JohannSearchIntel09 Shortest Path PDF

Shortest Path Search Algorithms with
Heuristic and Bidirectional Searches
Marcelo Johann
johann@inf.ufrgs.br
www.inf.ufrgs.br/~johann
Bidirectional Heuristic Shortest Path Search - Marcelo Johann - 2009 Visit to Intel: Slide 1
Outline
 Introduction
 Heuristic Search: properties of A*
 Bidirectional Search and the quest for Bi-A*
 LCS*
 principles, the algorithm, and how it performs
 Main Conclusions
 Pathological cases and open questions
 Application in grids and routing (mainly A*)
 6 ideas that we find significant…
Shortest Path Problem
In a locally finite graph G=(V,E), find the
shortest path between two designated nodes
source s and target t - smallest total cost.
v1 v5
v4 v6 t
s v3
v2 v8
v7
v9
Shortest Path Search

Starting at s, the method builds a search tree by
repetitive application of the successor operator
A node is expanded when we apply the successor
operation on it (the node becomes closed)
A node is generated when it is returned by the
successor operator (the node becomes open)
v1 v6
v4 v8 v0
v3
v2 s admissibility
v5 v9 v7 t
Breadth-First Search (BFS)
Expand first nodes that are closer to the source,
or at the same time nodes that are at the
same distance (FIFO).
Intermediate
search
target
source
Complete
search
First Shortest Path Search Methods

 BFS - Moore 1957 - Goes through all graph nodes,
starting with the ones closer to the source. Therefore, it
has: unit costs, FIFO structure, complexity O(n+|E|), for n
nodes.
 Shortest-path Search Algorithm - Dijkstra, 1959 - The
graph has arc costs and the algorithm seeks the path from
a source node to a target node. Complexity O(n2) or O(n
log n + |E|) with a priority queue.
 Maze Router - Lee, 1961 - An adaptation of BFS,
searching in a grid that represents a metal layer, by using
numbers from a sequence. After reaching the target, all
possible paths with the smallest cost can be identified
during the retrace phase. O (l2) for a circuit with side l.
The Maze Router
X X X X X X X X X X X X X X X X X X X X
X X
X X
X X
X X
X S X
X X
X X
X X
X X
X X X X X X X X X X X X X X X X
X X
X T X
X X
X X
X X
X X
X X
X X
The Maze Router

X X
X X
X X
X 2 X
X 2 1 2 X
X 2 X
X X
X X
X X
X X
X T X
X X
X X
X X
X X
X X
X X
The Maze Router
X X
X X
X 4 3 X
X 4 3 2 3 X
X 4 3 2 1 2 3 X
X 4 3 2 3 X
X 4 3 X
X X
X X
X X
X T X
X X
X X
X X
X X
X X
X X
The Maze Router

X X X X X X X
X X X X X X X X X X X X X
X 7 6 5 6 X
X 7 6 5 4
6 7 5 X
X 7 6 5 4 3
5 6 7 4 X
X 6 5 4 3 2 3
4 5 6 7 X
X 5 4 3 2 1 2
3 4 5 6 7 X
X 6 5 4 3 2 3
4 5 6 7 X
X 7 6 5 4 3
5 6 7 4 X
X 7 6 5 6 4 5 X
X 7 6 5 6 X
X X X X X
X X X X X X X X X X X
X X
X T X
X X
X X
X X
X X
X X
X X
The Maze Router
X X X X X X X X X
X 9 8 7 6 5 6 7 8X 9 10 11 12
X 8 7 6 5 4 5
12 6 7X 8 9 10 11
X 7 6 5 4 3 4
11 12 5 6X 7 8 9 10
X 6 5 4 3 2 3 4
10 11 12 5X 6 7 8 9
X 5 4 3 2 1 2 3
9 10 11 12 4X 5 6 7 8
X 6 5 4 3 2 3 4
10 11 12 5X 6 7 8 9
X 7 6 5 4 3 4
11 12 5 6X 7 8 9 10
X 8 7 6 5 4 5
12 6 7X 8 9 10 11
X 9 8 7 6 5 6 7 8X 9 10 11 12
X 10 9 X X X X X
X X X X XX X X X X
X 11 10 11 12 X
X 12 11 12 T X
X 12 X
X X
X X
X X
X X
X X
The Maze Router

X X X X X X X X X
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 5 4 3 2 1 2 3 4
13 14 X 5 6 7 8 9 10 11 12
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 10 9 X X X X X X X X X X X X X X X
X 11 10 11 12 13 14 15 16 X
X 12 11 12 13 14 15 16 T X
X 13 12 13 14 15 16 X
X 14 13 14 15 16 X
X 15 14 15 16 X
X 16 15 16 X
X 16 X
X X
The Maze Router
X X X X X X X X X
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 5 4 3 2 1 2 3 4
13 14 X 5 6 7 8 9 10 11 12
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 11 10 11 12 13 14 15 16 X
X 12 11 12 13 14 15 16 T X
X 13 12 13 14 15 16 X
X 14 13 14 15 16 X
X 15 14 15 16 X
X 16 15 16 X
X 16 X
X X
The Maze Router

X X X X X X X X X
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 5 4 3 2 1 2 3 4
13 14 X 5 6 7 8 9 10 11 12
X 6 5 4 3 2 3 4 5
14 15 X 6 7 8 9 10 11 12 13
X 7 6 5 4 3 4 5 6
15 16 X 7 8 9 10 11 12 13 14
X 8 7 6 5 4 5 6 16 7 X 8 9 10 11 12 13 14 15
X 9 8 7 6 5 6 7 8 X 9 10 11 12 13 14 15 16
X 11 10 11 12 13 14 15 16 X
X 12 11 12 13 14 15 16 T X
X 13 12 13 14 15 16 X
X 14 13 14 15 16 X
X 15 14 15 16 X
X 16 15 16 X
X 16 X
X X
Maze Router’s expansion g++
#include <iostream> void print(void);
#include <queue> int bfs(int source, int target)
using namespace std; {
#define SIDE 20 queue<int> q;
#define PLACE(x,y) ((y)*SIDE+(x)) q.push(source);
#define WEST(n) (n-1) while (!q.empty())
#define EAST(n) (n+1) {
#define NORTH(n) (n-SIDE) int n = q.front();
#define SOUTH(n) (n+SIDE) q.pop();
char Space[SIDE*SIDE]; if (n==target)
void init (void) return 1;
{ if (Space[n] != 'X')
for (int i=0; i<SIDE*SIDE; ++i) {
Space[i] = ' '; Space[n] = 'X';
for (int i=0; i<SIDE; ++i) print();
{ getchar();
Space[PLACE(i,0)] = 'X'; q.push(WEST(n));
Space[PLACE(i,SIDE-1)] = 'X'; q.push(EAST(n));
Space[PLACE(0,i)] = 'X'; q.push(NORTH(n));
Space[PLACE(SIDE-1,i)] = 'X'; q.push(SOUTH(n));
} }
for (int i=3; i<SIDE-3; ++i) }
Space[PLACE(i,10)] = 'X'; return 0;
} }
Outline
 Introduction
 LCS*
Heuristic Search (A*) [Hart et al. 1968]
Expand first more promissing nodes, according to:

f(n) = g(n) + h(n)
Also known as Best-First Search
Intermediate
stages
source target
g(n) h(n)
Complete
Search
Behaviour of f, g and h
1 2 3 4
g= g= g= g=
h= h= h= h=
f= c=1 f= c=2 f= c=2 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=0 g= g= g=
h= h=3 h= h= h=
f= c=1 f =3 c=2 f= c=2 f= c=2 f=
c=1 c=1 c=1 c=1
13 14 15 11
g= g= g= g=
h= h= h= h=
f= c=1 f= c=2 f= c=2 f=
Open List:
7
1 2 3 4
g=1 g= g= g=
h=4 h= h= h=
f =5 c=1 f= c=2 f= c=2 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g=1 g=0 g=2 g= g=
h=5 h=3 h=2 h= h=
f =5 c=1 f =3 c=2 f =4 c=2 f= c=2 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=1 g= g= g=
h=4 h= h= h=
f =5 c=1 f= c=2 f= c=2 f=
Open List:
8, 1, 13, 6
1 2 3 4
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g=1 g=0 g=2 g=4 g=
h=4 h=3 h=2 h=1 h=
f =5 c=1 f =3 c=2 f =4 c=2 f =5 c=2 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
Open List:
1, 13, 6, 9, 2, 14
1 2 3 4
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g=1 g=0 g=2 g=4 g=
h=4 h=3 h=2 h=1 h=
f =5 c=1 f =3 c=2 f =4 c=2 f =5 c=2 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
Open List:
1, 13, 6, 9, 2, 14
ties
1 2 3 4
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g=1 g=0 g=2 g=4 g=
h=4 h=3 h=2 h=1 h=
f =5 c=1 f =3 c=2 f =4 c=2 f =5 c=2 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=1 g=3 g= g=
h=4 h=3 h= h=
f =5 c=1 f =6 c=2 f= c=2 f=
Open List:
1, 13, 6, 9, 2, 14
But not critical as f * = 6
Quality of the estimator h
1 2 3
g=1 g=3 g=
h=4 h=3 h=
f =5 c=1 f =6 c=2 f=
Effect of the
c=1 c=1 c=1 estimation
6 7 8 9
g=1 g=0 g=2 g=4 efficiency
h=4 h=3 h=2 h=1
f =5 c=1 f =3 c=2 f =4 c=2 f =5
c=1 c=1 c=1
13 14 15
g=1 g=3 g=
h=4 h=3 h=
f =5 c=1 f =6 c=2 f=
The A* Algorithm
X X
X X
X X
X X
X 10 X
X X
X X
X
X
7 X
X
X X
X T X
X
X
3 X
X
X X
X X
X X
X X
h = estimation function
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X X
X X
X X
X T X
X X
X X
X X
X X
X X
X X
Min f = g + h Max g
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X 10 X
X 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Min f = g + h Max g
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Straingth path from S to T
The A* Algorithm
X X
X X
X X
X X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Try all possible straight paths
The A* Algorithm
X X
X X
X X
X 12 12 12 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X X
X T X
X X
X X
X X
X X
X X
X X
Maze router-like with obstacles
The A* Algorithm
X X
X 16 16 16 16 X
X 16 14 14 14 14 16 X
X 16 14 12 12 12 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 X X X X X X X X X X X X X X X
X 16 X
X 16 16 16 16 16 16 T X
X X
X X
X X
X X
X X
X X
Maze router-like with obstacles
The A* Algorithm
X X
X 16 16 16 16 X
X 16 14 14 14 14 16 X
X 16 14 12 12 12 12 14 16 Saved
X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 effort
X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 X
X 16 16 16 16 16 16 T X
X X
X X
X X
X X
X X
X X
Some Notation…
s, t - source and target nodes
k(n1,n2) - estimate to go from n1 to n2
k*(n1,n2) - actual cost of the shortest path from n1 to n2
h(n) = k(n,t)
h*(n) = k*(n,t)
g*(n) = k*(s,n)
Pa-b = path from a to b
Pa-b* = optimal path from a to b
f*(n) = cost of Ps-n* ∪ Pn-t* (going through n)
Properties of the Heuristic Estimator
 *): cost from n to t ≥ h(n)

Admissibility (*
ensures the shortest path h(n) t
n
Shortest path
from n to t
 Consistency: k(n1,n2) + k(n2,n3) ≥ k(n1,n3)

Every expanded node
n2
has optimal cost k(n1,n2 )
k(n2,n3 )
n3
known n1
k(n1,n3 )
g(n) = g*(n)
Properties of the A* Algorithm

 Continuity - the search has always an open node ∈ Ps-t*.
 Completeness - A* terminates.
 Admissibility - A* finds the optimal path.
 sufficient condition for expansion - f(n) < Ps-t* .
 Necessary condition for expansion - f(n) ≤ Ps-t* .
 Indeterminism on critical ties - biasing.
 Monotonicity - if n2 is expanded after n1 then
f(n2) ≥ f(n1 ).
 Dominance - A more informed A* expands less nodes.
 Excelence - A* is optimal among all unidirectional
algorithms that are no better informed, with a consistent
heuristic and no critical ties.
Outline
 Introduction
 LCS*
Bidirectional Search [Pohl 1971]

Two simultaneous search fronts
 Meeting node: recognized by both fronts
 Terminating condition: f(n) > min[f(m)]
 Leads to overlapping
Source search Target search
target
m
source
Meeting node
Unidirectional search
( to compare)
Bidirectional Search
Source search Target search
target
m
source
Meeting node
4 5
a b d
6
6 6 6
5 4
c e f
4 5
a b d
g=10 6
6 6 6
5 4
c e f
g=9
4 5
a b d
g=10 6
6 6 6
5 4
c e f
First Meeting
Heuristic Bidirectional Search
Goal: to combine the advantages of both
The challenge:
 Terminating condition
 Overlapping [Kwa 1989]
 The missing fronts problem
Goal The missing fronts

Bi-BFS
problem
Bi-A*
Wave-Shapping
One front looks at the other
Calculates distance to each node in the oposite front
f(n) = gs (n) + min[k(n,pi) + gt (pi)]
pi
t
s
gs(n) gt(pi)
n k(n,pi )
Requires quadratic time (or space)
Perimeter Search [Manzini 1995]
Two separate searches following one another

The first BFS and the second A*, usualy
First search, up to a
predetermined
perimeter pi
t
Second search with
front-to-front
gt(pi)
estimates
gs(n)
s k(n,pi )
n
It demonstrates the potencial of bidirectional heuristic

Outline
 Introduction
 LCS*
The LCS* Algorithm
Lowerbound Cooperative Search [ISA/CI2000]
Cooperation
 overall code is like BS* [Kwa 89]
 dynamic estimation (resistances e penalties)
 Function h of one front will be improved by the
other’s g values
 estimated values stored in references
completeness and admissibility proved [SBIA2000]
Conditions for expansion…
Sufficient Condition for Expansion

 A* expands all nodes with f(n) < Ps-t*
Necessary Condition for Expansion

 A* only expands nodes with f(n) ≤ Ps-t*
Missing fronts?
 No, this is not the main problem! [Kaindl 95,97]
The Pruning Power!
 The power of a heuristic admissible algorithm

is not in how fast it finds the goal but in how
efficiently it computes higher f(n) values for
the generated nodes.
Dynamic Estimation
 Resistence (min idea [Kaindl 96])
pi
Rt = Min[gt(pi) - k(pi,t)] gt(pi)
t
F(n) = f(n) + Rt k(pi ,t)
 A good idea for BFS, but

with A* the minimum values pi
are always at the opposite t

gt(pi)
side of the search k(pi ,t)
Dynamic Estimation
 Penalty (max idea [Kaindl 96])
Pt = Min[gt*(pi) + k(pi,s)]
pi
F(n) = gs(n) + Pt - ht(n)
gt(pi)
ht(pi) t
k(pi ,s)
ht(n) hs(n)
s
gs(n) n
 LCS* uses both, as penalties can

be smaller than resistances in
some cases
Cooperation and Visibility

Function h of one front will be improved by the
other’s g values
But h is under-estimated while g is over-estimated
In previous works:
 g and h were associated with the nodes
themselves
Visibility:
 estimated values stored in references
 single public closed set for both fronts allows
cooperation
LCS* is Admissible
Admissibility
 There are always references to nodes in the
optimal path in both search fronts;
 The algorithm terminates only after they meet;
Prunning Power
 Less nodes are expanded by increasing f(n) values
 Normal values are used in the open lists, but
single quantities Ω (resistance) and P (penalty)
are subtracted from Lmin (best cost so far)
LCS*’s code :
Too many details!

Ask me later on if you want…
Results: LCS* versus A* [Johann 2000]
Expanded nodes and actual runtime: typical
200 by 200 grids with random costs, mean 100,

1 minimum 50, maximum 300
0.8
ratio LCS*/A*
0.6 nodes
0.4 runtime
0.2
0
0
0
10
14
18
22
26
30
34
38
range of random costs
Specification of random costs
maximum
range
Center AC
(mean) costs
minimum
DC
cost
0
Expanded nodes in 2D grids
ε-Admissibility in A* and LCS*

 It is to find a path which is within ε from the
shortest one
 In heuristic search this is accomplished by
simply multiplying the original h function by
1+ε
 But in LCS*, you can also check how far
each meeting node is from a possible optimal
by comparing it to the smalest open f.
Testing ε-admissibility
LCS*/A* in 2D grids, regular cost regime
200 by 200 grids with random costs, mean 100,

3 minimum 50, maximum 300
2.5
ratio LCS*/A*
2
100%
1.5
96%
1
0.5
0
15 25 35 45 55 65 75 85 95
range of random costs
LCS* wins on harder mazes
100x100 random BFS, DFS and multi-DFS mazes
LCS* wins on random graphs
70 nodes, all 70x70 path problems
A* almost optimal in geometric graphs
70 nodes, all 70x70 path problems
Outline
 Introduction
 LCS*
Main Conclusions
 A* is way easier to implement. Always the
first choice.
 LCS* will usually lose when the area is
empty, as A* expands the smallest number
of nodes. (but it’s on easy instances)
 LCS* can get you a small constant
advantage, and can be significantly better
when the heuristic is bad, like in congested
routing, or when one of the pins (A* target)
is blocked.
Pathological cases
 When one of the pins is in a very congested
area, A* starting from this pin has an
advantage, for when the search leaves the
congested area, it goes direct to the target.
LCS* will always duplicate its efforts.
 Similar case happens with a barrier in the
middle. A* wins with 50% probability.
Open question
Direction choosing criteria:
 Alternating
 Smallest f
 Cardinality
 Inertial Cardinality
 Expanding from the most difficult first
But this contradicts others
Outline
 Introduction
 LCS*
Application in grids and routing
1. Do we need admissibility?
2. Degrees of freedom
3. Cost functions: add some processing
4. Optimizing data structures
5. How to avoid blocking pins
6. Grid properties and some ideas
1. Admissibility: do we need it?
 LCS* is a nice proved admissible algorithm
 But who cares?
 In most cases we don’t need full admissibility
 The advantage of LCS is that it can compute
ε-admissibility in two ways.
 The disadvantage is that if you don't need to
guarantee ε, then forget LCS* and make a
simple Bi-A* that terminates in the first
meeting.
2. Degrees of freedom
During expansion
On ties, use biasing, like in AMAZE
[Hentschke ISPD 2007]
And during retrace?

Need to record multiple pointers:
Uniform costs
1 2 3 4
g= g= g= g=
h= h= h= h=
f= c=1 f= c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g= g= g= g=
h= h= h= h= h=
f= c=1 f= c=1 f= c=1 f= c=1 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g= g= g=
h=4 h= h= h=
f =4 c=1 f= c=1 f= c=1 f=
Open List:
13
Critical Ties
1 2 3 4
g= g= g= g=
h= h= h= h=
f= c=1 f= c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g= g= g=
h= h=3 h= h= h=
f= c=1 f =4 c=1 f= c=1 f= c=1 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g= g=
h=4 h=3 h= h=
f =4 c=1 f =4 c=1 f= c=1 f=
Open List:
14,7
f(14) = f(7) = f *
Can choose what you want
1 2 3 4
g= g= g= g=
h= h= h= h=
f= c=1 f= c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g= g=
h= h=3 h=2 h= h=
f= c=1 f =4 c=1 f =4 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=
h=4 h=3 h=2 h=
f =4 c=1 f =4 c=1 f =4 c=1 f=
Open List:
15,8,7
Nodes close to the target first

1 2 3 4
g= g=3 g= g=
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=
h= h=3 h=2 h=1 h=
c=1 c=1 c=1 c=1
f= f =4 f =4 f =4 f=
worst
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
best
Open List:
11,9,8,7
Nodes close to the target first
1 2 3 4
g= g=3 g= g=
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=4
h= h=3 h=2 h=1 h=0
f= c=1 f =4 c=1 f =4 c=1 f =4 c=1 f =4
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
Open List:
15,7,2,3
Only four nodes expanded
But pay attention to:

1 2 3 4
g= g=3 g= g=
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=4
h= h=3 h=2 h=1 h=0
f= c=1 f =4 c=1 f =4 c=1 f =4 c=1 f =4
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
Open List:
15,7,2,3
Degree of freedom
1 2 3 4
g= g=3 g= g=
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=4
h= h=3 h=2 h=1 h=0
f= c=1 f =4 c=1 f =4 c=1 f =4 c=1 f =4
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
Open List:
15,7,2,3
Degree of freedom

1 2 3 4
g= g=3 g= g=
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=4
h= h=3 h=2 h=1 h=0
f= c=1 f =4 c=1 f =4 c=1 f =4 c=1 f =4
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
Open List:
15,7,2,3
Degree of freedom
1
g=
2
g=3
3
g=
4
g= options
h= h=3 h= h=
f= c=1 f =6 c=1 f= c=1 f=
c=1 c=1 c=1 c=1
6 7 8 9 10
g= g=1 g=2 g=3 g=4
h= h=3 h=2 h=1 h=0
f= c=1 f =4 c=1 f =4 c=1 f =4 c=1 f =4
c=1 c=1 c=1 c=1
13 14 15 11
g=0 g=1 g=2 g=3
h=4 h=3 h=2 h=1
f =4 c=1 f =4 c=1 f =4 c=1 f =4
Open List:
15,7,2,3
The A* Algorithm
X X
X X
X X
X X
X 10 X
X X
X X
X
X
7 X
X
X X
X T X
X
X
3 X
X
X X
X X
X X
X X
h = estimation function
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X X
X X
X X
X T X
X X
X X
X X
X X
X X
X X
Min f = g + h Max g
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X 10 X
X 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Min f = g + h Max g
The A* Algorithm
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Straingth path from S to T
The A* Algorithm
X X
X X
X X
X X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Biasing already considered

Degrees of freedom
The A* Algorithm
X X
X X
X X
X 12 12 12 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X 12 10 10 10 10 12 X
X X
X T X
X X
X X
X X
X X
X X
X X
But all this area will be expanded

anyway…
The A* Algorithm
X X
X 16 16 16 16 X
X 16 14 14 14 14 16 X
X 16 14 12 12 12 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 X
X 16 16 16 16 16 16 T X
X X
X X
X X
X X
X X
X X
After it escapes, A* again
The A* Algorithm
X X
X 16 16 16 16 X
X 16 14 14 14 14 16 X
X 16 14 12 12 12 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 14 12 10 10 10 10 12 14 16 X
X 16 X
X 16 16 16 16 16 16 T X
X X
X X
X X
X X
X X
X X
There are degrees of freedom in backtrace!
3. Cost Functions
 This is a delicate question for routing, but let's
just consider we may have different scenarios.
 So, things we can do:
 You can introduce noise (dither) in the cost
function to better randomize the paths. But
then you loose degrees of freedom.
 On the contrary, you can round the costs
down to low resolution, as to increase the
number of ties.
3. Cost Functions
Also, costs represent conflicting factors:
1. resource usage (how much metal this takes)

2. Path quality, = delay
3. Congestion
So they can be weighted for net, factor, layer

It helps taking better paths with less effort
A simple visual example…
A simple visual example…
4. Optimizing Data Structures

A* expands nodes with increasing values of f.
If the maximum cost per arc is known, we have
a limited range for f values at the same time in
the open list.
Therefore, we can implement it as an array of
lists.
Also, the ordering of each list can be calculated
only when a new f value is needed.
5. To avoid blocking pins
Typically the number of alternate paths is proportional to the
distance from a given pin.
So the probability of getting blocked is higher close to the pins.
We can minimize blocking by reserving some escapes for a pin.
But fixed reservations compromise freedom, and should be used
only when there is a single way out.
For multiple possible accesses, a technique of alternate
reservations was proposed by Johann/Reis in 1995:
 Place an alternate reservation in one position for each pin, maintaining a list of
other options to it. When drawing a connection during retrace, if this connection
passes on an alternate reservation, eliminate the reservation and check for the
other options. If there are more than one still available, chose one of then to hold
an alternate reservation. On the other hand, if there is only one way out remaining,
make it a fixed reservation. In both cases, no change is made to the path being
recorded in the current retrace.
6. Grid Properties
Let us look at this again…
X X
X X
X X
X X
X 10 X
X 10 X
X 10 X
X 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
6. Grid Properties
Its clear that this area could not reach t with cost f=10.
X X
X X
X X
X X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
6. Grid Properties
This is still an open problem…
X X
X X
X X
X X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X 10 10 10 10 X
X X
X T X
X X
X X
X X
X X
X X
X X
Shortest Path Search Algorithms with
Heuristic and Bidirectional Searches
y ou !
h an k
T
Marcelo Johann
johann@inf.ufrgs.br
www.inf.ufrgs.br/~johann

JohannSearchIntel09 Shortest Path PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

JohannSearchIntel09 Shortest Path PDF

Загружено:

Авторское право:

Доступные форматы

Shortest Path Search Algorithms with

Heuristic and Bidirectional Searches

Shortest Path Search

First Shortest Path Search Methods

The Maze Router

The Maze Router

The Maze Router

The Maze Router

Expand first more promissing nodes, according to:

Straingth path from S to T

Try all possible straight paths

Maze router-like with obstacles

Maze router-like with obstacles

 *): cost from n to t ≥ h(n)

 Consistency: k(n1,n2) + k(n2,n3) ≥ k(n1,n3)

Properties of the A* Algorithm

Bidirectional Search [Pohl 1971]

Source search Target search

Source search Target search

Goal The missing fronts

Requires quadratic time (or space)

Two separate searches following one another

It demonstrates the potencial of bidirectional heuristic

Conditions for expansion…

Sufficient Condition for Expansion

Necessary Condition for Expansion

 The power of a heuristic admissible algorithm

F(n) = f(n) + Rt k(pi ,t)

 A good idea for BFS, but

are always at the opposite t

side of the search k(pi ,t)

 LCS* uses both, as penalties can

Cooperation and Visibility

Too many details!

Expanded nodes and actual runtime: typical

200 by 200 grids with random costs, mean 100,

Specification of random costs

ε-Admissibility in A* and LCS*

LCS*/A* in 2D grids, regular cost regime

200 by 200 grids with random costs, mean 100,

LCS* wins on harder mazes

100x100 random BFS, DFS and multi-DFS mazes

70 nodes, all 70x70 path problems

A* almost optimal in geometric graphs

70 nodes, all 70x70 path problems

Application in grids and routing

And during retrace?

Nodes close to the target first

But pay attention to:

But pay attention to:

Straingth path from S to T

Biasing already considered

But all this area will be expanded

After it escapes, A* again

There are degrees of freedom in backtrace!

1. resource usage (how much metal this takes)

So they can be weighted for net, factor, layer

A simple visual example…

4. Optimizing Data Structures

Вам также может понравиться

LCS/A in 2D grids, regular cost regime