Problem Solving and Search

Geraint A. Wiggins Professor of Computational Creativity Queen Mary, University of London

What is problem solving?

An agent can act by
establishing goals that result in solving a problem considering sequences of actions that might achieve those goals

To automate this, we need to formulate the problem The process of considering action sequences within a problem
formulation is called search
problem-solving agents problem types problem formulation example problems search strategies

This part of the module is about

An Agent on Holiday
Example
An agent is in Arad, Romania, at the end of a touring holiday. It has a ticket to y out of
Bucharest the following day. The ticket is non-refundable, the agents visa is about to expire and there are no seats available on ights for six weeks.

The agents performance measure includes

cost of ticket avoidance of arrest and/or deportation improvement of suntan seeing more sights

Lots of actions are possible to score highly on the performance

Agent can reject any actions that fail to meet this goal

measure, but the seriousness of the agents situation means that the agent should adopt the goal of getting to Bucharest

A problem-solving agent
Goals help to organise behaviour
goal formulation is the rst step to problem solving a goal is a desired state of (part of) the world
though it does not usually specify the state of the whole world alternatively, a goal can be thought of as the set of world states in which it is satised

Actions cause transitions between world states

The agent has to work out which action(s) will get it to the goal state

Problem formulation is exhaustive analysis of possible actions and states The agent needs to consider possible actions and the resulting states
put left foot forward 18 inches?? get cab to next town??

A problem-solving agent

Our holiday agent has adopted the goal of driving to Bucharest There are three roads out of Arad: to Sibiu, Timisoara and Zerind
None of these reaches the goal. Which road should the agent follow?

Without additional knowledge about Romania, the agent cannot choose

the best road to take: it can only make a random choice get to and which actions it can perform

However, if the agent has a map of Romania, it knows which states it can

Exercise: route planning on a map

Exercise
Oradea Neamt Zerind Iasi Arad Sibiu Fagaras Vaslui Timisoara Rimnicu Vilcea Pitesti Hirsova Mehadia Urziceni Bucharest Dobreta Craiova Giurgiu Eforie

Lugoj

a route from Arad to Bucharest. Find Find a route from Aradto [Link] this the optimal route? How did you go the optimal the route? Is thisabout ndingroute? How did you go about nding the route?
ELE611 Articial Intelligence Department of Electronic Engineering Problem Solving & Search 6

A Problem-Solving Agent

The agent can use the map information

consider subsequent stages of a hypothetical journey through each of the 3
nearest towns, to try to nd a route that eventually leads to Bucharest

Once it has found a path on the map from Arad to Bucharest, it can
achieve its goal locally by driving to each town on the route agent can decide what to do by searching
utility (goal states)

More generally, with several immediate options of unknown value, the

assessing different possible sequences of actions that lead to states of known choosing the best one

Design Process
To design a problem-solving program
FORMULATE:
Formulate problem, goal and starting conditions SEARCH:
Use search algorithm to nd action sequence that realises goal
from starting conditions

EXECUTE:

Use solution to guide actions

To design a problem-solving agent

Do the above in the context of a changing world

THE ART OF AI PROBLEM SOLVING IS DECIDING WHAT GOES INTO THE DESCRIPTION OF THE OPERATORS AND STATES, AND WHAT IS LEFT OUT

Search Algorithms

A search algorithm takes a problem as input, formulated as

a set of goals to be achieved a set of possible actions to be applied

It returns a solution in the form of an action sequence or plan Once a solution is found, the chosen actions can be carried out
the execution phase

In the agent context, this becomes complicated because of interaction

between agents and other dynamic things in the world
so we consider the search on its own rst

Well-dened problems

Basic problem elements are states and actions Formally, we need

initial state: the world the agent starts in operators: describe actions in terms of states reached by applying them goal state: the world that the agent is trying to be in

The initial state and operators dene the state space of the problem To determine when the goal state is reached, we need
goal test: a boolean function over states, true if input is the goal state

the set of all states reachable from the initial state and any sequence of actions

Searching for solutions

A solution is a sequence of operators leading from start to goal state Given a problem formulation, we have to map out the state space
while we are searching the state space its sometimes called the search space

Generate action sequences, using an agenda:

start at initial state check whether in goal state
if yes, were done

otherwise, consider other states by applying the operators to the present state
to generate a new set of states
This is called expanding the state

Search Trees


Its helpful to think of the state space as structured into a tree
actions correspond with arcs Choosing states and actions I states correspond with nodes
Sibiu
Oradea Neamt Zerind Arad Sibiu Fagaras

Arad

Timisoara

Zerind

Arad
Iasi

Rimnicu Vilcea Fagaras Oradea

Vaslui

Timisoara

Rimnicu Vilcea Pitesti Hirsova Mehadia Urziceni Bucharest

Lugoj

Note that we dont allow loops

Dobreta Craiova Giurgiu Eforie

Search Trees


Searching this way takes us to Bucharest in 3 steps
but only because of the order we picked the cities in
Arad

Choosing states and actions I

Oradea Neamt

Sibiu Arad
Iasi Arad Sibiu Fagaras

Timisoara

Zerind

Zerind

Rimnicu Vilcea Fagaras Oradea

Vaslui

Timisoara

Rimnicu Vilcea Pitesti

Sibiu

Bucharest

Lugoj

Hirsova Mehadia Urziceni Bucharest Dobreta Craiova Giurgiu Eforie

Clear thinking: state node!

States of the world and nodes in the search tree are not the same thing
A state is a (representation of) a problem conguration A node is a data structure used by a search algorithm
constitutes part of a search tree includes parent, action, children, depth (for IDS), path cost (for UCS)

States do not formally have parents, actions, children, depth, or path cost!

The Expand function creates new nodes, lling in the various elds and
using the operators of the problem formulation to approximate the corresponding states

Search Algorithm Framework

function AgendaSearch( start-state, actions, goal-test ) returns action-sequence or failure static: seq, an action sequence, initially empty agenda, a state sequence, initially contains start-state loop do if Empty(agenda) then return failure if goal-test(First(agenda)) then return seq agenda Queuing-Fn(Rest(agenda), Expand(First(agenda), actions)) seq Append( seq, Action(First(agenda))) end loop

The Queuing-Fn function determines what kind of search we are doing The Expand function produces a set of states achieved by applying all possible
actions to the given state

This is generally called Agenda-based search

Search & the Queuing-Fn

Searching this way takes us to Bucharest in 3 steps
but only because of the order we picked the cities in
Arad Timisoara Zerind

Choosing states and actions I

Oradea Neamt

Sibiu

Zerind Iasi Arad Sibiu Fagaras Vaslui Timisoara Rimnicu Vilcea Pitesti Hirsova Mehadia Urziceni Bucharest Dobreta Craiova Giurgiu Eforie

Lugoj

if wed chosen rightmost order, we would have gone via Timisoara, Dobreta and Pitesti wasting time and fuel

Search & the Queuing-Fn

This is an example of Depth-First Search (DFS)
we construct the search tree by going down it, only going across it when going
down fails

depth-rst search can fail when the tree is innitely deep

the space contains cycles or is generated from continuous values

depth-rst search depends (crucially) on the order in which operators are

applied/nodes are expanded
it may not nd things that are present it may nd long action sequences before short ones (usually, thats unhelpful)

In terms of agendas, we implement DFS with a Queuing-Fn that

removes the current node from the agenda appends its children to the front of the agenda note that the agenda is rotating all the time, always dealing with its rst item

Search & the Queuing-Fn

An alternative is Breadth-First Search (BFS)
we construct the search tree by going across it, repeatedly, replacing each node
with its children

breadth-rst search guarantees to nd the solution if there is one breadth-rst search guarantees to nd the shortest action sequence but it causes a combinatorial explosion of nodes in the agenda, and so is
unsuitable for very large search spaces
most real AI problems have very large search spaces

In terms of agendas, we implement BFS with a Queuing-Fn that

removes the current node from the agenda appends its children to the back of the agenda note that the agenda is rotating all the time, always dealing with its rst item

Breadth-rst search in Romania

Here the different depths are given different colours
we can see the different layers of the tree as the search proceeds Choosing states and actions I and its clear that the shorted solution, if there is one, will be found
Oradea Neamt Zerind Iasi Arad Sibiu Fagaras Vaslui Timisoara Rimnicu Vilcea Pitesti Hirsova Mehadia Urziceni Bucharest Dobreta Craiova Giurgiu Eforie

Lugoj

Measuring search effectiveness

To determine the utility of a particular path, we need
cost: a function that assigns a cost to a path (to differentiate between solutions)

To evaluate a search strategy or algorithm, need to ask

does it nd a solution? is it a good solution (one with a low path cost)? is it the best solution (optimal)? how much time and memory are required to nd a solution?

Effectiveness of search depends on the objectives. Compare

searching for a route to Bucharest (many routes may be satisfactory) searching for the next prime number (must be a true prime, and theres only 1)

Measuring search effectiveness

Now consider the efciency of the planned route
going via Sibiu and Fagaras has the fewest towns Exercise: Arad to Bucharest using UCS but going via Sibiu Pitesti is shorter
71
Oradea Neamt

75
Arad

Zerind

151

87
Iasi

140
Sibiu

92 99
Fagaras Vaslui Rimnicu Vilcea

118 80
Timisoara

111 70

Lugoj

97 146

Pitesti

211

142 98
Urziceni

Mehadia

101 138

85

Hirsova

75
Dobreta

86
Bucharest

120
Craiova

90
Giurgiu Eforie

Straight!line distance to Bucharest Arad 366 Bucharest 0 Craiova 160 Dobreta 242 Eforie 161 Fagaras 178 Giurgiu 77 Hirsova 151 Iasi 226 Lugoj 244 Mehadia 241 Neamt 234 Oradea 380 Pitesti 98 Rimnicu Vilcea 193 Sibiu 253 Timisoara 329 Urziceni 80 Vaslui 199 Zerind 374

Measuring search effectiveness

We can compute the utility of a route by a function which implements
Exercise: Arad to Bucharest using UCS
here, we might use distance for each road
Oradea Neamt

the performance measure, in terms of the problem formulation

71

75
Arad

Zerind

151

87
Iasi

140
Sibiu

92 99
Fagaras Vaslui Rimnicu Vilcea

118 80
Timisoara

111 70

Lugoj

97 146

Pitesti

211

142 98
Urziceni

Mehadia

101 138

85

Hirsova

75
Dobreta

86
Bucharest

120
Craiova

90
Giurgiu Eforie

Straight!line distance to Bucharest Arad 366 Bucharest 0 Craiova 160 Dobreta 242 Eforie 161 Fagaras 178 Giurgiu 77 Hirsova 151 Iasi 226 Lugoj 244 Mehadia 241 Neamt 234 Oradea 380 Pitesti 98 Rimnicu Vilcea 193 Sibiu 253 Timisoara 329 Urziceni 80 Vaslui 199 Zerind 374

Search & the Queuing-Fn

We can implement this using Uniform-Cost Search (UCS)
Also called Best-First Search

we construct the search tree by picking the best node in the agenda after each
expansion, according to the utility function, expanding it, and so on

best-rst search guarantees to nd the solution if there is one best-rst search guarantees to nd the best action sequence in terms of utility but it cannot guarantee to use less memory than BFS
most real AI problems have very large search spaces

In terms of agendas, implement UCS/BestFS with a Queuing-Fn that

removes the current node from the agenda appends its children to the agenda and then sorts the whole agenda by the utility
function
and stores the cost-so-far at each node in the agenda, to avoid recalculation

note that the agenda is rotating all the time, always dealing with its rst item

Best-rst search (UCS) in Romania

Using UCS to consider real distances
Here BFS would get us wrong answer Exercise: Arada to Bucharest using UCS
146 71 75
75
Arad Zerind Oradea Neamt

UCS is correct, but spends time searching routes which are obviously wrong
Straight!line distance to Bucharest Arad 366 Bucharest 0 Craiova 160 Dobreta 242 Eforie 161 Fagaras 178 Giurgiu 77 Hirsova 151 Iasi 226 Lugoj 244 Mehadia 241 Neamt 234 Oradea 380 Pitesti 98 Rimnicu Vilcea 193 Sibiu 253 Timisoara 329 Urziceni 80 Vaslui 199 Zerind 374

151

87
Iasi

140
Sibiu

118
Timisoara

99

239

92
Vaslui

Fagaras

140 229
111 70 299 75
Dobreta

80

118

160

Rimnicu Vilcea

257
97 146 138 120
Craiova Pitesti

211

142

Lugoj

450
Mehadia

444

101

85

98
Urziceni

Hirsova

358 Bucharest
90

86

374

306

Giurgiu

Eforie

Formulating the problem

Now lets look at the full specication of this problem
State space: cities (20) Goal test: in Bucharest?

Initial state: in Arad Exercise: Arad to Bucharest using U Operator: drive between two cities
71
Oradea Neamt

Strai to Bu
87
Iasi

Path cost: one of

75
Arad

Zerind

151

140
Sibiu

total milage total expected travel time step count in search

92 99
Fagaras Vaslui Rimnicu Vilcea

118 80
Timisoara

111 70

Lugoj

97 146

Pitesti

211

142 98
Urziceni

Many details are left out

abstraction

Mehadia

101 138

85

Hirsova

75
Dobreta

86
Bucharest

120
Craiova

90
Giurgiu Eforie

Arad Buch Crai Dob Efor Faga Giur Hirs Iasi Lugo Meh Neam Orad Pites Rim Sibiu Tim Urzi Vasl Zeri

Analysing search spaces

A strategy is dened by its order of node expansion Uninformed strategies use only information in the problem formulation
Sometimes called blind search or brute-force search
Breadth-rst search Depth-rst search Best-rst search/Uniform-cost search (this feels informed, but isnt, by this denition)

Informed or heuristic search strategies a quality measure not strictly in the

problem formulation to guide the order in which nodes are expanded
e.g., An informed agent might notice that the goal state, Bucharest, is southeast
Algorithm A and Algorithm A*

of Arad, and that only Sibiu is in that direction, so it is likely to be the best choice

Analysing search spaces for cost

A strategy is evaluated along the following dimensions
completeness
time complexity
space complexity
optimality

does it always nd a solution if one exists? number of nodes generated maximum number of nodes in memory at once does it always nd a least-cost solution?

Time and space complexity are measured in terms of the search tree
b
maximum branching factor of the search tree (bushiness of tree) d
depth of the least-cost solution m
maximum depth of the state space (may be )

Even the best strategies need to avoid repeated nodes

trade-off between time and space (see Russell and Norvig)

Analysing search algorithms for cost

So far, weve looked at Depth rst, Breadth rst, and Uniform Cost
search in practice
now, the theory

Analysing search algorithms for cost

Depth rst search (DFS)

Complete? No, DFS fails in innite depth search spaces and spaces with loops Time? O(bm): terrible if m is much larger than d, but if search tree is bushy, may
be much faster than breadth rst

Space? O(bm), i.e., space linear in length of action sequence! Need only store a
single path from the root to the leaf node, along with remaining unexpanded sibling nodes for each node on path

Optimal? No Pros: For problems with many solutions, DFS may be faster than BFS because it Cons: can get stuck down wrong search path

has a good chance of nding a solution after exploring only a small portion of the whole space

Analysing search algorithms for cost

Breadth rst search Properties of breadth rst search
Complete? Yes, all nodes are examined if b is nite Complete: if b is nite. Time? 1 + bYes, 2all nodes. .are examinedd), i.e., exponential in b + b + b3 + . + bd = O(b
Time: 1 + b + b2 + b3 + . . . + bd = O(bd), i.e., exponential in b. Space? O(bd) keeps every node in memory Space: O(bd) (keeps every node in memory). Optimal? Only if cost = 1 per step, not in general Optimal: Yes (if cost = 1 per step); not optimal in general. If there are several If there are several solutions it will nd the shallowest goal rst solutions it will nd the shallowest goal rst. Drawbacks: Memory requirement to store nodes areaabig problem Drawbacks: Memory requirements to store nodes is big problem.
Assumingbb== 10, expansion1000 nodes/sec, and 100 bytes/node: Assuming 10, expansion of of 1000 nodes/sec, and 100 bytes/node:
Depth Nodes Time Memory 0 1 1 millisecond 100 bytes 2 111 0.1 seconds 11 kilobytes 4 11,111 11 seconds 1 megabyte 6 6 10 18 minutes 111 megabytes 8 108 31 hours 11 gigabytes 10 10 10 128 days 1 terabyte 12 12 10 35 years 111 terabytes 14 14 10 3500 years 11,111 terabytes

Analysing search algorithms for cost

Uniform-cost (Best rst) search

Shares dynamic properties of DFS and BFS Requires step cost information in problem formation and path cost in agenda Time? Number of nodes with path cost less than optimal solution Space? Number of nodes with path cost less than optimal solution Optimal? Yes nds the cheapest solution as long as step cost > 0 First solution reported (but not necessarily encountered) is always the cheapest
does not explore the whole tree and avoids looping but can explore the same path multiple times if there is more than one route through a node

Depth bounded search: IDS

Iterative-deepening search is a compromise between DFS and BFS

Do DFS, but include a depth bound, forcing BFS-like behaviour In the agenda-based algorithm, we keep track of the depth of each node in the
agenda, then
Start with depth-cutoff = 1 Search, expanding only nodes with depth < depth-cutoff if unsuccessful, increment depth and repeat

Exercise: IDS on holiday

Execute IDS for the Romanian holiday problem, on paper

71
Oradea Neamt

Exercise: Arad to Bucharest using UCS

Arad Bucharest Craiova Dobreta Eforie Fagaras Giurgiu Hirsova Iasi Lugoj Mehadia Neamt Oradea Pitesti Rimnicu Vilcea Sibiu Timisoara Urziceni Vaslui Zerind

Straight!line distanc to Bucharest

87
Iasi

75
Arad

Zerind

151

140
Sibiu

92 99
Fagaras Vaslui Rimnicu Vilcea

118 80
Timisoara

111 70

Lugoj

97 146

Pitesti

211

142 98
Urziceni

Mehadia

101 138

85

Hirsova

75
Dobreta

86
Bucharest

120
Craiova

90
Giurgiu Eforie

Draw the search tree created by UCS. How many nodes are expanded

Analysing search algorithms for cost

Iterative Deepening Search

Complete? Yes Time? (d+1)b0 +db1 +(d1)b2 +...+ bd = O(bd) Space? O(bd) Optimal? Yes, if step cost = 1 Pros: combines the benets of BFS and DFS
optimal and complete (like BFS) modest memory requirements (like DFS) can be modied to explore uniform-cost tree

Cons: some (perhaps many) states are expanded multiple times

Although this appears wasteful, the overhead is actually quite small for most cases

Informed Search
Informed search methods use problem specic information about the state
space to generate solutions more effectively
Sometimes called Best First Search

Extra knowledge is provided by a heuristic which returns a number

describing the desirability of expanding a given node
an approximation since we usually dont know best node to expand nodes expanded in order of desirability (like UCS) most desirable processed rst

The aim is to nd low cost solutions

typically evaluate estimated cost of solution and try to minimise it measure incorporates estimate of cost from a state to the closest goal state
try to expand the node closest to the goal try to expand the node on the least cost solution path

Greedy, or hill-climbing, search

The simplest kind of informed search estimates the cost from the current
UCS Exercise: Arad to Bucharest using UCS
note the difference with UCS, where we use the actual cost so far
Oradea 71 Oradea

node to the solution and uses that to sort its agenda Exercise: Arad to Bucharest using

Straight!line distance Straight!line distance to Bucharest to Bucharest Neamt Arad 366 71 Neamt Arad 366 0 Bucharest 87 Zerind 151 Bucharest 0 160 Craiova 87 75 Zerind 151 Craiova 160 242 Iasi 75 Dobreta Iasi Dobreta 242 161 Arad Eforie 140 92 Arad Eforie 161 178 140 Fagaras Sibiu 92 Fagaras 99 Fagaras 178 77 Giurgiu Sibiu Fagaras 118 99 Giurgiu 77 151 Hirsova Vaslui 118 80 Hirsova 151 226 Vaslui Iasi 80 Rimnicu Vilcea Iasi Lugoj 226 244 Timisoara Rimnicu Vilcea Lugoj 244 241 Timisoara Mehadia 142 211 Mehadia 111 241 234 Neamt 142 Pitesti 97 Lugoj 211 111 Neamt 234 380 Oradea Pitesti 97 Lugoj 70 98 Oradea 380 98 Pitesti Hirsova 85 70 146 98 Pitesti 101 Rimnicu Vilcea 98 193 Mehadia Urziceni Hirsova 85 146 101 193 253 Sibiu 86 Rimnicu Vilcea Mehadia Urziceni 75 138 Bucharest Sibiu 86 253 329 Timisoara 75 138 120 Bucharest Dobreta Timisoara 329 80 Urziceni 90 120 Dobreta Urziceni 80 199 Vaslui Craiova 90 Eforie Giurgiu Vaslui Craiova 199 374 Zerind Eforie Giurgiu Zerind 374

Greedy, or hill-climbing, search

Most greedy search has an agenda of length 1
all but the best node is thrown away, at every cycle but it can be implemented with an agenda of any length the shorter the agenda, the more likely the search is to get caught in local
minima, where there is a temporary dip in the solution quality
throwing away solutions means you cant recover if you go astray

In the route-nding problem, a good heuristic is the straight-line distance

to the goal
as the crow ies this idea, that a geometrical distance estimates a distance along a graph, is the basis
of many heuristics

But because it doesnt account for cost so far, it can lead us astray

Algorithm A
Algorithm A combines the best of UCS and Greedy search
UCS had cost so far Greedy had estimated cost remaining or heuristic

Algorithm A has cost so far + estimated cost remaining

This is usually notated as f(n) = g(n) + h(n)

f is the overall cost/evaluation/utility function g is the cost-so-far function h is the heuristic

Algorithm A

is complete but can be sub-optimal in node expansion

Algorithm A*
Algorithm A* uses a heuristic which is admissible
this means that it never over-estimates the distance to the closest solution in any graph representation of space, geometric distance has this property

A* is optimal
in the sense that we are guaranteed a solution if there is one in the sense that the fewest possible nodes will be expanded, given this problem
formulation
NB re-formulating may produce better search

Exercise
Using the data on the earlier slide, apply A* to the Arad-Bucharest route problem
in what order are the nodes expanded?

Analysing search algorithms for cost

Proving optimality of A*
Suppose there is a suboptimal goal G2 in the agenda A (standard proof ) Optimality of
Let n be an unexpanded node on a shortest path to an optimal goal G1 but h(G2) = 0, as its a goal

f(G2) = g(G2)+ h(G2) Suppose some suboptimal goal G2 has been generated and is in the g(G2) > g(G1)
because G2 is sub-optimal
n Start

n be an unexpanded node on a shortest path to an optimal goal G1

g(G1) f(n)
because h is admissible

G2

Because f(G2) f(n), A* will never select G2 for expansion

2 2 2 The more accurate h is, the fewer unnecessary nodes will be expanded

f (G ) = g(G ) > g(G1) f (n)

since h(G ) = 0 since G2 is suboptimal since h is admissible

Analysing search algorithms for cost

One way to characterise the quality of a heuristic is the effective
branching factor (b*)
if
the total number of nodes expanded by A* for a particular problem is N the solution depth is d

then
b* is the branching factor that a uniform tree of depth d would have to have in order to contain N nodes

N =1+b*+(b*)2 +...+(b*)d
e.g., if A* nds a goal at depth 5 using 52 nodes, the effective branching factor is 1.91

Usually the effective branching factor exhibited by a given heuristic is

fairly constant over a large range of problems
This can be used for empirical evaluation of a heuristics usefulness

Exercise: The 8s-Puzzle

Choosing a heuristic function for a more abstract problem

the 8s-puzzle
typical solution 20 steps branching factor 3 (2 if space in corner, 4 in middle, 3 otherwise) exhaustive search to depth 20 examines 320 3.5 109 nodes

A* Heuristic must never over-estimate number of steps to goal

h1 = number of tiles in wrong position h2 = sum of Manhattan/city-block distances of tiles from their goals

Choosing a heuristic for 8s-puzzle

d 2 4 6 8 10 12 14 16 18 20 22 24 Search Cost IDS A*(h1 ) A*(h2 ) 10 6 6 112 13 12 680 20 18 6384 39 25 47127 93 39 364404 227 73 3473941 539 113 1301 211 3056 363 7276 676 - 18094 1219 - 39135 1641 Eective Branching Factor IDS A*(h1 ) A*(h2 ) 2.45 1.79 1.79 2.87 1.48 1.45 2.73 1.34 1.30 2.80 1.33 1.24 2.79 1.38 1.22 2.78 1.42 1.24 2.83 1.44 1.23 1.45 1.25 1.46 1.26 1.47 1.27 1.48 1.28 1.48 1.26

Table comparing the search costs and eective branching factors for IDS and A* using h1 and h2 . Data are averaged over 100 instances of the 8-puzzle, for various solution lengths.

better than h1 which is which better than uninformed search h2 is Clearly, h is better than h betteristhan uninformed search Is h always better than h ? AlwaysYes. A*heuristicwill expand fewer nodesso long as it doesnt over use a using h with higher values, than A* using h .
2 1 2 1 2 1

estimate use a heuristic with higher values, as long as it doesnt overestimate. Always

Further Reading

In Chapters 3 & 4 of the textbook you can read more about search
including
Bidirectional search Searching with partial information Hill-climbing search and simulated annealing Genetic algorithms Online search agents