Estimating Travel Cost Distributions of Paths in Road Networks using Dual-Input LSTMs
Authors
Nielsen, Christopher Hansen ; Randers, Simon Makne
Term
4. term
Education
Publication year
2020
Submitted on
2020-12-29
Pages
8
Abstract
Fremskridt i sensorteknologi giver nu detaljerede data om, hvor «dyrt» det er at køre på veje, og hvordan denne omkostning varierer. Sådanne data gør det muligt at beskrive usikkerhed som fulde sandsynlighedsfordelinger i stedet for blot et enkelt gennemsnit. I dette arbejde estimerer vi fordelingen af rejseomkostning for hele ruter i et vejnet ved hjælp af floating car data (GPS-spor fra kørende køretøjer). Vi undersøger to måder at repræsentere fordelinger på: histogrammer og Gaussian Mixture Models (GMM’er). Givet en rute og et afgangstidspunkt estimerer vi rutenes rejseomkostningsfordeling med en dual-input long short-term memory (DI-LSTM) model, en type neuralt netværk til sekvensdata. For hver vejstrækning kombinerer modellen to inputfordelinger: fordelingen for den aktuelle strækning (edge) og den akkumulerede fordeling for ruten indtil denne strækning, som overføres fra den forrige DI-LSTM-enhed. Denne sammenfletning styres af to nye porte (gates), der bestemmer, hvordan de to input blandes. Forsøg på et stort datasæt af køretøjstrajektorier viser, at DI-LSTM overgår en klassisk LSTM, især for længere ruter. Vi finder desuden, at GMM’er ofte er bedre egnet end histogrammer som fordelingsrepræsentation.
Advances in sensor technology now provide detailed data about the “cost” of traveling on roads and how that cost varies. Such data make it possible to model uncertainty as full probability distributions rather than a single average. This work estimates the distribution of travel cost for entire paths in a road network using floating car data (GPS traces from moving vehicles). We examine two ways to represent distributions: histograms and Gaussian Mixture Models (GMMs). Given a path and a departure time, we estimate the path’s travel-cost distribution with a dual-input long short-term memory (DI-LSTM) model, a type of neural network for sequence data. At each road segment, the model combines two input distributions: the distribution for the current segment (edge) and the accumulated distribution for the path up to that segment, carried over from the previous DI-LSTM unit. Two new gates control how these inputs are merged. Experiments on a large vehicle-trajectory dataset show that DI-LSTM outperforms a classic LSTM, especially on longer paths. We also find that GMMs are often better suited than histograms for representing distributions.
[This abstract was generated with the help of AI]
Documents