Reasoning in Artificial Intelligence

1. Deductive reasoning
2. Inductive Reasoning:
3. Abductive reasoning
4. Common Sense Reasoning
5. Monotonic Reasoning
6. Non-monotonic Reasoning
Analogical reasoning
 Analogy

• At its most fundamental level, an analogy is a comparative relationship

between two (or more) different things or ideas.
• Analogies draw attention to how two unlike things are similar. As a cognitive
tool, they allow us to infer the properties or predict the behavior of
an unknown entity based on its similarities to a known entity. Analogical
reasoning is one of the most important cognitive tools we use to structure
our understanding of and beliefs about the world. Below are some of the
crucial contexts in which analogies function:

• Cognitive Development: Analogies play an important role in human cognitive

and linguistic development; children compare their knowledge
about past experiences to make predictions about present or future
experiences.

• Models and Maps: Mental and physical models are created for the express
purpose of illustrating analogies. Maps and models depend on their
similarities to and differences from whatever it is they are supposed to
represent. A map of Europe would not be practical if it was the size of the
continent itself; instead, a good map has consistent and accurate
spatial relationships between points of interest.
• Scientific Method: Scientists use inductive reasoning, which involves
recognizing analogies between things that are known and things that
are hypothesized but not yet known. If the unknown thing is similar in
important respects to a known thing, they make inferences about the
unknown thing based on its similarities to the known thing. The Italian
scientist Galileo used simple Archimedean machines like pendulums and
inclined planes as analogies for more complex phenomena like freefall.
While developing his laws of planetary motion, Johannes Kepler was
inspired by analogies from musical theory, like harmony, octave, and pitch.

• Literary Devices: Poets and storytellers use specific forms of analogy to

create vivid imagery. A metaphor is an analogy that states that one
thing is another thing to make some feature explicit. A simile is an analogy
that compares two things by stating that they are like each other.
• "The clock is a merciless master" is an example of a metaphor.
• "The sickness came like a thief" in the night is an example of a simile.

• Legal Precedent: When courts rule on a case, they set precedent. The ever-
growing body of legal precedent is used as a reference when courts are
presented with legal disputes. Present cases are compared to past cases,
and if a case is analogous to precedent, courts tend to rule as they ruled in
the past.
 Analogical Reasoning

• In informal terms, to reason by analogy is merely to compare two unlike

things and to make inferences based on the resulting analysis. Logicians and
philosophers, however, have a stricter definition of analogical reasoning:
reasoning by analogy means to use an analogical argument to form a
conclusion. An analogical argument is a formal argumentative form, the
structure of which is described in the following section.
• The primary difference between deductive arguments and analogical
arguments is that if a deductive argument is valid and sound, the truth of the
conclusion is guaranteed. Analogical arguments are called ampliative,
meaning that the conclusion may be false even if the argument is valid and
sound.
• Although analogical arguments are not truth-guaranteeing, they allow
the possibility of inference, which leads to new knowledge. For this
reason, analogical arguments are extremely important in domains like
science.
Structure of Analogical Reasoning
Suppose that a child graduates from elementary school and moves on to high
school. Although she doesn't know for certain what it will be like, she can
reason analogically to infer certain likelihoods about her upcoming experience.
Formally, her analogical argument might look as follows:
P. Both schools have teachers, a gymnasium, and classrooms.
P. The elementary school has a football field.
C. Therefore, the high school also has a football field.

Although the student can know with certainty that both premises are true, analogical
arguments do not guarantee the truth of their conclusions. A slightly modified version of
the example demonstrates how this is the case:

P. Both schools have teachers, a gymnasium, and classrooms.

P. The elementary school has cubbies for coats and lunchboxes.
C. Therefore, the high school has cubbies for coats and lunchboxes.

The above is a valid analogical argument, and its premises are true. But chances are,
the high school will have lockers instead of cubbies.

The relata of an analogical argument are called the source domain and the target
domain. In the example, the elementary school is the source domain and high school is
the target domain; known properties of the source domain are inferred about the target
domain based on known similarities.
Type of Reasoning
Solution
Structural Similarities
Recognition Process
Constraint Satisfaction
 Overview
• Constraint satisfaction problems (CSPs) need solutions that satisfy all
the associated constraints. Look into the definition and examples of
constraint satisfaction problems and understand the process of
converting problems to CSPs, using examples.
• Consider a Sudoku game with some numbers filled initially in some
squares. You are expected to fill the empty squares with numbers
ranging from 1 to 9 in such a way that no row, column or a block has a
number repeating itself. This is a very basic constraint satisfaction
problem. You are supposed to solve a problem keeping in mind some
constraints. The remaining squares that are to be filled are known as
variables, and the range of numbers (1-9) that can fill them is known as
a domain. Variables take on values from the domain. The conditions
governing how a variable will choose its domain are known as
constraints.
 Overview
A constraint satisfaction problem (CSP) is a problem that requires its solution
within some limitations or conditions also known as constraints. It consists of the
following:

•A finite set of variables which stores the solution (V = {V1, V2, V3,....., Vn})
•A set of discrete values known as domain from which the solution is picked (D =
{D1, D2, D3,.....,Dn})
•A finite set of constraints (C = {C1, C2, C3,......, Cn})

Please note, that the elements in the domain can be both continuous and discrete
but in AI, we generally only deal with discrete values.

Also, note that all these sets should be finite except for the domain set. Each
variable in the variable set can have different domains. For example, consider the
Sudoku problem again. Suppose that a row, column and block already have 3, 5
and 7 filled in. Then the domain for all the variables in that row, column and block
will be {1, 2, 4, 6, 8, 9}.
Popular Problems with CSP

The following problems are some of the popular problems that

can be solved using CSP:

1.CryptArithmetic (Coding alphabets to numbers.)

2.n-Queen (In an n-queen problem, n queens should be placed
in an nXn matrix such that no queen shares the same row,
column or diagonal.)
3.Map Coloring (coloring different regions of map, ensuring
no adjacent regions have the same color)
4.Crossword (everyday puzzles appearing in newspapers)
5.Sudoku (a number grid)
6.Latin Square Problem
• Algorithms for CSPs
– Backtracking (systematic search)
– Constraint propagation (k-consistency)
– Variable and value ordering heuristics
– Backjumping and dependency-directed backtracking
Motivating example: 8 Queens
Place 8 queens on a chess board such
That none is attacking another.

Generate-and-test, with no
redundancies → “only” 88 combinations

8**8 is 16,777,216
Motivating example: 8-Queens

After placing these two queens, it’s

trivial to make the squares we can
no longer use
 What more do we need for 8 queens?

• Not just a successor function and goal test

• But also
– a means to propagate constraints
imposed by one queen on the others
– an early failure test
→ Explicit representation of constraints and
constraint manipulation algorithms
 Informal definition of CSP
• CSP = Constraint Satisfaction Problem, given
(1) finite set of variables
(2) each with domain of possible values (often finite)
(3) set of constraints limiting values variables can take
on
• Solution is an assignment of a value to each variable
such that the constraints are all satisfied
• Tasks might be to decide if a solution exists, to find a
solution, to find all solutions, or to find “best solution”
according to some metric (objective function).
 Example: 8-Queens Problem
• Eight variables Xi, i = 1..8 where Xi is the row
number of queen in column i
• Domain for each variable {1,2,…,8}
• Constraints are of the forms:
–Not on same row:
Xi = k ➔ Xj  k for j = 1..8, ji
–Not on same diagonal
Xi = ki, Xj = kj ➔|i-j| | ki - kj| for j = 1..8, ji
 Example: Task Scheduling
T1

T2 T4

T3

Examples of scheduling constraints:

• T1 must be done during T3
• T2 must be achieved before T1 starts
• T2 must overlap with T3
• T4 must start after T1 is complete
 Example: Map coloring
Color the following map using three colors
(red, green, blue) such that no two adjacent
regions have the same color.

D A
B

C
 Map coloring
• Variables: A, B, C, D, E all of domain RGB
• Domains: RGB = {red, green, blue}
• Constraints: AB, AC,A  E, A  D, B  C, C 
D, D  E
• A solution: A=red, B=green, C=blue, D=green,
E=blue

E E
D A => D A
B B
C C
 Brute Force methods
• Finding a solution by a brute force solve(A,B,C,D,E) :-
color(A),
search is easy color(B),
– Generate and test is a weak method color(C),
color(D),
– Just generate potential combinations and color(E),
test each not(A=B),
not(A=B),
• Potentially very inefficient not(B=C),
not(A=C),
–With n variables where each can have one not(C=D),
of 3 values, there are 3n possible solutions not(A=E),
not(C=D).
to check
• There are ~190 countries in the world, color(red).
color(green).
which we can color using four colors color(blue).

• 4190 is a big number!

4**190 is 2462625387274654950767440006258975862817483704404090416746768337765357610718575663213391640930307227550414249394176L
 Example: SATisfiability
• Given a set of propositions containing
variables, find an assignment of the variables
to {false, true} that satisfies them.
• For example, the clauses:
– (A  B  C)  ( A  D)
– (equivalent to (C → A)  (B  D → A)
are satisfied by
A = false, B = true, C = false, D = false
• Satisfiability is known to be NP-complete,
so in the worst case, solving CSP problems
requires exponential time
 Real-world problems
CSPs are a good match for many practical problems that arise in
the real world

• Scheduling • Graph layout

• Temporal reasoning • Network management
• Building design • Natural language
• Planning processing
• Optimization/satisfaction • Molecular biology /
• Vision genomics
• VLSI design
Definition of a constraint network (CN)
A constraint network (CN) consists of
• a set of variables X = {x1, x2, … xn}
– each with associated domain of values {d1,d2,…dn}
– domains are typically finite
• a set of constraints {c1, c2 … cm} where
– each defines a predicate which is a relation over a
particular subset of variables (X)
– e.g., Ci involves variables {Xi1, Xi2, … Xik} and
defines the relation Ri  Di1 x Di2 x … Dik
Running example: coloring Australia
NT
Q
WA

SA NSW

• Seven variables: {WA,NT,SA,Q,NSW,V,T}

• Each variable has the same domain: {red, green, blue}
• No two adjacent variables have the same value:
WANT, WASA, NTSA, NTQ, SAQ, SANSW,
SAV,QNSW, NSWV
Unary & binary constraints most common
Binary constraints
T1
NT
Q
WA T2 T4
SA NSW
T3
V
T

• Two variables are adjacent or neighbors if

connected by an edge or an arc
• Possible to rewrite problems with higher-order
constraints as ones with just binary constraints
• Reification
 Formal definition of a CN
• Instantiations
–An instantiation of a subset of variables S
is an assignment of a value in its domain to
each variable in S
–An instantiation is legal iff it does not
violate any constraints
• A solution is an instantiation of all of the
variables in the network
 Typical tasks for CSP
• Solution related tasks:
–Does a solution exist?
–Find one solution
–Find all solutions
–Given a metric on solutions, find the best one
–Given a partial instantiation, do any of the
above
• Transform the CN into an equivalent CN
that is easier to solve
 Binary CSP
• A binary CSP is a CSP where all constraints
are binary or unary
• Any non-binary CSP can be converted into a
binary CSP by introducing additional variables
• A binary CSP can be represented as a
constraint graph, with a node for each
variable and an arc between two nodes iff
there’s a constraint involving the two variables
– Unary constraints appear as self-referential arcs
Running example: coloring Australia
NT
Q
WA

SA NSW

• Seven variables: {WA,NT,SA,Q,NSW,V,T}

• Each variable has the same domain: {red, green, blue}
• No two adjacent variables have the same value:
WANT, WASA, NTSA, NTQ, SAQ, SANSW,
SAV,QNSW, NSWV
A running example: coloring Australia
NT
Q
WA
SA NSW

V
T

• Solutions are complete and consistent assignments

• One of several solutions
• Note that for generality, constraints can be expressed
as relations, e.g., WA ≠ NT is
(WA,NT) in {(red,green), (red,blue), (green,red), (green,blue),
(blue,red),(blue,green)}
 Challenges for constraint reasoning
• What if not all constraints can be satisfied?
– Hard vs. soft constraints
– Degree of constraint satisfaction
– Cost of violating constraints
• What if constraints are of different forms?
– Symbolic constraints
– Numerical constraints [constraint solving]
– Temporal constraints
– Mixed constraints
 Challenges for constraint reasoning
• What if constraints are represented intensionally?
– Cost of evaluating constraints (time, memory,
resources)
• What if constraints, variables, and/or values change
over time?
– Dynamic constraint networks
– Temporal constraint networks
– Constraint repair
• What if multiple agents or systems are involved in
constraint satisfaction?
– Distributed CSPs
– Localization techniques
Metareasoning
• Metareasoning is a general AI term for “thinking about thinking” within a
computational system.

• While reasoning algorithms are used to make decisions, a metareasoning

algorithm is used to control a reasoning algorithm or to select among a set of
reasoning algorithms, determining which decision-making method should be
used under different circumstances.

• A classic example of metareasoning is to determine whether a reasoning

algorithm should stop or continue in a given context.

• Metareasoning can be described as in Fig. 1, where reasoning occurs at the

Object Level based on observations at the Ground Level, and the decisions
made at the Object Level are enacted at the Ground Level.

• For example, a sensored alarm might sound at the Ground Level when an
algorithm at the Object Level determines from sensor input that an intruder
was present (e.g., this algorithm may sound an alarm when two or more
motion events are detected within a 10-s window).
Fig.1 Classic decision‒action loop diagram of metareasoning, where
reasoning happens at the object level to select the actions that will happen
at the ground level, and metareasoning happens at the meta-level to control
what occurs at the object level
• Metareasoning then occurs when information from the Object Level is
observed and altered at the Meta-Level.

• In the previous example, an algorithm at the Meta-Level might adjust the

sensitivity of the alarm if it is triggered too often, causing battery issues
(e.g., this Meta-Level algorithm might impose a new algorithm at the Object
Level that sounds the alarm only when three or more motion events are
detected within a 10-s window).
 Metareasoning within a single agent and
multi-agent
• Metareasoning can occur within a single agent, as depicted in Fig. 1, or it can
occur within a multi-agent system (MAS) as you can see in figure (2-a).

• Metareasoning is often used in a multi-agent setting to optimize the

performance of an entire system, and there are many options for how it is
implemented with different consequences for resources such as time and
compute power.

• For example, agents within an MAS may perform their metareasoning

independently and communicate at the Object Level, which may be a good
solution when communication is costly and coordination is a low priority.

• When coordination is more important, independently metareasoning agents

may communicate at the Meta-Level to jointly determine how they will
independently metareason.
Fig.2(a) An MAS system where Fig.2(b) An MAS where each agent’s
metareasoning occurs independently metareasoning communicates and
for each agent. coordinates with each other agent’s
metareasoning.
 An MAS with a single centralized and multiple
separate metareasoning agents
• Metareasoning can also be performed in a more centralized fashion by separate
metareasoning agents (Fig.3-a). As communication resources allow, the best
coordination and metareasoning is expected to come from a single centralized
metareasoning agent (Fig. 3-b).

Fig.3-a An MAS with multiple separate Fig.3-b An MAS with a single

metareasoning agents centralized metareasoning agent.
• Systems also vary in the object of their metareasoning. Single agent metareasoning
is often used to control algorithm halting or switching and applied to a wide variety
of fields, including scheduling and planning, heuristic search, and object detection.

• Within MASs, metareasoning is often used to control communication and

resources within the systems, including controlling communication frequency or
content, or assigning tasks.

An additional concern in metareasoning is how much learning or metareasoning

should happen online versus offline. Because online metareasoning can be costly
in terms of time and computation, offline policies are often maximized to the
extent that they do not unduly impair system accuracy.

Broadly speaking, ToM (Theory of Mind) is a form of metareasoning, or

“thinking about thinking.” As described in Fig. 1, however, metareasoning
is performed through monitoring and controlling the Object Level,
whereas ToM involves making inferences from what is happening at the
Ground Level without direct access to the Object Level (e.g., an agent’s
beliefs).
• While metareasoning is already widely used in single- and multi-agent
systems to improve performance, ToM approaches arguably have not been
explored as deeply as a method to improve performance of an artificially
intelligent agent.

• This is almost certainly in part because ToM is more closely tied to human
cognition, which places strong restrictions on plausible ToM models and
biases research toward human applications.

• Additionally, ToM itself is still somewhat controversial (e.g., Who has it?
When is it acquired? Under what conditions is it exercised?) but it holds
promise for creating more-transparent (if not authentically human) systems,
especially systems reasoning with multiple sources of information and with
differing provenance and certainty.

• In particular, recent computational ToM approaches, which use simpler,

heuristic definitions of ToM, may be the best source of innovation in this
field.