Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed...

14
1 Artificial Intelligence, Spring, 2009 Informed Search & Exploration Instructor: Tsung-Che Chiang [email protected] Department of Computer Science and Information Engineering National Taiwan Normal University 2 Artificial Intelligence, Spring, 2009 Outline Informed Search Strategies Heuristic Functions Local Search & Optimization Problems Online Search Agents and Unknown Environments Summary

Transcript of Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed...

Page 1: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

1

Artificial Intelligence, Spring, 2009

Informed Search &Exploration

Instructor: Tsung-Che [email protected]

Department of Computer Science and Information EngineeringNational Taiwan Normal University

2Artificial Intelligence, Spring, 2009

Outline

Informed Search StrategiesHeuristic FunctionsLocal Search & Optimization ProblemsOnline Search Agents and Unknown

EnvironmentsSummary

Page 2: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

2

3Artificial Intelligence, Spring, 2009

Informed Search Strategies

Can we select the node from the fringe in amore efficient way?

Artificial Intelligence: A Modern Approach, 2nd ed., Figure 3.19

4Artificial Intelligence, Spring, 2009

Informed Search Strategies

Best-first search selects the node forexpansion according to an evaluationfunction f(n).

There is a whole family of best-firstsearch algorithms with different evaluationfunctions.

A key component of them is a heuristicfunction h(n).

h(n) = estimated cost of the cheapest pathfrom node n to a goal node

Page 3: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

3

5Artificial Intelligence, Spring, 2009

Informed Search Strategies

Example of a heuristic functionThe straight line distance

6Artificial Intelligence, Spring, 2009

Informed Search Strategies

Heuristic functions are the most commonform to impart additional knowledge.

There is one constraint:for a goal node n, h(n) = 0.

Page 4: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

4

7Artificial Intelligence, Spring, 2009

Greedy Best-first Search

Greedy best-first search expands the nodethat is closest to the goal, namely usingf(n) = h(n).

Assume we use the straight-line distanceas the heuristic function.

Exercise: Apply the greedy best-first search.

8Artificial Intelligence, Spring, 2009

Greedy Best-first Search

253

176

Artificial Intelligence: A Modern Approach, 2nd ed., Figure 4.2

366

h(n), straight-line distance

Page 5: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

5

9Artificial Intelligence, Spring, 2009

Greedy Best-first Search

The solution is not optimal.Notice that the name reveals that it just tries

to get as close to the goal as it can.

10Artificial Intelligence, Spring, 2009

Greedy Best-first Search

ProblemsFrom Iasi to Fagaras

Dead end?Infinite loop?

Page 6: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

6

11Artificial Intelligence, Spring, 2009

Greedy Best-first Search

Analysis It is neither optimal nor complete.The worst-case time and space complexity is

O(bm), where m is the maximum depth of thesearch space.

The complexity can be reduced substantially,and the amount of reduction depends on theproblem and the quality of the heuristic.

12Artificial Intelligence, Spring, 2009

A* Search

It is the most well-known form of best-first search.

It evaluates the nodes by

f(n) = g(n) + h(n), where

g(n) = the cost to reach the node h(n) = the estimated cost from the node to the

goal

f(n) = estimated cost of thecheapest solution through n

Page 7: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

7

13Artificial Intelligence, Spring, 2009

A* Search

Exercise

14Artificial Intelligence, Spring, 2009

A* Search

366=366+0

393=140+253

415=239+176 413=220+193

417=317+100

Artificial Intelligence: A Modern Approach, 2nd ed., Figure 4.3

1

23

4

5

0

Page 8: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

8

15Artificial Intelligence, Spring, 2009

A* Search

A* using TREE-SEARCH is optimal if h(n) isadmissible. It requires that h(n) never overestimates the

cost to reach the goal. Consequently, f(n) never overestimates the true

cost of a solution through n.

In our route-finding example, the straight-line distance is an admissible heuristic.

16Artificial Intelligence, Spring, 2009

A* Search

Proof of optimalityG2 is a suboptimal goal node appearing on the

fringe. n is a fringe node on an optimal path.C* is the cost of the optimal solution.

f(G2) = g(G2) + h(G2) = g(G2) > C*f(n) = g(n) + h(n) C* < f(G2)

G2 will not be expanded and A*must return an optimal solution.

Page 9: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

9

17Artificial Intelligence, Spring, 2009

A* Search

What if we use GRAPH-SEARCH?There are two ways to keep the optimality.

(1) Extend GRAPH-SEARCH so that it discardsthe more expensive of any two paths found tothe same node. The extra bookkeeping is messy.

(2) Ensure that the optimal path to anyrepeated state is always the FIRST onefollowed. An extra requirement of h(n) –consistency

(monotonicity)

18Artificial Intelligence, Spring, 2009

A* Search

A* using GRAPH-SEARCH is optimal if h(n)is consistent. n: node n: successor generated from n by action a c(n, a, n): cost from n to n

h(n) c(n, a, n) + h(n)

Page 10: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

10

19Artificial Intelligence, Spring, 2009

A* Search

Every consistent heuristic function is alsoadmissible.

It is quite hard to find heuristics that areadmissible but not consistent.

The straight-line distance heuristic isconsistent.

h(n) c(n, a, n) + h(n)

n

goal

n

h(n)

h(n)

c(n, a, n)

20Artificial Intelligence, Spring, 2009

A* Search

If h(n) is consistent, then the values of f(n) alongany path are nondecreasing.

f(n) = g(n) + h(n)= g(n) + c(n,a,n) + h(n)g(n) + h(n) = f(n)

It follows that A* using GRAPH-SEARCH expandsthe nodes in nondecreasing order of f(n). Hence,the first goal being expanded must be an optimalsolution.

Page 11: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

11

21Artificial Intelligence, Spring, 2009

A* Search

The fact that f-cost are nondecreasing along anypath means that we can draw contours in the statespace.

Artificial Intelligence: A Modern Approach, 2nd ed., Figure 4.4

All nodes inside thecontour labeled 420 havef(n) be at most 420.

380

400

420

22Artificial Intelligence, Spring, 2009

A* Search

With uniform-cost search (A* using h(n) =0), the bands will be “circular”around thestart state.

With more accurate heuristics, the bandbecome more narrowly focused around theoptimal path.

A* expands all nodes with f(n) < C*. It might expand some of the nodes right on the

contour with f(n) = C* before selecting a goalnode.

Page 12: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

12

23Artificial Intelligence, Spring, 2009

A* Search

Intuitively, the first solution found mustbe an optimal one.Nodes in all subsequent contours will have

higher f-cost and thus higher g-cost (h=0).

It is also intuitive that A* search iscomplete. It requires that there is only finitely many

nodes with cost no greater than C*. step cost exceeding some finite finite b

24Artificial Intelligence, Spring, 2009

A* Search

A* is optimally efficient for any heuristicfunction.No optimal algorithm can guarantee to expand

fewer nodes than A*.This is because any algorithm that does not

expand all nodes with f(n) < C* runs the risk ofmissing the optimal solution.

A* is complete, optimal, and optimallyefficient. What else can you ask for?

Page 13: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

13

25Artificial Intelligence, Spring, 2009

A* Search

Complexity For most problems, the number of nodes within

the goal contour is still exponential in thelength of the solution,

unless …

))((log)()( ** nhOnhnh h*(n) is the true costfrom n to the goal.

26Artificial Intelligence, Spring, 2009

A* Search

Space requirement is more serious thanthe time requirement.

It’s impractical to insist on the optimality.One can use heuristic functions that are more

accurate but not admissible.

Next, we will discuss algorithms thatovercome the space problem withoutsacrificing the optimality.

Page 14: Informed Search & Explorationweb.ntnu.edu.tw/~tcchiang/ai/3_Informed_search-1.pdf · Informed Search Strategies Can we select the node from the fringe in a more efficient way? Artificial

14

27Artificial Intelligence, Spring, 2009

Demonstration

BFS: http://www.youtube.com/watch?v=NpeEWq09mMw

DFS: http://www.youtube.com/watch?v=Xjppe7LfvnQ

A* Algorithm with Python-Pygame:http://www.youtube.com/watch?v=FNRfSQDF7TA

A-Star Pathfind Test:http://www.youtube.com/watch?v=vxX5A3-I88I

A* Personality-guided Pathfinding :http://www.youtube.com/watch?v=GA5vhBkH29I