Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.

006 Fall 2011

Lecture 18: Shortest Paths IV - Speeding

• Bidirectional search

• Goal directed search - potentials and landmarks

Readings
Wagner, Dorothea, and Thomas Willhalm. ”Speed-up Techniques for Shortest-Path
Computations.” Proceedings of the 24th International Symposium on Theoretical
Aspects of Computer Science (STACS 2007): 23-36. Read up to Section 3.2.

DIJKSTRA single-source, single-target

Initialize()
Q ← V [G]
while Q = φ
do u ← EXTRACT MIN(Q) (stop if u = t!)
for each vertex v t Adj[u]
do RELAX(u, v, w)

Observation: If only shortest path from s to t is required, stop when t is removed

1
Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.006 Fall 2011

S t

backward search
forward search

Figure 1: Bi-directional Search Idea.

Bi-Directional Search
Note: Speedup techniques covered here do not change worst-case behavior, but reduce
the number of visited vertices in practice.

Bi-D Search

Alternate forward search from s

Algorithm terminates when some vertex w has been processed, i.e., deleted from the
queue of both searches, Qf and Qb
Subtlety: After search terminates, find node x with minimum value of df (x) + db (x).
x may not be the vertex w that caused termination as in example to the left!
Find shortest path from s to x using Πf and shortest path backwards from t to x
using Πb . Note: x will have been deleted from either Qf or Qb or both.

2
Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.006 Fall 2011

u u’
3

3
3

s t
5
5

Figure 2: Bi-D Search Example.

Minimum value for df (x) + db (x) over all vertices that have been processed in at least
one search (see Figure 3):

df (u) + db (u) = 3 + 6 = 9

df (u' ) + db (u' ) = 6 + 3 = 9
df (w) + db (w) = 5 + 5 = 10

3
Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.006 Fall 2011

Goal-Directed Search or A∗
Modify edge weights with potential function over vertices.

w (u, v) = w (u, v) − λ(u) + λ(v)

Search toward target as shown in Figure 4:

Correctness
w(p) = w(p) − λt (s) + λt (t)
So shortest paths are maintained in modified graph with w weights (see Figure 5).
To apply Dijkstra, we need w(u, v) ≥ 0 for all (u, v).
Choose potential function appropriately, to be feasible.

Landmarks
Small set of landmarks LCV . For all u ∈ V, l ∈ L, pre-compute δ(u, l).
(l)
Potential λt (u) = δ(u, l) − δ(t, l) for each l.
(l)
CLAIM: λt is feasible.

Feasibility

(l) (l)
w(u, v) = w(u, v) − λt (u) + λt (v)
= w(u, v) − δ(u, l) + δ(t, l) + δ(v, l) − δ(t, l)
= w(u, v) − δ(u, l) + δ(v, l) ≥ 0 by the Δ -inequality
(l)
λt (u) = max λt (u) is also feasible
lEL

4
Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.006 Fall 2011

u u’ Backward u u’
Forward 3 3

3 df (u) = 3
3 3 db(u’) = 3 3
s
s t
5 5 t db(t) = 0
df (s) = 0 5 5

w df (w) = 5
w db (w) = 5

df (u’) = 6 db(u’) = 3
u u’ u u’
Forward 3 Backward 3

3 3 3
3 db(u) = 6
df (u) = 3
s s t
t
5 5 db(t) = 0
df (s) = 0 5 5

df (w) = 5 w db (w) = 5
w

u u’ u u’ db(u’) = 3
Forward 3 Backward 3
df (u’) = 6 df (u’) = 6
3 3 3
3 df (u) = 3 df (u) = 3
db(u) = 6
s s t
t
5 5 db(t) = 0
5 df (t) = 10 df (s) = 0 5
df (s) = 0 df (t) = 10
db (s) = 10
df (w) = 5 w db (w) = 5
w
df (w) = 5
deleted from both queues
so terminate!

Figure 3: Forward and Backward Search and Termination.

5
Lecture 18 Shortest Paths IV: Speeding Up Dijkstra 6.006 Fall 2011

v’ v
5 5
increase decrease
go uphill go downhill

Figure 4: Targeted Search

p
s t

p’

Figure 5: Modifying Edge Weights.

6
MIT OpenCourseWare
http://ocw.mit.edu

6.006 Introduction to Algorithms

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.