da yan, zhou zhao and wilfred ng the hong kong university of science and technology

34
Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Upload: oscar-cook

Post on 17-Dec-2015

215 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Da Yan, Zhou Zhao and Wilfred NgThe Hong Kong University of Science and Technology

Page 2: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

2

Page 3: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

IntroductionHow to define an optimal meeting

point on road networks?

9 km

6 km 3 kmSix

people

Min-max OMP

arg minx [maxi dist(qi, x)]

q2

q3

q4

q5

q6

q1

x

3

Page 4: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

IntroductionHow to define an optimal meeting

point on road networks?

9 km

3 km

Min-sum OMP

arg minx [∑i dist(qi, x)]

q2

q3

q4

q5

q6

q1

x

Our focus

4

Page 5: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

IntroductionNetwork distance dN(p1, p2) is the

length of the shortest path connecting p1 and p2

Problem DefinitionGiven a set of query points Q ={q1, q2, …,

qn} on a road network G = (V, E), an optimal meeting point (OMP) query returns the point x’ = arg minx [∑i dN(qi, x)]

5

Page 6: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

IntroductionApplications:

Minimizing the total travel cost for a group of people who want to find a location for gathering

Helping a travel agency decide the location for a tourist bus to pick up the tourists

Part of AI for computer players in strategy games such as WorldofWarcraft

6

Page 7: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

7

Page 8: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Related WorkMin-sum OMP query in the

Euclidean space: a.k.a. the Weber problem

Min-sum OMP in the Euclidean space: geometric median of Q ={q1, q2, …, qn}

No closed form formula existsSolved by gradient descent methods

8

Page 9: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Related WorkMin-sum OMP query on a road

network G = (V, E)Split-point-based methodA split point is defined for each

query point qi ∈ Q and each edge (u, v) ∈ E 4 km

6 km

12 km 14 km

u v

qi 9

Page 10: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Related WorkExisting work has proved the

following:An OMP must exist among all the

split pointsSearch space: |Q| · |E|Computational cost:

Split point evaluationMin-sum distance computation

10

Page 11: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

11

Page 12: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Baseline AlgorithmTheorem 1: Given an OMP query

with query point set Q on a road network G = (V, E), an OMP must exist among V ∪ Q

Search space: |V| + |Q|Computational cost:

Only min-sum distance computation

|Q| · |E| |V| + |Q|

Search Space

12

Page 13: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Baseline AlgorithmMain idea:

Suppose that no point in Q is on edge (u, v)

uv x x'

δ

q2q1 q6

q5q4q3

q8

q7

Part of q1’s shortest path to x

+ 2 ·δ − 6 ·δshorter

13

Page 14: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Baseline AlgorithmTheorem 1 only relies on the fact

that the road network G = (V, E) is a graph

Edge length can refer toPhysical distanceTravel delay……

Can we do more ?

14

Page 15: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

15

Page 16: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningExisting method:

Cut the whole road network into partitions

Check only those split points that are in the smallest enclosing partition

16

Page 17: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningFirst Trial

Collect into a set P those end points of all the edges which the query points in Q are on

Compute the convex hull of the point set P u

v

a

bOMP

17

Page 18: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningFirst Trial

Collect into a set P those end points of all the edges which the query points in Q are on

Compute the convex hull of the point set P

18

Page 19: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningSecond Trial

Collect into a set P those end points of all the edges which the query points in Q are on

Compute the convex hull of the point set P as H

Find the shortest path for each pair of neighboring points on H, add all the points on these paths into a set S

Compute the convex hull of the point set S

19

Page 20: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningSecond Trial

Only fails to return the OMP in 5 of the 79900 queries tested in total

Sum-of-distances values of these meeting points are all within 0.1% more than the smallest value

u

v

a

b

20

Page 21: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Convex-Hull-Based PruningAll vertices in V are organized as a kd-

treeCheck only those vertices that are in

the convex hull of the point set SRange query on kd-tree, refinement

step using ccw(p1, p2, p3) = (p2.x − p1.x) (p3.y − p1.y) − (p3.x − p1.x) (p2.y − p1.y)p0

p1pin

pout

21

Page 22: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

22

Page 23: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Fast Greedy AlgorithmRoad network is a metric spaceSum-of-distances function: almost

convex

23

Page 24: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Fast Greedy AlgorithmCompute the center of gravity of Q as (xc, yc)

NN query on vertex kd-tree with query point (xc, yc), find its NN vertex vnn

Initialize the current vertex as vnn

In each iteration, find the neighboring vertex vmin with minimum sum-of-distances valueIf the value is smaller than the current one, set the

current vertex as vmin

Otherwise, terminate

24

Page 25: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

25

Page 26: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsDatasets:

Road network datasets of 46 states in US

http://data.geocomm.com/catalog/USQuery points are randomly generated in

a rectangular query windowGiven OMP x and a result meeting point

x’, Ratio(x’) = [∑i dN(qi, x’) − ∑i dN(qi, x)] / ∑i dN(qi, x)

26

Page 27: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsEffect of window size

27

Page 28: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsEffect of window size

28

Page 29: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsEffect of |Q|

29

Page 30: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsEffect of |Q|

30

Page 31: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ExperimentsEffect of |Q|

31

Page 32: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

OutlineIntroductionRelated WorkBaseline AlgorithmConvex-Hull-Based PruningFast Greedy AlgorithmExperimentsConclusion

32

Page 33: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

ConclusionOur baseline algorithm substantially

reduce the search space of the OMP query from |Q| · |E| to |V| + |Q|

Our two-phase convex-hull-based pruning technique is accurate and effective

We develop an extremely efficient greedy algorithm to find a high-quality near-optimal meeting point

33

Page 34: Da Yan, Zhou Zhao and Wilfred Ng The Hong Kong University of Science and Technology

Thank you!

34