A Novel Approach for Heterogeneous Task Allocation in Mobile Crowd Sensing Using Deep Reinforcement Learning Based on Reverse Stackelberg Game

Abstract

Today, the cloud space has become the main platform for communication and interactions between users. With the advent of Internet of Things (IOT) technology, cloud systems face a huge volume of requests every day that need to be implemented in real time and efficiently. There are many situations where the exact determination of the status of the requests and how to implement them face serious challenges. In fact, in the IOT environment with the large variety of users and the modeling of many heterogeneous requests from them, it is crucial to adopt a flexible approach. Therefore, deriving a suitable scheduling mechanism that can minimize both task execution delay and cloud resource utilization is of great importance. One of the most common methods offered to manage the huge activities created in the network is the use of resources placed in the edge nodes of the network under the title of Mobile Crowd Sensing. However, in various applications of mobile crowd sensing, such as in the field of health or in intelligent transportation systems, the structure of requests sent by actors has a very high diversity. This diversity in requested activities and the high number of requests suggest the need for proper management of available resources. In this article, the Reverse Stackelberg game theory method consisting of fuzzy logic is used in order to quickly adopt an effective strategy for the optimal allocation of resources by gaining experience from its past performance. In this regard, in order to achieve the desired quality in assigning heterogeneous tasks among users, deep learning is used, which has useful features such as being online and highly adaptable. On the other hand, in order to perform tasks that require low delay, information about the location of the user and mobile, it is necessary to use the capabilities of the fog processing environment in interaction with the cloud space so that the capacity of the end layers of the network can be used well. The obtained results show that by using the proposed approach, more than 35% of CPU usage cost is saved compared to other state of the art methods.