Research On Mobility Prediction Based On Path And Trip Planning

In recent years,with the evolution of Internet of things,cloud computing,mobile Internet,wireless communication network and other technologies,"smart tourism" makes use of Internet and other technologies and portable terminal devices to customize tourism products for tourists and improve their travel experience.Among them,the user mobility prediction and trip planning are widely concerned.This thesis studies the mobility prediction method based on path semantics and the personalized trip planning problem based on reinforcement learning.The two subproblems are studied in depth respectively,and the deficiencies of the research content and the possible improvement directions are described.Firstly,the thesis summarizes the relevant technologies and research contents of mobility prediction,and analyzes the classification and characteristics of spatiotemporal features,as well as summarizes the traditional travel salesman problem and the reinforcement learning theory related to this paper.Secondly,aiming at the problem that traditional location prediction based on region location or point of interest ignores the details of movement,which can lead to difficult to weigh on prediction accuracy and geographical precision,a framework of mobility prediction based on the path semantic is designed,including the path sequence extraction and model training.The path sequence extraction algorithm not only eliminates the trajectory noise introduced by the region position,but also avoids the conflict between the region granularity and the prediction accuracy,and realizes the accurate reconstruction of the original trajectory.In the extraction of path sequences,the candidate junctions are extracted by drawing a similar point diagram of the original GPS trajectory.Furthermore,redundant junctions are filtered by clustering algorithm,and the similar trajectories between junctions are finally merged to typical paths.In this thesis,a long-short term memory neural network is used to model and predict the constructed trajectory sequence.The experiment shows that the prediction based on path is much higher than that based on regional location on geographical precision and improves the accuracy to some extent.Thirdly,the traditional trip planning methods not fully considering the low efficiency caused by the expansion of the scale of interest points and users' travel preferences.In this thesis,under the premise of multiple cost constraints,a sequence-to-sequence based neural network combining the pointer mechanism is introduced to optimize the trip through deep reinforcement learning algorithm A3C.In contrast,the general heuristic search algorithm is easy to fall into the local optimal solution,and search efficiency is not high.Experimental results show that the proposed trip planning algorithm has good convergence,and the solution quality and search efficiency are improved.At last,the thesis is summarized and the development trend in the future is pointed out,some improvement directions are put forward for deficiencies,laying a foundation for further optimization.