Research And Implementation Of Mobile Crowd Sensing Task Allocation Algorithm Based On Deep Reinforcement Learnin

Mobile crowdsensing is a new mode for collecting and mining data and intelligent decision-making with mobile intelligent devices.The key to the high performance of MCS is the efficient method of task allocation.In recent years,the combination of deep reinforcement learning and mobile crowdsensing has made the related tasks of machine learning possible on large-scale smart mobile devices.Deep reinforcement learning has also proven to have significant advantages in using large data to train object recognition,classification and future event prediction models.The traditional algorithm(greedy algorithm or ant algorithm)assumes that workers and tasks are static.It’s not fit for the scene where the position and time of workers and tasks change continuously.In addition,the existing methods usually make decisions by the central server based on the collected information,which usually leads to leakage of workers’ privacy.The introduction of deep reinforcement learning algorithms can effectively solve the shortcomings of traditional allocation algorithms in policy fairness and model efficiency.Based on past experience,it can continuously adjust the probability distribution of strategy selection in each decision-making process by selecting actions and interacting with the environment to obtain corresponding status and rewards,and ultimately achieve global optimum.In view of the above issues,this paper will study from the following two aspects:(1)A task allocation method based on deep reinforcement learning(DRL)with privacy protection.Firstly,aiming to maximize the two-way benefits of workers and platforms and realizes Nash equilibrium,the task allocation is modeled as a dynamic programming problem of multi-objective optimization.Secondly,the model based on proximal policy optimization(PPO)of DRL for training and learning model parameters is proposed.Finally,we use the local differential privacy method to add random noise to the sensitive information of workers to protect privacy.The central server trains the whole model to obtain the optimal allocation strategy.In this paper,the astringency,revenue and task cover rate are experimentally evaluated.The results show that the proposed method has significant improvement in different indexes,and can protect the privacy of workers,compared with the traditional methods and other DRL based methods.(2)Task assignment for deep reinforcement learning based on federated learning(FL).Based on traditional privacy protection technologies and task assignment frameworks,a federated learning framework is applied.In each iteration,workers train local models on smart mobile devices with the sensory data they collect.Then upload local model parameters to the server and update the global model through parameter aggregation.The server will broadcast the updated global model locally to update the local model.Through comparative experiments under simulated datasets,the model is superior to the traditional MCS task assignment algorithm in convergence,model loss,model training accuracy,total cost and total overhead,and can reduce the risk of privacy leaks while guaranteeing good performance.