Method for calculating the trajectory likelihood of recurrent neural network based on output state constraints

Abstract

The present invention belongs to the field of trajectory calculation technology, in particular to a method for calculating the trajectory likelihood of a recurrent neural network based on output state constraints. The method includes the following steps: a training phase of modeling a trajectory by using a recurrent neural network with output state limitation and training the model parameters according to historical trajectory data; and an online query phase of predicting the likelihood of moving to the next road section of each road section of the inputted trajectory according to the trained model. The method of the invention utilizes the strong expression ability of the deep neural network to learn the long-distance dependence which cannot be captured based on the Markov model and thus can more accurately model the trajectory data.

Description

technical field [0001] The invention belongs to the technical field of trajectory calculation, in particular to a method for calculating the likelihood probability of a trajectory of a recurrent neural network based on output state constraints. Background technique [0002] The popularity of mobile GPS devices has promoted the development of location-based services, and trajectory calculation has also emerged as the times require. With the acquisition of a large amount of trajectory data, it is possible to establish a probability model for these trajectory data. Trajectory model is a very important and fundamental problem in trajectory calculation. Probabilistic modeling of trajectories can help solve application problems related to driving paths, such as route recommendation, road condition prediction, trajectory prediction, frequent pattern mining, anomaly detection, etc., and has very large application scenarios. Existing trajectory probability modeling techniques are m...

AI Technical Summary

Problems solved by technology

[0006] There are also some works that use the model of reverse reinforcement learning to probabilistically model the trajectory. The main problem of these methods is that using Markov decision process for modeling and using the model of reverse reinforcement learning, although it can capture the future information, but still cannot get rid of the Markov property; although the model using maximum entropy reverse reinforcement learning can capture long-distance dependencies, it is constrained by the lack of model parameters and the poor expressiveness of the model.

Images