Utilizing Mixture Density Networks for Travel TimeProbability Distribution Predictions
Authors
Kirkeby, Laurids Vinther ; Holm, Mikkel Elkjær
Term
4. term
Education
Publication year
2020
Submitted on
2020-06-12
Pages
21
Abstract
This thesis investigates whether travel time distributions for road segments can be predicted from floating car data using a Mixture Density Network (MDN), which combines deep neural networks with mixture models. Rather than outputting a single mean travel time, the model estimates a full probability distribution, enabling route choice that balances speed and reliability. We use 1.1 billion GPS observations from Denmark (April 2012–December 2014), map-matched to an OpenStreetMap road network with OSRM, aggregated to a topological graph, and reduced to about 42.5 million edge traversals. The model is trained on spatial, temporal, and augmented features including segment length, road type, speed limit, distances to points of interest, and neighborhood embeddings. Training minimizes negative log-likelihood, and the predicted distributions feed pathfinding algorithms to construct routes between origin–destination node pairs under different risk preferences. We also outline experiments comparing against common machine learning models and assessing feature contributions; within this excerpt, the approach achieves a low negative log-likelihood and successfully generates routes, with full results and discussion provided later in the thesis.
Dette speciale undersøger, om rejsetidsfordelinger for vejstrækninger kan forudsiges ud fra flydende bildata ved hjælp af en Mixture Density Network (MDN), som kombinerer dybe neurale netværk med blandingsmodeller. I stedet for kun at forudsige en gennemsnitlig rejsetid estimerer vi hele sandsynlighedsfordelingen, så rutevalg kan afveje hurtighed og pålidelighed. Vi anvender 1,1 mia. GPS-observationer indsamlet i Danmark (april 2012–december 2014), der map-matches til et OpenStreetMap-vejnetsværk med OSRM, aggregeres til et topologisk grafniveau og reduceres til cirka 42,5 mio. kantpassager. Modellen trænes med rumlige, tidslige og supplerende features, herunder vejsegmentlængde, vejtype, hastighedsgrænse, afstande til points of interest og neighborhood-embedding. Træningen optimeres med negativ log-likelihood, og de forudsagte fordelinger kan anvendes i rutealgoritmer til at konstruere ruter mellem oprindelses- og destinationsknuder med ønsket usikkerhedsprofil. Vi skitserer også eksperimenter mod almindelige maskinlæringsmodeller og en vurdering af featurebidrag; i dette uddrag rapporteres en lav negativ log-likelihood og vellykket rute-generering, mens fulde resultater og diskussion forefindes senere i afhandlingen.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords