into platform based sensing strategies and platform and user interaction based sensing strategies based on different perspectives.At the same time,platform based sensing strategies can be divided into offline sensing strategies and online sensing strategies.Platform and user interaction based sensing strategies can be divided into data driven sensing strategies and other interaction based sensing strategies.Both schemes have their advantages and disadvantages.In addition,future research directions have also been discussed based on existing strategies and theories in other fields.(2)A RL-based online participant selection scheme incorporating area coverage ratio and degree in mobile crowdsensing(PSARE)is proposed.Considering the cold start problem of MCS online tasks in reality,firstly,this thesis proposes an online participant selection scheme using the Q-Learning in reinforcement learning.Then,in order to avoid the state explosion problem of Q-Learning in practical applications,a two-stage Q-Learning idea is proposed.The first stage classifies participants based on their trajectories,and the second stage selects the participants with the highest coverage quality in the class in the form of ε-greedy to execute the task.In addition,in measuring the coverage quality of MCS,both the coverage ratio of the general area and the coverage degree of the hotspot area are considered.The experiment shows that the proposed scheme has higher overall and timeslot level coverage ratio and coverage degree than other comparison schemes on the real dataset,and has good perceptual effects.(3)In the execution of MCS tasks,participants are influenced by random environments and other participants,which makes it difficult for them to obtain stable income.In order to solve this problem,this thesis proposes a multi-agent reinforcement learning based mobile crowd sensing strategy(MARCS).Firstly,the proposed scheme depicts the trajectory of participants and uploads the depicted trajectory and sensing data collected by the participants.Secondly,the activity factor of each participant is calculated through the trajectory,where the activity factor represents the level of activity of the participant in the task.Then,the platform uses the activity factor and sensing data to calculate data quality,and assigns rewards to each participant based on data quality through the Shapley value method.Finally,the multi-agent reinforcement learning method MADDPG is used to provide each participant with the optimal strategy that is beneficial for their income.The experiment shows that the proposed scheme has better results than other schemes in terms of participant income and platform efficiency. |
