Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Low-Rank Cyclostationarity Predictive Routing Is Almost as Good as Real-Time Data-based Routing

Oriel-Singer, Ilai-Bistritz, Giseung-Park, Woohyeon-Byeon, Youngchul-Sung, Amir-Leshem

TL;DR

This work develops a spatiotemporal predictor based on a low-rank decomposition of the traffic matrix and the temporal subspace coefficients that incurs an average excess travel time of less than 1.5 minutes and matches that of a near-real-time predictor.

Abstract

Dynamic shortest-path routing, using real-time traffic data, enables path selection responsive to evolving conditions. Nevertheless, transportation planning tasks such as adaptive congestion pricing, fleet routing, and long-term operational decisions rely on offline traffic estimators. To address this problem, we develop a spatiotemporal predictor based on a low-rank decomposition of the traffic matrix and the temporal subspace coefficients. Using a recent large-scale measurement campaign over the Seoul road network, we show that our proposed predictor incurs an average excess travel time of less than 1.5 minutes. Moreover, our predictor's tail of the excess travel time distribution matches that of a near-real-time predictor. Results based on one year of traffic data are also demonstrated in simulations.

Low-Rank Cyclostationarity Predictive Routing Is Almost as Good as Real-Time Data-based Routing

TL;DR

This work develops a spatiotemporal predictor based on a low-rank decomposition of the traffic matrix and the temporal subspace coefficients that incurs an average excess travel time of less than 1.5 minutes and matches that of a near-real-time predictor.

Abstract

Dynamic shortest-path routing, using real-time traffic data, enables path selection responsive to evolving conditions. Nevertheless, transportation planning tasks such as adaptive congestion pricing, fleet routing, and long-term operational decisions rely on offline traffic estimators. To address this problem, we develop a spatiotemporal predictor based on a low-rank decomposition of the traffic matrix and the temporal subspace coefficients. Using a recent large-scale measurement campaign over the Seoul road network, we show that our proposed predictor incurs an average excess travel time of less than 1.5 minutes. Moreover, our predictor's tail of the excess travel time distribution matches that of a near-real-time predictor. Results based on one year of traffic data are also demonstrated in simulations.
Paper Structure (10 sections, 9 equations, 4 figures, 2 tables, 1 algorithm)

This paper contains 10 sections, 9 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. problem formulation
  4. Spatiotemporal Predictor
  5. Numerical Results
  6. Preprocessing the data
  7. Training and test sets
  8. Analysis and comparison
  9. Conclusions
  10. Supplementary material

Figures (4)

  • Figure 1: Distribution and tail behavior of routing regret. Left: Cyclostationary predictors (red/purple) exhibit regret distributions closely matching 10-minute-lag (blue) and static (brown) benchmarks. Right: log CCDF of tail regret on log scale.
  • Figure 2: 10% upper quantile regret as a function of path length (number of hops in the unweighted shortest path per OD pair). Only path lengths with more than 1,000 samples are shown to ensure stable quantile estimates.
  • Figure 3: MDL criterion vs. number of modes $k$. The minimum is at $k=73$ (dashed line), but the curve nearly plateaus around $k=25$.
  • Figure 4: Power spectral density of temporal singular vectors. Modes 1 and 2 are dominated by daily periodicity ($1\,\text{day}^{-1}$), while Mode 7 exhibits a stronger weekly component ($1/7\,\text{day}^{-1}$).