Chapter 7: Logical Agents

Dr. Daisy Tang

In which we design agents that can form representations of the world,

use a process of inference to derive new representations about the
world, and use these new representations to deduce what to do.
Motivation of Knowledge-Based Agents
 Humans know things and do reasoning, which are important
for artificial agents

 Knowledge-based agents benefit from knowledge expressed in

general forms, combining information to suit various purposes

 Knowledge and reasoning is important when dealing with

partially observable environments

 Understanding natural language requires inferring hidden

states

 Flexibility for accepting new tasks

CS 420: Artificial Intelligence 2

Knowledge Bases

 Knowledge base = set of sentences in a formal language

 Declarative approach to building an agent (or other system):

 Tell it what it needs to know
 Then it can Ask itself what to do - answers should follow from the
KB

 Both “Tell” and “Ask” involve inference – deriving new

sentences from old

CS 420: Artificial Intelligence 3

A Simple Knowledge-Based Agent

 Details hidden in three functions

 Similar to agent with internal state in Chapter 2

 Agents can be viewed at the knowledge level

i.e., what they know, regardless of how they are implemented

 Or at the implementation level

 i.e., data structures in KB and algorithms that manipulate them

CS 420: Artificial Intelligence 4

Agent’s Requirements
 The agent must be able to:
 Represent states, actions, etc.
 Incorporate new percepts
 Update internal representations of the world
 Deduce hidden properties of the world
 Deduce appropriate actions

CS 420: Artificial Intelligence 5

Our New Toy Problem –
Wumpus World

CS 420: Artificial Intelligence 6

Wumpus World PEAS
 Performance measure
 gold +1000, death -1000
 -1 per step, -10 for using the arrow

 Environment
 4x4 grid of rooms
 Agent always starts at [1,1]
 Wumpus and gold randomly chosen
 Each square other than start is pit
with 0.2 probability

CS 420: Artificial Intelligence 7

Wumpus World PEAS
 Actuators: Left turn, Right turn,
Forward, Grab, Release, Shoot, Climb
 Shooting kills wumpus if you are facing it
 Shooting uses up the only arrow
 Grabbing picks up gold if in same square
 Releasing drops the gold in same square
 Climbing out of the cave from [1, 1]

 Sensors (percepts): Stench, Breeze,

Glitter, Bump, Scream
 Squares adjacent to wumpus are smelly
 Squares adjacent to pit are breezy
 Glitter iff gold is in the same square
 Bump when agent walks into a wall
 When wumpus is killed, it screams

CS 420: Artificial Intelligence 8

Wumpus World Properties
 Fully observable?
 No – only local perception
 Deterministic?
 Yes – outcomes exactly specified
 Episodic?
 No – sequential at the level of actions
 Static?
 Yes – Wumpus and Pits do not move
 Discrete?
 Yes
 Single-agent?
 Yes – Wumpus is essentially a natural feature

CS 420: Artificial Intelligence 9

 Exploring a Wumpus World
Agent’s initial knowledge base contains the rules of the environment: it knows that
it’s at [1,1] and it’s safe at [1,1]

percept: [none, none, none, none, none]

CS 420: Artificial Intelligence 10

Exploring a Wumpus World

percept: [none, breeze, none, none, none]

CS 420: Artificial Intelligence 11

Exploring a Wumpus World

percept: [none, breeze, none, none, none]

CS 420: Artificial Intelligence 12

Exploring a Wumpus World

our cautious agent will go back and to [2,1]

CS 420: Artificial Intelligence 13

Exploring a Wumpus World

percept: [stench, none, none, none, none]

CS 420: Artificial Intelligence 14

Exploring a Wumpus World

percept: [none, none, none, none, none]

CS 420: Artificial Intelligence 15

Exploring a Wumpus World

CS 420: Artificial Intelligence 16

Exploring a Wumpus World

percept: [stench, breeze, glitter, none, none]

CS 420: Artificial Intelligence 17

Logical Reasoning
 In each case where the agent draws a
conclusion from the available information,
that conclusion is guaranteed to be correct
if the available information is correct

 The above is the fundamental of logical

reasoning

CS 420: Artificial Intelligence 18

Logic in General
 Logics are formal languages for representing
information such that conclusions can be drawn
 Syntax defines the sentences in the language
 Semantics define the "meaning" of sentences;
 i.e., define truth of a sentence in a world

 E.g., the language of arithmetic

 x+2 ≥ y is a sentence; x2+y > {} is not a sentence
 x+2 ≥ y is true iff the number x+2 is no less than the
number y
 x+2 ≥ y is true in a world where x = 7, y = 1
 x+2 ≥ y is false in a world where x = 0, y = 6

CS 420: Artificial Intelligence 19

Entailment
 Now, we have a notion of truth, we are ready to
study logical reasoning, which involves logical
entailment between sentences
 Entailment means that one thing follows from
another:
KB ╞ α

 Knowledge base KB entails sentence α if and only if

α is true in all worlds where KB is true
 E.g., the KB containing “the Giants won” and “the Reds
won” entails “Either the Giants won or the Reds won”
 E.g., x+y = 4 entails 4 = x+y
 Entailment is a relationship between sentences (i.e.,
syntax) that is based on semantics

CS 420: Artificial Intelligence 20

Models
 We use the term “model” to represent “possible world”
 We say m is a model of a sentence α if α is true in model m

 M(α) is the set of all models of α

 Then KB ╞ α iff M(KB)  M(α)

 E.g. KB = Giants won and Reds
won, α = Giants won

CS 420: Artificial Intelligence 21

Entailment in the Wumpus World
 Situation after detecting
nothing in [1,1], moving
right, breeze in [2,1]

 Consider possible models

for KB assuming only pits

 3 Boolean choices  8
possible models

CS 420: Artificial Intelligence 22

Wumpus Models

CS 420: Artificial Intelligence 23

Wumpus Models

 KB = wumpus-world rules + observations

CS 420: Artificial Intelligence 24

Wumpus Models

 KB = wumpus-world rules + observations

 α1 = "[1,2] is safe", KB ╞ α1, proved by model checking
 enumerate all possible models to check that α is true in all models in
which KB is true
CS 420: Artificial Intelligence 25
Wumpus Models

 KB = wumpus-world rules + observations

CS 420: Artificial Intelligence 26

Wumpus Models

 KB = wumpus-world rules + observations

 α2 = "[2,2] is safe“, KB ╞ α2?
CS 420: Artificial Intelligence 27
Wumpus Models

 KB = wumpus-world rules + observations

 α2 = "[2,2] is safe", KB ╞ α2
CS 420: Artificial Intelligence 28
Soundness and Completeness
 An inference algorithm that derives only
entailed sentences is called sound

 An inference algorithm is complete if it can

derive any sentence that is entailed

 If KB is true in the real world, then any

sentence derived from KB by a sound
inference procedure is also true in the real
world

CS 420: Artificial Intelligence 29

Grounding Problem
 Grounding: how do we know that KB is true
in the real world?
 Agent’s sensors create the connection
 Meaning and truth of percept sentences through
sensing and sentence construction
 General rules
 From perceptual experience

CS 420: Artificial Intelligence 30

Propositional Logic: Syntax
 Propositional logic is the simplest logic – illustrates
basic ideas

 The proposition symbols S1, S2 etc are sentences

 If S is a sentence, S is a sentence (negation)
 If S1 and S2 are sentences, S1  S2 is a sentence
(conjunction)
 If S1 and S2 are sentences, S1  S2 is a sentence
(disjunction)
 If S1 and S2 are sentences, S1  S2 is a sentence
(implication)
 If S1 and S2 are sentences, S1  S2 is a sentence
(biconditional)

CS 420: Artificial Intelligence 31

Propositional Logic: Semantics
Each model specifies true/false for each proposition symbol

E.g. P1,2 P2,2 P3,1

false true false

With these symbols, 8 possible models, can be enumerated automatically.

Rules for evaluating truth with respect to a model m:

S is true iff S is false

S1  S2 is true iff S1 is true and S2 is true
S1  S2 is true iff S1is true or S2 is true
S1  S2 is true iff S1 is false or S2 is true
i.e., is false iff S1 is true and S2 is false
S 1  S2 is true iff S1S2 is true and S2S1 is true

Simple recursive process evaluates an arbitrary sentence, e.g.,

P1,2  (P2,2  P3,1) = true  (true  false) = true  true = true

CS 420: Artificial Intelligence 32

Truth Tables for Connectives

CS 420: Artificial Intelligence 33

Wumpus World Sentences
 Let Pi,j be true if there is a pit in [i, j]
 Let Bi,j be true if there is a breeze in [i, j]
 R1: P1,1
 R2: B1,1
 R3: B2,1

 "Pits cause breezes in adjacent squares“

 R4: B1,1  (P1,2  P2,1)
 R5: B2,1  (P1,1  P2,2  P3,1)

 The KB consists of the above 5 sentences. It can

also be considered as a single sentence – the
conjunction R1  R2  R3  R4  R5

CS 420: Artificial Intelligence 34

Inference
 The aim of logical inference is to decide
whether KB╞ Ɑ for some sentence Ɑ.

 Is P2,2 entailed?

 Our first algorithm for inference will be a

direct implementation of the definition of
entailment:
 Enumerate the models, and check that Ɑ is true in
every model in which KB is true

CS 420: Artificial Intelligence 35

Truth Tables for Inference

P1,2 is false, cannot decide on p2,2

α1 = "[1,2] is safe"

CS 420: Artificial Intelligence 36

Inference by Enumeration
 Depth-first enumeration of all models is sound and complete

 For n symbols, time complexity is O(2n), space complexity is O(n)

PL-TRUE? returns true if a sentence holds within a model.

CS 420: Artificial Intelligence 37

Proof Methods In General
 Proof methods divide into (roughly) two kinds:
 Model checking
 truth table enumeration (always exponential in n)
 improved backtracking, e.g., Davis--Putnam-Logemann-
Loveland (DPLL)
 heuristic search in model space (sound but incomplete)
 e.g., min-conflicts-like hill-climbing algorithms
 Application of inference rules
 Legitimate (sound) generation of new sentences from old
 Proof = a sequence of inference rule applications
 Can use inference rules as operators in a standard search
algorithm
 Typically require transformation of sentences into a normal
form

CS 420: Artificial Intelligence 38

Validity and Satisfiability
 A sentence is valid if it is true in all models, also known as
tautology
 e.g., True, A  A, A  A, (A  (A  B))  B

 Validity is connected to inference via the Deduction Theorem

 KB ╞ α if and only if (KB  α) is valid

 A sentence is satisfiable if it is true in some model

 e.g., A  B, C

 A sentence is unsatisfiable if it is true in no models

 e.g., A  A

 Satisfiability is connected to inference via the following:

 KB ╞ α if and only if (KB  α) is unsatisfiable
 Thus proof by contradiction

CS 420: Artificial Intelligence 39

Equivalence Rules
 Two sentences are logically equivalent iff they are true in the
same models: α ≡ ß iff α╞ β and β╞ α

CS 420: Artificial Intelligence 40

Inference Rules
From Can Derive Abbreviation for rule
R, R  S S Modus Ponens- mp
R  S, S´ R´ Modus Tollens- mt
R, S RΛS Conjunction-con
RΛS R, S Simplification- sim
R RVS Addition- add

 R1: ¬P1,1
 R2: B1,1  P1,2 V P2,1
 R3: B2,1  P1,1 V P2,2 V P3,1
 R4: ¬B1,1
 R5: B2,1

 Can we prove ¬P1,2?

CS 420: Artificial Intelligence 41

Now, Your Turn
From Can Derive Abbreviation for rule
R, R  S S Modus Ponens- mp
R  S, S´ R´ Modus Tollens- mt
R, S RΛS Conjunction-con
RΛS R, S Simplification- sim
R RVS Addition- add

 Prove ¬P2,1 and P3,1 given:

R1: P1, 1
R 2 : B1, 1  P1, 2  P 2, 1
R3 : B 2, 1  P1, 1  P 2, 2  P 3, 1
R 4 : B1, 1
R5 : B 2, 1
R 6 : P 2, 2
CS 420: Artificial Intelligence 42
Proof Problem as Search Problem
 Initial state: initial knowledge base
 Actions:
 set of actions consists of all the inference rules applied to
all the sentences that match the left half of the inference
rule
 Result: add the sentence in the right half of the inference
rule
 Goal: the goal is a state that contains the sentence we are
trying to prove

 Practically, finding a proof can be more efficient because the

proof can ignore irrelevant propositions, no matter how many
of them there are

CS 420: Artificial Intelligence 43

Resolution
 The inference rules can be applied whenever suitable
premises are found in the KB

 The conclusion of the rule must follow regardless of what else

is in the KB

 The above rules are sound, but if the available inference rules
are inadequate, then it’s not complete

 Now, we introduce a single inference rule, resolution, that

gives a complete inference algorithm when coupled with any
complete search algorithm

CS 420: Artificial Intelligence 44

Example
 Recall our previous proof of P3,1 given:
R1: P1, 1
R 2 : B1, 1  P1, 2  P 2, 1
R3 : B 2, 1  P1, 1  P 2, 2  P 3, 1
R 4 : B1, 1
R5 : B 2, 1
R6 : P 2, 2
... Unit resolution
R9 : P1, 1  P 2, 2  P 3, 1
R10 : P 2, 2  P 3, 1 Resolution rule: ¬P1,1 resolves with P1,1 in R9

R11: P 3, 1 Resolution rule: ¬P2,2 resolves with P2,2 in R10

CS 420: Artificial Intelligence 45

Resolution Rules
 Unit resolution
 li and m are complementary literals
l 1 …  l k, m
l 1  …  l i-1  l i+1  …  lk

 Generalized resolution rule

l 1 …  l k, m1  …  m n
l 1  …  l i-1  l i+1  …  lk  m1  …  m j-1  m j+1 ...  m n

 where li and mj are complementary literals

CS 420: Artificial Intelligence 46

Soundness and Completeness
 Soundness of resolution inference rule:
 if l i is true, then mj is false
 hence (m1  …  m j-1  m j+1 ...  mn) must be true
 if l i is false, then
 (l1  …  l i-1  l i+1  …  lk) must be true
 no matter what
 (l1  …  l i-1  l i+1  …  lk) V (m1  …  m j-1  m j+1 ...  mn) is
true

 Any complete search algorithm, applying only the

resolution rule, can derive any conclusion entailed
by any knowledge base in propositional logic
CS 420: Artificial Intelligence 47
 A resolution-based theorem prover can, for
any sentences α and β in propositional logic,
decide whether α╞ β

CS 420: Artificial Intelligence 48

Resolution and CNF
 Resolution rule applies only to
disjunctions of literals

 Every sentence of propositional

logic is logically equivalent to a
conjunction of disjunctions of
literals

 Conjunctive Normal Form

(CNF): P1,3  P2,2, P2,2

 conjunction of disjunctions of
literals clauses P1,3
 E.g., (A  B)  (B  C  D)

CS 420: Artificial Intelligence 49

Conversion to CNF
B1,1  (P1,2  P2,1)

1. Eliminate , replacing α  β with (α  β)(β  α).

(B1,1  (P1,2  P2,1))  ((P1,2  P2,1)  B1,1)

2. Eliminate , replacing α  β with α β.

(B1,1  P1,2  P2,1)  ((P1,2  P2,1)  B1,1)

3. Move  inwards using de Morgan's rules and double-

negation:
(B1,1  P1,2  P2,1)  ((P1,2  P2,1)  B1,1)

4. Apply distributive law ( over ) and flatten:

(B1,1  P1,2  P2,1)  (P1,2  B1,1)  (P2,1  B1,1)

CS 420: Artificial Intelligence 50

Resolution Algorithm
 To show that KB╞ α, we show that (KB ^ α) is unsatisfiable,
proof by contradiction

 First, (KB ^ α) is converted into CNF

 Then the resolution rule is applied to the resulting clauses
 Each pair that contains complementary literals is resolved to
produce a new clause, which is added to the set if it’s not present
 Process continues until:
 no new clauses can be added, thus KB does not entail α
 two clauses resolve to yield {}, thus KB entails α

 {} is a disjunction of no disjuncts is equivalent to False, thus

the contradiction

CS 420: Artificial Intelligence 51

Resolution Algorithm
 Proof by contradiction, i.e., show that KBα is
unsatisfiable

CS 420: Artificial Intelligence 52

Resolution Example
 KB = (B1,1  (P1,2 P2,1))  B1,1 α = P1,2

CS 420: Artificial Intelligence 53

Horn Clauses
 Why horn clauses?
 In many practical situations, however, the full
power of resolution is not needed
 Real-world KBs often contain some restricted
clauses – Horn clauses

 Horn Form (restricted)

 KB = conjunction of Horn clauses
 Horn clause = disjunction of literals of which at
most one is positive
 E.g., (L1,1 V Breeze V B1,1)

CS 420: Artificial Intelligence 54

Properties of Horn Clauses
 Every horn clause can be written as an implication whose
premise is a conjunction of positive literals and whose
conclusion is a single positive literal
 E.g., (L1,1 ^ Breeze)  B1,1

 Modus Ponens (for Horn Form): complete for Horn KBs

α1, … ,αn, α1  …  αn  β
β
 Definite clause: horn clauses with exactly one positive literal
 The positive literal is called the head and the negative literals
form the body
 Definite clauses form the basis for logic programming

 Can be used with forward chaining or backward chaining

 These algorithms are very natural and run in linear time

CS 420: Artificial Intelligence 55

Forward Chaining
 Idea:
 fire any rule whose premises are satisfied in the KB
 add its conclusion to the KB, until query is found

and-or
graph

CS 420: Artificial Intelligence 56

Forward Chaining Algorithm

if p = q then return true

if inferred[p] = false then
inferred[p] = true
for each clause c in KB where p is in c.Premise do
decrement count[c]
if count[c] = 0 then add c.Conclusion to agenda

 Forward chaining is sound and complete for Horn KB

CS 420: Artificial Intelligence 57

Forward Chaining Example

CS 420: Artificial Intelligence 58

Forward Chaining Example

CS 420: Artificial Intelligence 59

Forward Chaining Example

CS 420: Artificial Intelligence 60

Forward Chaining Example

CS 420: Artificial Intelligence 61

Forward Chaining Example

CS 420: Artificial Intelligence 62

Forward Chaining Example

CS 420: Artificial Intelligence 63

Proof of Soundness
 Every inference is an application of Modus
Ponens

CS 420: Artificial Intelligence 64

Proof of Completeness
 FC derives every atomic sentence that is entailed
by KB
1. FC reaches a fixed point where no new atomic sentences
are derived
2. Consider the final state as a model m, assigning
true/false to symbols; true for each symbol inferred,
false otherwise
3. Every definite clause in the original KB is true in m
a1  …  ak  b, assuming the opposite, then it contradicts
“fixed point”
4. Hence, every entailed atomic sentence will be derived

CS 420: Artificial Intelligence 65

Backward Chaining
 Idea: work backwards from the query q:
 to prove q by BC
 check if q is known already, or
 prove by BC all premises of some rule concluding q

 Avoid loops: check if new subgoal is already on the

goal stack

 Avoid repeated work: check if new subgoal

 has already been proved true, or

 has already failed

CS 420: Artificial Intelligence 66

Backward Chaining Example

CS 420: Artificial Intelligence 67

Backward Chaining Example

CS 420: Artificial Intelligence 68

Backward Chaining Example

CS 420: Artificial Intelligence 69

Backward Chaining Example

CS 420: Artificial Intelligence 70

Backward Chaining Example

CS 420: Artificial Intelligence 71

Backward Chaining Example

CS 420: Artificial Intelligence 72

Backward Chaining Example

CS 420: Artificial Intelligence 73

Backward Chaining Example

CS 420: Artificial Intelligence 74

Backward Chaining Example

CS 420: Artificial Intelligence 75

Backward Chaining Example

CS 420: Artificial Intelligence 76

Forward vs. Backward Chaining
 FC is data-driven, automatic, unconscious
processing
 e.g., object recognition, routine decisions

 May do lots of work that is irrelevant to the goal

 BC is goal-driven, appropriate for problem-solving,

 e.g., Where are my keys? How do I get into a PhD
program?

 Complexity of BC can be much less than linear in

size of KB

CS 420: Artificial Intelligence 77

Inference-Based Agents in the Wumpus World

 A wumpus-world agent using propositional logic:

P1,1
W1,1
Bx,y  (Px,y+1  Px,y-1  Px+1,y  Px-1,y)
Sx,y  (Wx,y+1  Wx,y-1  Wx+1,y  Wx-1,y)
W1,1  W1,2  …  W4,4
W1,1  W1,2
W1,1  W1,3
…

 64 distinct proposition symbols, 155 sentences

CS 420: Artificial Intelligence 78
Expressiveness Limitation
 KB contains "physics" sentences for every single
square

 For every time t and every location [x,y],

Lx,y  FacingRight t  Forward t  Lx+1,y

 Rapid proliferation of clauses

CS 420: Artificial Intelligence 79

Summary
 Logical agents apply inference to a knowledge base to derive new
information and make decisions

 Basic concepts of logic:

 syntax: formal structure of sentences
 semantics: truth of sentences based on models
 entailment: necessary truth of one sentence given another
 inference: deriving sentences from other sentences
 soundness: derivations produce only entailed sentences
 completeness: derivations can produce all entailed sentences

 Wumpus world requires the ability to represent partial and negated

information, reason by cases, etc.

 Resolution is complete for propositional logic

Forward, backward chaining are linear-time, complete for Horn
clauses

 Propositional logic lacks expressive power

CS 420: Artificial Intelligence 80