Data Driven Analysis in Transportation Systems

Transcription

1 Data Driven Analysis in Transportation Systems Ugur Demiryurek, Ph.D. Associate Director, Integrated Media Systems Center (IMSC) Viterbi School of Engineering University of Southern California Los Angeles, CA

2 Facilitating an infrastructure for acquiring, processing, storing and querying real-time and historical transportation datasets 2

3 Team Government LA-Metro: Los Angeles County Metropolitan Transportation Authority RIITS: Regional Integrated Intelligent Transportation System USC Price School of Public Policy Stevens Institute Industrial Engineering Marshall School of Business Industry M\ a National Science Foundation (NSF) I 3

4 Rethinking Congestion Texas Transportation Institute (TTI): Best know source for congestion and mobility studies 4

5 Rethinking Congestion The reports are static (quarterly) Corridors have very different behaviors Spatial and temporal granularity of average values are high level (e.g., one hour) Not customized for needs of local authorities 5

6 Demo 6

7 Studied 28 corridors in Los Angeles County Evaluated Planning-time index E.g., Planning-time index is 1.60 and the free-flow travel time is 15 minutes, then Planning-time = = 24 minutes Travel-time index Hours delayed Gallons wasted Rethinking Congestion 7

8 Rethinking Congestion Scatter plots of the TTI measures compared with the IMSC measures. The straight lines indicate where measures by TTI equal measures by IMSC. Total number of data points in each table is 28. The triangle points show the side of the deviation where most observations are. 8

9 Main findings Rethinking Congestion TTI Underestimated: Planning-time index TTI Overestimated: Travel-time, delays, gallons wasted Planning Time Index Travel Time Index Delay Personhour Gallons Wasted 17% -8% -23% -7% Comparisons of measures between TTI and IMSC dataset 9

10 Rethinking Congestion There are more extreme congestion conditions than expected during the peak hours Our estimated planning time indexes are 17 percent larger than the TTI estimates Drivers should be prepared to leave more slack time during peak hour trips But overall (for 24 hours a day and 7 days a week) congestion is less than expected Annual delay and total congestion cost to be around 2/3 of the stated value from TTI Q. An, U.Demiryurek, P. Gordon, J. Moore, C. Shahabi, Los Angeles Congested Corridor Study and Comparisons with Texas Transportation Institute Congestion Estimates, Industrial and Systems Engineering Research Conference (ISERC), San Juan, Puerto Rico, May

11 ClearPath The weight of an edge is a function of time, i.e., timedependent. Arrival-time to an edge determines the travel-time on that edge. 8 :30 AM 5:00 PM 11

12 ClearPath USC Stevens fund ClearPath, a spin-off from USC R&D $2+ million research funding in last 3 years 10 years of research on route planning Access to massive traffic sensor data Intellectual Property 3 patents: 2 filed and 1 provisional Team Complementary expertise in engineering and business 12

13 ClearPath Inrix Google Maps NAVTEQ Waze ClearPath Basic Route guidance Shortest distance Road-type selector Real-time traffic GUI interface Static POI Crowdsourcing Predictive Paths * Anticipate event traffic * FastPath POI * *Patent Pending 13

14 ClearPath Value Proposition Improvement over Google Map s paths: 18% 64 hours per year saved by average commuter $840 Average fuel consumption savings 14

15 Demo 15

16 ClearPath Goals: Exact fastest path in large time-dependent road networks Fast response time Fast precomputation Low space overhead s Efficient update operations d U. Demiryurek, F. Banaei Kashani, C. Shahabi, A. Ranganathan: Online Computation of Fastest Path in Time-Dependent Networks. In SSTD A. Akdogan, U. Demiryurek, C. Shahabi: Voronoi-Based Geospatial Query Processing with MapReduce. In IEED CloudCom 2010 (Best paper award) 16

17 ClearPath Dreyfus vs. Time-dependent A* d d Dreyfus s s TD A*Search increasing cost v j s v i h(v) =(v i,d) d Optimality Condition: h(v) should not overestimate the actual distance between v and d 17

18 ClearPath Challenge: Finding Heuristic Function ht(vi,d) <= D(vi,d,t) in TD Networks v j s v i h t (v i,d) d The distance (travel-time) between any node v and d is time-dependent. 18

19 Time-dependent A* Search Naïve Heuristic Function: D EUC ( v, d) max( speed ) i Euclidean distance between v and d divided by the maximum speed among the edges Guaranteed to be a lower-bound as the distance between v and d is never overestimated Problem: It is a very loose bound, hence yields insignificant performance improvement 19

20 Time-dependent A* Search Goal: Find a h(v i ) that will never overestimate the time-dependent traveltime between v i and d. This is necessary for Exact results h(v i ) should be as close as possible to actual distances for Efficient processing of fastest path computation 20

21 Time-dependent A* Search Goal: Find a h(v i ) that will never overestimate the time-dependent traveltime between v i and d. This is necessary for Exact results h(v i ) should be as close as possible to actual distances for Efficient processing of fastest path computation Approach: Step 1: Partition the road network into non-overlapping partitions (Offline) Step 2: Precompute h(v i ) using distances in and between the nonoverlapping partitions (Offline) 21

22 Time-dependent A* Search Step 1: Partition the road network using network hierarchies Border Nodes S 1 S 2 S 3 S 4 S 5 S 6 S 7 S 10 Highway Edges Arterial Edges S 8 S 9 S 11 22

23 Time-dependent A* Search Step 2: Compute intra and inter distance labels Intra: fastest path in Lower-bound Graph G from each node v i to border nodes and border nodes to v i Inter : fastest path in Lower-bound Graph G between border nodes b i b k v i d b j b l S i Only store the minimum of the intra and inter distance lables LTT( vi, bi ) arg min( LTT( vi, bi ), LTT( vi, bj )) LTT( bl, d) arg min( LTT( bk, d), LTT( bl, d)) LTT( bi, bk ) arg min( LTT( bi, bk ), LTT( bi, bl ), LTT( bj, bk ), LTT( bj, bl )) S j 23

24 Time-dependent A* Search Fast h(vi,d) computation h(v i,d) is computed by simple table look-ups (nanoseconds) b 19 b 17 b 13 b 23 b 2 b 1 b 14 b 12 n 16 S 1 n 3 S n 7 b 29 b 4 b 9 b 30 b 7 b 8 b 19 b 32 b 10 b 33 S 4 Border Node Partition Border Node-to- Distance Border-to- Partition Border Node h(n 7,d) = n b 6 5 S 1 b 3 b 23 14,5 b S 23 1,S,7 3 n b 7 9 S 1 b 12 b 4 18,6 b S 9 1,S, b 1 b S 4,S 1 n 16. S 4. b 17.,3 b 33,6,5 24 nb n b k u,.. n S k b u,,x b v,y

25 Future Directions More on event impact analysis Games, concerts, Obama visit New corridor analysis Light rail opening Express Lane opening ClearPath Field tests and surveys New features (Delivery,Multimodal) 25

26 Questions? 26