COMP 424 - Artificial Intelligence

Informed Search

Instructor: Jackie CK Cheung (jcheung@[Link])

Readings: R&N Ch 3.5-3.7 (3rd ed); 3.5 – 3.6 (4th ed)
 Announcements
• Midterm scheduled for Feb 18, 2022
• Online MyCourses Quiz; you’ll have 48 hours to complete;
designed to take ~80 minutes
• Final projects to be done in pairs
• Syllabus & office hours posted on Ed

Upcoming:
• Quiz 1 will be posted by end of week
• Study guide information and first SG (you will have plenty of time
to find a group and organize them)
• A1 will be posted early next week

COMP-424: Artificial intelligence 2

 Overview of today’s class
• Using heuristics to inform (guide) the search

• Informed search algorithms

• Best-First (Greedy) Search
• Heuristic Search
• A* Search

• Learning heuristic functions

COMP-424: Artificial intelligence 3

 Eight-Puzzle

How would you solve an eight-puzzle? Do you really perform

an uninformed search as we discussed last class, or do
some configurations seem more promising than others?

COMP-424: Artificial intelligence 4

 Informed Search
• Uninformed search methods expand nodes based on
distance from the start node, 𝑑(𝑛𝑠𝑡𝑎𝑟𝑡 , 𝑛)
Easy: we always know that!
• What about expanding based on distance to the goal,
𝑑(𝑛, 𝑛𝑔𝑜𝑎𝑙)?
• If we know 𝑑(𝑛, 𝑛𝑔𝑜𝑎𝑙) exactly?
Easy: just expand the nodes that minimize 𝑑(𝑛, 𝑛𝑔𝑜𝑎𝑙 )
• If we do not know 𝑑(𝑛, 𝑛𝑔𝑜𝑎𝑙) exactly, can we use intuition
about this distance?
• We will call this intuition a heuristic 𝒉 𝒏 .

COMP-424: Artificial intelligence 5

 Example
• We are very close to the goal:

COMP-424: Artificial intelligence 6

 Example: Travelling through Romania
What is good heuristic for path planning problems?

COMP-424: Artificial intelligence 7

 Example heuristic: path planning
• What is a reasonable heuristic?
• The straight-line distance between two places.

• Is it always right?
• Clearly not - roads are rarely straight!

• But it’s roughly correct.

COMP-424: Artificial intelligence 8

 Example heuristic: eight-puzzle
• What would be a good heuristic for this problem?

COMP-424: Artificial intelligence 9

 Example heuristic: eight-puzzle
• Two options: ℎ1(𝑛) = number of misplaced tiles = 7
ℎ2(𝑛) = total Manhattan distance
= (2+3+3+2+4+2+0+2) = 18
• Which is better?

COMP-424: Artificial intelligence 10

 Example heuristic: eight-puzzle
• Two options: ℎ1(𝑛) = number of misplaced tiles = 7
ℎ2(𝑛) = total Manhattan distance
= (2+3+3+2+4+2+0+2) = 18
• Which is better? Intuitively, ℎ2(𝑛) seems better : varies more
across state space and its estimate is closer to the true cost.

COMP-424: Artificial intelligence 11

 Where do heuristics come from?
Possible strategies:
• Prior knowledge about the problem

• Exact solution to a relaxed version of the problem

• E.g., If rules of eight-puzzle are changed to allow a tile to move
anywhere. Then ℎ1(𝑛) gives the exact distance to goal.
• E.g., If rules of eight-puzzle are changed to allow a tile to move to
any adjacent square. Then ℎ2(𝑛) gives the exact distance to goal.

• Learning from prior experience

• More on this later

COMP-424: Artificial intelligence 12

 Best-First Search
• Algorithm: expand the most promising node according to
the heuristic.

• Example:

• What is the sequence of nodes visited in Best-First Search?

• What is the sequence of nodes visited in Uniform-Cost Search?

COMP-424: Artificial intelligence 13

 Best-First Search
• Algorithm: expand the most promising node according to
the heuristic.

• Example:

• What is the sequence of nodes visited in Best-First Search?

START -> A -> C -> GOAL
• What is the sequence of nodes visited in Uniform-Cost Search?
START -> A -> B -> C -> GOAL

COMP-424: Artificial intelligence 14

 Best-First Search
• Algorithm: expand the most promising node according to
the heuristic.

• If the heuristic is always 0, and we break ties by preferring

shallowest depth, best-first search is the same as breadth-
first-search.
• Best-first search contrasts with uniform-cost search:
Uniform cost search considers the cost-so-far.
Best-first search considers the cost-to-go.
• Best-first search is a greedy method.
maximize short-term advantage without
worrying about long-term consequences.

COMP-424: Artificial intelligence 15

 Properties of Best-First Search
• In worst case: exponential time/space complexity (same as
Breadth-First S).
O(bd) (b=branching factor, d=solution depth).
• A good heuristic can help a lot!
O(bd) space complexity (same as DFS).

• Completeness:
• Not always complete. Can get stuck in loops.
• Complete in finite spaces, if we check to avoid repeated states.

• Not optimal! (as seen in the example.) => Can we fix this?

COMP-424: Artificial intelligence 16

 Fixing Best-First Search
• What’s the problem with best-first search?
• It is too greedy: does not take into account the cost so far!

• Let’s use both of the following:

• Let 𝑔 be the cost of the path so far
• Let ℎ be a heuristic function (estimated cost to go)

COMP-424: Artificial intelligence 17

 Fixing Best-First Search
• What’s the problem with best-first search?
• It is too greedy: does not take into account the cost so far!

• Let’s use both of the following:

• Let 𝑔 be the cost of the path so far
• Let ℎ be a heuristic function (estimated cost to go)

• Heuristic search: best-first search, greedy with respect to

𝑓 =𝑔+ℎ
• Think of 𝑓 = 𝑔 + ℎ as the estimate of the cost of the current path.

COMP-424: Artificial intelligence 18

 Heuristic search
• Algorithm:
• At every step, take node 𝑛 from the front of the queue.
• Add to queue successors 𝑛’ with priorities: 𝑓(𝑛’) = 𝑔(𝑛’) + ℎ(𝑛’)
• Terminate when a goal state is popped from the queue.

• Previous example:

COMP-424: Artificial intelligence 19

 Example

Priority queue: {(S,h(S))}

COMP-424: Artificial intelligence 20

 Example

Priority queue: {(S,h(S))}

{(S,h(S)), (B,4), (A,7)}
{(S,h(S)), (B,4), (C,5), (A,7)}
{(S,h(S)), (B,4), (C,5), (A,7), (G,9)}
{(S,h(S)), (B,4), (C,5), (A,7), (G,8), (G,9)} Found!

COMP-424: Artificial intelligence 21

 Example

Priority queue: {(S,h(S))}

{(S,h(S)), (B,4), (A,7)}
{(S,h(S)), (B,4), (C,5), (A,7)}
{(S,h(S)), (B,4), (C,5), (A,7), (G,9)}
{(S,h(S)), (B,4), (C,5), (A,7), (G,8), (G,9)} Found!

Remark: Continue expanding nodes after goal is found if there

are unexpanded nodes that have lower cost than current path
to goal.
COMP-424: Artificial intelligence 22
 Is heuristic search optimal?
• Example:

h(A)=?

COMP-424: Artificial intelligence 23

 Is heuristic search optimal?
• Example:

h(A)=?

• Short answer: NO!

• In this example, any heuristic ℎ(𝐴) > 2 leads to suboptimal
path (and if ℎ(𝐴) = 2, it depends on tie breaking).
We must put conditions on the choice of heuristic to guarantee
optimality.

COMP-424: Artificial intelligence 24

 Admissible heuristics
• Let ℎ∗ (𝑛) be the shortest path from n to any goal state.
• A heuristic h is called an admissible heuristic if:
ℎ 𝑛 ≤ ℎ∗ 𝑛 , ∀𝑛.
• Admissible heuristics are optimistic.

• Trivial case of an admissible heuristic: ℎ 𝑛 = 0, ∀𝑛.

• In this case, we get uniform-cost search.

• If ℎ is optimistic, we must have ℎ(𝑠) = 0, 𝑠𝐺, the goal

set.

COMP-424: Artificial intelligence 25

 Examples of admissible heuristics
• Robot navigation?

• 8-puzzle?

COMP-424: Artificial intelligence 26

 Examples of admissible heuristics
• Robot navigation? straight-line distance to goal.

• 8-puzzle? h1: number of misplaced tiles.

h2: sum of Manhattan distances for each tile
to its goal position.

• In general, if we get a heuristic by solving a relaxed

version of the problem, we will obtain an admissible
heuristic.

COMP-424: Artificial intelligence 27

 A* search
• Algorithm: Heuristic search with an admissible heuristic.
• Let g be the cost of the path so far.
• Let h be an admissible heuristic function.
• Perform greedy search with respect to:
𝑓 =𝑔+ℎ

COMP-424: Artificial intelligence 28

 Consistency
• An admissible heuristic h is called consistent (or
monotone*) if for every state s and successor s’:
h(s)  c(s,s’) + h(s’).

• Think of h as getting “more precise” as it gets closer to the goal.

• This is a version of triangle inequality*, so heuristics that

respect this inequality are metrics*.
• Consistency is a slightly stronger property than admissibility.
*Note: Look-up these terms
if they are not familiar.

COMP-424: Artificial intelligence 29

 Is A* complete?
• Suppose that the state space graph is finite, h is
consistent, and all costs c(s,s’)>0.
• This implies that f cannot decrease along any path:
f(s) = g(s) + h(s)  g(s) + c(s,s’) + h(s’) = f(s’)
• If a solution exists, it must have bounded cost.
• In this case, A* cannot get stuck in cycles, because the f-score of a
re-expanded node will keep going up, eventually exceeding the
solution’s.
• Hence, A* will have to find the solution! So, it is complete.

Note:
• Non-decreasing path costs and c(s,s’)>0 mean that it is never
necessary to re-expand a node, as the first time it is encountered
by A* already gives you a lowest cost path to it.
COMP-424: Artificial intelligence 30
 Fixing inconsistent heuristics

• Make a small change to A*:

Instead of 𝑓(𝑠’) = 𝑔(𝑠’) + ℎ(𝑠’), use
𝑓(𝑠’) = max{ 𝑔(𝑠’) + ℎ(𝑠’), 𝑓(𝑠) }

• With this change, 𝑓 is non-decreasing along any path, and

the previous argument applies.

COMP-424: Artificial intelligence 31

 Is A* optimal?
• Proof by contradiction. Suppose a suboptimal goal G2 is in
queue.
• Let n be an unexpanded node on a shortest path to optimal
goal G1. start
• We have:
f(G2) = g(G2) since h(G2)=0 n
> g(G1) since G2 is suboptimal
G1 G2
≥ f(n) since h is admissible

• Since f(G2) > f(n), A* will select n for expansion before G2.
• Similarly, all nodes on the optimal path will be chosen before
G2, so G1 will be reached before G2.
COMP-424: Artificial intelligence 32
 Dominance
• If ℎ2(𝑛) ≥ ℎ1(𝑛), for all 𝑛 (both admissible), then ℎ2
dominates ℎ1
Intuition: ℎ2(𝑛) is more informative than ℎ1(𝑛)

• Eight-puzzle example:
• d=14 Iterative depth search = 3,473,941 nodes
A*(h1) = 539 nodes
A*(h2) = 113 nodes

• d=24 Iterative depth search = too many nodes

A*(h1) = 39,135 nodes
A*(h2) = 1,641 nodes

COMP-424: Artificial intelligence 33

 Properties of A*
• Complete!
• Optimal!

• Exponential worst-case time/space complexity.

• But with a perfect heuristic, complexity is O(bd), because we only
expand nodes on optimal path.
• With a good heuristic, complexity is often sub-exponential.

• Optimally efficient: with a given ℎ, no other optimal

search algorithm is guaranteed to expand fewer nodes
than A*.
• Technicality: up to tie-breaking

COMP-424: Artificial intelligence 34

 Iterative Deepening A* (IDA*)
• Same trick as we used before to avoid memory problems.
• Algorithm is basically depth-first search, but using the f-
value to decide in which order to consider the
descendants of a node.
• Use an f-value limit, rather than a depth limit.

COMP-424: Artificial intelligence 35

 Iterative Deepening A* (IDA*)
• Same trick as we used before to avoid memory problems.
• Algorithm is basically depth-first search, but using the f-
value to decide in which order to consider the
descendants of a node.
• Use an f-value limit, rather than a depth limit.

• IDA* has the same properties as A* but uses less memory.

• In order to avoid always expanding new nodes, old ones
can be remembered if memory permits (this version is
known as Simplified Memory-bounded A*, SMA*)

COMP-424: Artificial intelligence 36

 Iterative Deepening A* example

• Set f1 = 4  only S, B are searched

(no other nodes are put in the
queue, because they exceed the
cutoff threshold.)

• Set f2 = 8  now S, A, B, C, G, are

all searched.

COMP-424: Artificial intelligence 37

 Learning Heuristic Functions
• Can we automatically construct a better heuristic function
for eight-puzzles?
• Idea: use previous experience of playing other eight-
puzzles to help us learn ℎ.
• Need:
• Features that describe a particular state configuration
𝑥1 (𝑛), 𝑥2 (𝑛), …
• Form of heuristic function to be learned
ℎ(𝑛) = 𝑐1 𝑥1 (𝑛) + 𝑐2 𝑥2 (𝑛)
• Target for the heuristic function

COMP-424: Artificial intelligence 38

 Example
1

• Actual cost to goal is 2.

• Let 𝑥1 (𝑛) be number of misplaced tiles (= 2)
• 𝑥2 𝑛 be number of pairs of incorrectly adjacent tiles
• Pick 𝑐1 and 𝑐2 s.t. ℎ 𝑛 = 𝑐1 𝑥1 𝑛 + 𝑐2 𝑥2 𝑛 = 2
• Actually, we would try to fit an entire dataset of plays, not just
this particular case.

COMP-424: Artificial intelligence 39

 Summary of informed search
• Insight: use knowledge about the problem, in the form of
a heuristic.
• The heuristic is a guess for the remaining cost to the goal.
• A good heuristic can reduce search time from exponential to
almost linear.
• Best-first search is greedy with respect to the heuristic,
not complete and not optimal.
• Heuristic search is greedy with respect to f=g+h, where g
is the cost so far and h is the estimated cost to go.
• A* is heuristic search where h is an admissible heuristic.
• A* is complete and optimal.
• A* is a key AI algorithm

COMP-424: Artificial intelligence 40