Research On Data Collecting Method Based On Grouping Multi-agent Deep Reinforcement Learning In Mobile Crowdsensing

Mobile Crowdsensing is a promising sensing paradigm that recruits users carrying sensing devices to cooperatively perform sensing tasks.The system is usually composed of three parts: users,platforms and task requesters.It has wide coverage,low deployment cost and good scalability.However,for some special scenarios,users cannot complete the task.For example,people with sensing devices could not enter the target task areas,tasks require high-precision data or computing power.To this end,unmanned aerial vehicles(UAVs)as the powerful sensing devices equipped with different kinds of high-precision sensors could be used to replace user participation and carry out the special crowdsensing tasks.In this paper,we focus on scheduling UAVs to sense the task Point-of-Interests(Po Is)with different frequency coverage requirements.Specifically,different from the previous works that only focus on the data collection amount of Po Is,each Po I in our scenario needs to be collected with certain frequency coverage requirements,which means sensing more than once every period of time is meaningless.Therefore,UAVs need to continuously patrol among Po Is during the task cycle.In addition,there are several charging stations in the area which can charge the UAVs.Such problem is novel and can be applied to many specific scenarios,such as epidemic monitoring,precision agriculture,environmental monitoring,etc.When scheduling UAVs to perform the task,we not only pay attention to the coverage of frequency requirements,but also the geographical fairness.We need to collect more diverse data for global analysis,rather than just focusing on some Po Is.Meanwhile,because of the limited initial energy reserve,UAVs should keep balancing between data collection and energy charging.Charging at different times and different positions has a great influence.To sum up,to accomplish the sensing task,the scheduling strategy needs to consider the coverage requirement,geographic fairness and energy charging simultaneously.In our problem,multiple UAVs work together to complete the task cooperatively.The flight strategy of each UAV is affected by others,so the scheduling strategy needs to consider the complex interaction among UAVs.Therefore,we solve our problem based on a multi-agent deep reinforcement learning method MADDPG.It can schedule UAVs distributively.We re-adjust the reward function by adding a new function which returns an immediate reward by considering coverage time and other UAVs’ actions for dealing with the frequency coverage requirements.As the number of agents increases,the training time is too long to converge.To this end,we propose a grouping multi-agent deep reinforcement learning method GMADDPG.It groups all UAVs into some teams by a distance-based clustering algorithm(DCA),then it regards each team as an agent.In this way,G-MADDPG solves the problem that the number of agents is too large.After determining the scheduling strategy of each team,an equal pattern that each UAV patrols along the route of its team with a certain starting time difference is proposed.Moreover,G-MADDPG can control the trade-off between training time and result accuracy flexibly by adjusting the number of teams.We conduct extensive simulations to verify the performance of our algorithm and compare it with three baselines based on three evaluation indicators.The results show that our scheduling strategy has better performance and is flexible in balancing training time and result accuracy.