CS 4700:

Foundations of Artificial Intelligence

Bart Selman
[email protected]

Module: Knowledge, Reasoning, and Planning

Part 1

Logical Agents
R&N: Chapter 7

1
 A Model-Based Agent

Re
qu
ire
s:
K no
wl
ed
ge
a nd
R ea
so
n in
g

2
 Knowledge and Reasoning
Knowledge and Reasoning:
humans are very good at acquiring new information by
combining raw knowledge, experience with reasoning.
AI-slogan: “Knowledge is power” (or “Data is power”?)

Examples:
Medical diagnosis --- physician diagnosing a patient
infers what disease, based on the knowledge he/she
acquired as a student, textbooks, prior cases
Common sense knowledge / reasoning ---
common everyday assumptions / inferences
e.g., “lecture starts at four” infer pm not am;
when traveling, I assume there is some way to get from the
airport to the hotel.
3
Logical agents:
Agents with some representation of the
complex knowledge about the world / its environment,
and uses inference to derive new information from that
knowledge combined with new inputs (e.g. via perception).

Key issues:
1- Representation of knowledge
What form? Meaning / semantics?
2- Reasoning and inference processes
Efficiency.

4
 Knowledge-base Agents

Key issues:
– Representation of knowledge  knowledge base
– Reasoning processes  inference/reasoning

Knowledge base = set of sentences in a formal language

representing facts about the world(*)

(*) called Knowledge Representation (KR) language

5
 Knowledge bases
Key aspects:
– How to add sentences to the knowledge base
– How to query the knowledge base

Both tasks may involve inference – i.e. how to derive new

sentences from old sentences

Logical agents – inference must obey the fundamental

requirement that when one asks a question to the knowledge
base, the answer should follow from what has been told to the
knowledge base previously. (In other words the inference
process should not “make things” up…)

6
 A simple knowledge-based agent

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
–
7
KR language candidate:
logical language (propositional / first-order) combined
with a logical inference mechanism

How close to human thought? (mental-models / Johnson-

Laird).

What is “the language of thought”?

Greeks / Boole / Frege --- Rational thought: Logic?
Why not use natural language (e.g. English)?

We want clear syntax & semantics (well-defined

meaning), and, mechanism to infer new information.
Soln.: Use a formal language.
8
Consider: to-the-right-of(x,y)
9
10
11
 Procedural style:

Knowledge-based alternative:

12
13
14
The “symbol grounding problem.” 15
16
17
18
 Note: KB defines exactly the set
of worlds we are interested in.

I.e.: Models(KB) Models( ) 19

20
21
Note: (1) This was Aristotle’s original goal ---
Construct infallible arguments based purely
on the form of statements --- not on the “meaning”
of individual propositions.
(2) Sets of models can be exponential size or worse,
compared to symbolic inference (deduction).
22
23
24
25
 Intermezzo: Reflections on Inference,
Knowledge, and Data ca. 2012
In the logical agent perspective, we take the “knowledge-based”
approach.
We’ll have a knowledge base, capturing our world knowledge in a
formal language. We’ll use an inference procedure to derive new
information.
Representation has well-defined syntax and semantics / meaning.

How far are actual AI systems in the knowledge / reasoning

paradigm?

Specifically, where do Deep Blue, Watson, and Siri fit?

What about IR, Google Translate, and Google’s Knowledge Graph?
And, “shallow semantics” / Semantic Web?
26
IBM’s Jeopardy! playing system.
Watson

27
Chris Welty from IBM on Watson and “understanding”

http://spectrum.ieee.org/podcast/at-work/innovation/what-is-toronto

See 4 min mark for discussion on “understanding.”

Till around min 11.
28
 Key aspects: Watson
1) Limited number of categories of types of questions
2) Well-defined knowledge sources (Wikipedia; Encyclopedias;
Dictionaries etc.) Not: WWW (Contrast: Google Knowl. Graph)
3) Around 70 “expert” modules that handle different aspects of the
possible question and answers. 1000+ hypotheses
4) Final result based on a confidence vote.
5) Limited language parsing. Lot based on individual words (like
IR).
6) Form of IR “+” (parsing; limited knowledge graph: “books have
authors” “movie has actors and director” “verb represents an
action” etc. )
7) Some game theory to reason about whether to answer

To some extent: Watson “understands”!

(It’s actually a bit of a problem for IBM… Clients expect too much!)
29
 Semantic Web
Data.gov in process of converting CSV data into Resource Description
Framework
– RPI Tetherless World Constellation - Jim Hendler, Berners-Lee
– RDF - triples (facts), class / subclass
– OWL - defined classes, constraints
Directly grows out of KR work in
AAAI community 1985 - 2005
– Ron Brachman, Hector Levesque,
Deborah McGuinness, Alex Borgida,
Peter Patel-Schneider, Franz Baader,
Francesco Donini, ...

31
 AI Knowledge-
Data-
Knowledge
Inference
Intensive
NLU Triangle
Common Sense 2012
Computer Vision
20+yr GAP!

M
Google’s Knowl. Graph

ac
hin
Semantic Web Watson
Verification Google Transl.

e
Le
Siri
Robbin’s Conj.

ar
nin
4-color thm. Google Search (IR)
Deep Blue

g
Inference/Se
Data
arch
Intensive
Intensive
32
Back to Knowledge-Based Agents…

33
 Illustrative example: Wumpus World
Performance measure (Somewhat whimsical!)
– gold +1000,
– death -1000
(falling into a pit or being eaten by the wumpus)
– -1 per step, -10 for using the arrow
Environment
– Rooms / squares connected by doors.
– Squares adjacent to wumpus are smelly
– Squares adjacent to pit are breezy
– Glitter iff gold is in the same square
– 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
– Randomly generated at start of game. Wumpus only senses current room.
Sensors: Stench, Breeze, Glitter, Bump, Scream [perceptual inputs]
Actuators: Left turn, Right turn, Forward, Grab, Release, Shoot 34
 Wumpus world characterization
Fully Observable No – only local perception

Deterministic Yes – outcomes exactly specified

Static Yes – Wumpus and Pits do not move

Discrete Yes

Single-agent? Yes – Wumpus is essentially a “natural feature.”

35
 Exploring a wumpus world

The knowledge base of the agent

consists of the rules of the
Wumpus world plus the percept
“nothing” in [1,1]

Boolean percept
feature values:
<0, 0, 0, 0, 0>
None, none, none, none, none
Stench, Breeze, Glitter, Bump, Scream
36
 World “known” to agent
at time = 0.

T=0 The KB of the agent consists of

None, none, none, none, none the rules of the Wumpus world plus
the percept “nothing” in [1,1].
Stench, Breeze, Glitter, Bump, Scream
By inference, the agent’s knowledge
base also has the information that
[2,1] and [1,2] are okay.
Added as propositions. 37
 Further exploration
T=0
T=1

P?

V A/B P?

None, breeze, none, none, none

None, none, none, none, none
A – agent
V – visited
Stench, Breeze, Glitter, Bump, Scream B - breeze
@ T = 1 What follows?
Where next? Pit(2,2) or Pit(3,1)
38
 T=3
4

3 W
S P? S
2 P
1 P? P
1 2 3 4
Stench, none, none, none, none
Stench, Breeze, Glitter, Bump, Scream

Where is Wumpus?

Wumpus cannot be in (1,1) or in (2,2) (Why?) Wumpus in (1,3)

Not breeze in (1,2)  no pit in (2,2); but we know there is
pit in (2,2) or (3,1)  pit in (3,1)
39
We reasoned about the possible states
the Wumpus world can be in, given our W
percepts and our knowledge of the rules P
of the Wumpus world.
I.e., the content of KB at T=3. P

What follows is what holds true in all those worlds that

satisfy what is known at that time T=3 about the
particular Wumpus world we are in.
Example property: P_in_(3,1)
Models(KB) Models(P_in_(3,1))

Essence of logical reasoning:

Given all we know, Pit_in_(3,1) holds.
(“The world cannot be different.”) 40
 Formally: Entailment
Knowledge Base (KB) in the Wumpus World 
Rules of the wumpus world + new percepts
Situation after detecting nothing in [1,1],
moving right, breeze in [2,1]. I.e. T=1.

Consider possible models for KB with respect to

the cells (1,2), (2,2) and (3,1), with respect to T=1
the existence or non existence of pits

3 Boolean choices 
8 possible interpretations
(enumerate all the models or
“possible worlds” wrt Pitt location)

41
Is KB consistent with all
8 possible worlds? Worlds
that violate KB
(are inconsistent
with what we
know)

KB = Wumpus-world rules + observations (T=1)

Q: Why does world violate KB? 42

So, KB defines Entailment in Wumpus World
all worlds that
we hold possible.
Queries: we want to know the properties of those worlds.
That’s how the semantics of logical entailment is defined.

Models of the KB and α1

Note: \alpha_1
holds in more
models than KB.
That’s OK, but we
don’t care about
those worlds.
KB = Wumpus-world rules + observations
α1 = "[1,2] has no pit", KB ╞ α1
– In every model in which KB is true, α1 is True (proved by
43
“model checking”)
KB = wumpus-world rules + observations
Wumpus models
α2 = "[2,2] has no pit", this is only True in some
of the models for which KB is True, therefore KB ╞ α2

Model Checking
Models of α2

A model of KB where α2 does NOT hold!

44
 Entailment via
“Model Checking”
Inference by Model checking –
We enumerate all the KB models and check if α1 and α2 are
True in all the models (which implies that we can only use it
when we have a finite number of models).

I.e. using semantics directly.

Models(KB) Models( )

45
 Example redux: More formal

P?

V A/B P?

None, none, none, none, none None, breeze, none, none, none
Stench, Breeze, Glitter, Bump, Scream A – agent
V – visited
B - breeze
How do we actually encode background
knowledge and percepts in formal language?
46
 Wumpus World KB
Define propositions:
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].

Sentence 1 (R1):  P1,1 [Given.]

Sentence 2 (R2):  B1,1 [Observation T = 0.]
[Observation T = 1.]
Sentence 3 (R3): B2,1

"Pits cause breezes in adjacent squares”

Sentence 4 (R4): B1,1  (P1,2  P2,1)
Sentence 5 (R5): B2,1  (P1,1  P2,2  P3,1)
etc.
Notes: (1) one such statement about Breeze for each square.
(2) similar statements about Wumpus, and stench
and Gold and glitter. (Need more propositional
47
letters.)
 What about Time? What about Actions?
Is Time represented?
No!
Can include time in propositions:
Explicit time Pi,j,t Bi,j,t Li,j,t etc.
Many more props: O(TN2) (Li,j,t for agent at (i,j) at time t)
Now, we can also model actions, use props: Move(i ,j ,k ,l ,t)
E.g. Move(1, 1, 2, 1, 0)
What knowledge axiom(s) capture(s) the effect of an Agent
move?
)
Move(i, j, k, l ,t) (: L(i, j, t+1) Æ L(k, l, t+1))

Is this it?
What about i, j, k, and l?
What about Agent location at time t? 48
Improved: Move implies a change in the world state;
a change in the world state, implies a move occurred!
Move(i, j, k, l ,t) , (L(i, j, t) Æ : L(i, j, t+1) Æ L(k, l, t+1))
For all tuples (i, j, k, l) that represent legitimate possible moves.
E.g. (1, 1, 2, 1) or (1, 1, 1, 2)

Still, some remaining subtleties when representing time and

actions. What happens to propositions at time t+1 compared to at
time t, that are *not* involved in any action?
E.g. P(1, 3, 3) is derived at some point.
What about P(1, 3, 4), True or False?
R&N suggests having P as an “atemporal var” since it cannot change over

time. Nevertheless, we have many other vars that can change over time,
called “fluents”.
Values of propositions not involved in any action should not
change! “The Frame Problem” / Frame Axioms R&N 7.7.1 49
 Successor-State Axioms
Axiom schema:
F is a fluent (prop. that can change over time)

Ft+1 , (ActionCausesFt Ç (F(t) : ActionCausesNotFt))

For example:

i.e. L_1,1 was “as before” with [no movement action or bump into wall]
or resulted from some action (movement into L_1,1.

50
 Actions and inputs up to time 6 Some example inferences
Note: includes turns! Section 7.7.1 R&N

Define “OK”:

In milliseconds, with modern SAT solver. 51

 Alternative formulation: Situation Calculus
R&N 10.4.2

No explicit time. Actions are what changes the world

from “situation” to “situation”. More elegant, but
still need frame axioms to capture what stays the same.
Inherent with many representation formalisms: “physical”
persistance does not come for free! (and probably shouldn’t) 52
 Inference by enumeration / “model checking”
Style I
The goal of logical inference is to decide whether KB╞ α, for some .
For example, given the rules of the Wumpus World, is P22
entailed? Relevant propositional symbols:
R1:  P 1,1 ?
R2: B 1,1
Models(KB) Models( P22 )
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)

Inference by enumeration. We have 7 relevant symbols

Therefore 27 = 128 interpretations.
Need to check if P22 is true in all of the KB models
(interpretations that satisfy KB sentences).

Q.: KB has many more symbols. Why can we restrict ourselves

to these symbols here? But, be careful, typically we can’t!! 53
 All equivalent
Prop. / FO Logic
1) KB ² ® entailment
2) M(KB) µ M(®) by defn. / semantic proofs / truth tables
“model checking” /enumeration
(style I, R&N 7.4.4)
3) ² (KB ) ®) deduction thm. R&N 7.5

4) KB ` ® soundness and completeness

logical deduction / symbol pushing
proof by inference rules (style II)
e.g. modus ponens (R&N 7.5.1)
5) (KB Æ : ®) is inconsistent Proof by contradiction
use CNF / clausal form
Resolution (style III, R&N 7.5)
SAT solvers (style IV, R&N 7.6)
most effective
54
 Proof techniques

M(KB) µ M(®) by defn. / semantic proofs / truth tables

“model checking”
(style I, R&N 7.4.4) Done.

KB ` ® soundness and completeness

logical deduction / symbol pushing
proof by inference rules (style II)
e.g. modus ponens (R&N 7.5.1)
(KB Æ : ®) is inconsistent Proof by contradiction
use CNF / clausal form
Resolution (style III, R&N 7.5)
SAT solvers (style IV, R&N 7.6)
most effective
55
 Aside
Standard syntax and semantics for propositional
logic. (CS-2800; see 7.4.1 and 7.4.2.)

Syntax:

56
Semantics

Note: Truth value of a sentence is built from its

parts “compositional semantics”

57
 Logical equivalences

(*)

(*) key to go to clausal (Conjunctive Normal Form)

Implication for “humans”; clauses for machines.
de Morgan laws also very useful in going to clausal form.
58
KB at T = 1: Style II: Proof by inference rules
R1: P1,1
R2: B1,1
Modus Ponens (MP)
R3: B2,1

R4: B1,1  (P1,2  P2,1)

R5: B2,1  (P1,1  P2,2  P3,1)
How can we show that KR ² : P1,2 ? P?
R4: B1,1  (P1,2  P2,1)
It follows (bicond elim)
V A/B P?
B1,1 ) (P1,2  P2,1) Æ (P1,2  P2,1) ) B1,1
And-elimination (7.5.1 R&N):
(P1,2  P2,1) ) B1,1 Wumpus world
By contrapositive:
at T = 1
: B1,1 ) (: (P1,2 Ç P2,1))
Thus (de Morgan):
: B1,1 ) (:P1,2 Æ : P2,1)
By Modus Ponens using R2, we get
(:P1,2 Æ : P2,1)
Finally, by And-elimination, we get:
: P1,2 59
 Length of Proofs
Why bother with inference rules? We could always use a truth table
to check the validity of a conclusion from a set of premises.
But, resulting proof can be much shorter than truth table method.

Consider KB:
p_1, p_1  p_2, p_2  p_3, …, p_(n-1)  p_n

To prove conclusion: p_n

Inference rules: n-1 MP steps Truth table:

2n

Key open question: Is there always a short proof for any valid
conclusion? Probably not. The NP vs. co-NP question.
(The closely related: P vs. NP question carries a $1M prize.)

60
First, we need a conversion to Conjunctive Style III: Resolution
Normal Form (CNF) or Clausal Form.

Let’s consider converting R4 in clausal form:

R4: B1,1  (P1,2  P2,1)
We have:
B1,1 ) (P1,2 Ç P2,1)
which gives (implication elimination):
P?
(: B1,1 Ç P1,2 Ç P2,1)

V A/B P?
Also
(P1,2  P2,1) ) B1,1
which gives: Wumpus world
(: (P1,2 Ç P2,1) Ç B1,1)
at T = 1
Thus,
(: P1,2 Æ : P2,1) Ç B1,1
leaving,
(: P1,2 Ç B1,1 )
(: P2,1 Ç B1,1 )

(note: clauses in red) 61

 KB at T = 1:
R1: P1,1
R2: B1,1
R3: B2,1

R4: B1,1  (P1,2  P2,1)

R5: B2,1  (P1,1  P2,2  P3,1)

KB at T=1 in clausal form:

P?
R1: P1,1
V A/B P?
R2: B1,1
R3: B2,1
Wumpus world
R4a: : B1,1 Ç P Ç P2,1
1,2 at T = 1
R4b: : P1,2 Ç B1,1
R4c: : P2,1 Ç B1,1

R5a: : B2,1 Ç P1,1 Ç P2,2 Ç P3,1

R5b: : P1,1 Ç B2,1
R5c: : P2,2 Ç B2,1
62
R5d: : P3,1 Ç B2,1
 How can we show that KR ² : P1,2 ?

Proof by contradiction:
Need to show that (KB Æ P1,2) is
inconsistent (unsatisfiable).
Resolution rule:

(® Ç p) and (¯ Ç : p)

gives resolvent (logically valid conclusion):

(® Ç ¯)

If we can reach the empty clause, then KB

is inconsistent. (And, vice versa.)
63
KB at T=1 in clausal form:
R1: P1,1
R2: B1,1
R3: B2,1

R4a: : B1,1 Ç P1,2 Ç P2,1

R4b: : P1,2 Ç B1,1
R4c: : P2,1 Ç B1,1 P?
R5a: : B2,1 Ç P1,1 Ç P2,2 Ç P3,1 V A/B P?
R5b: : P1,1 Ç B2,1
R5c: : P2,2 Ç B2,1 Wumpus world
R5d: : P3,1 Ç B2,1 at T = 1
Show that (KB Æ P1,2) is inconsistent.
(unsatisfiable)
R4b with P1,2 resolves to B1,1,
which with R2, resolves to the empty clause, .
So, we can conclude KB ² : P1,2.
64
(make sure you use “what you want to prove.”)
KB at T=1 in clausal form:
R1: P1,1
R2: B1,1
R3: B2,1
R4a: : B1,1 Ç P Ç P2,1
1,2
R4b: : P1,2 Ç B1,1
R4c: : P2,1 Ç B1,1
P?
R5a: : B2,1 Ç P1,1 Ç P2,2 Ç P3,1 P?
V A/B
R5b: : P1,1 Ç B2,1
R5c: : P2,2 Ç B2,1
Wumpus world
R5d: : P3,1 Ç B2,1
Note that R5a resolved with R1, and at T = 1
then resolved with R3, gives (P2,2 Ç P3,1).

Almost there… to show KB ² (P2,2 Ç P3,1) , we need to show

KB Æ (: (P2,2 Ç P3,1)) is inconsistent.
So, show KB Æ : P2,2 Æ : P3,1 is inconsistent.
This follows from (P2,2 Ç P3,1); because in two more resolution
steps, we get the empty clause (a contradiction). 65
 Length of Proofs
What is hard for resolution?
Consider:
Given a fixed pos. int. N

(P(i,1) Ç P(i,2) Ç … P(i,N)) for i = 1, … N+1

(: P(i,j) Ç : P(i’,j)) for j= 1, … N;

i = 1, … N+1;
i’ = 1, …N+1; j =/= j’
What does this encode?
Think of: P(i,j) for “object i in location j”

Pigeon hole problem…

Provable requires exponential number of resolution
steps to reach empty clause (Haken 1985). Method “can’t count.”
66
Instead of using resolution to show that
Style IV: SAT Solvers

KB Æ : ® is inconsistent,

modern Satisfiability (SAT) solvers operating on the clausal form

are *much* more efficient.
The SAT solvers treat the set of clauses as a set of constraints
(disjunctions) on Boolean variables, i.e., a CSP problem!
Current solvers are very powerful. Can handle 1 Million+
variables and several millions of clauses.

Systematic: Davis Putnam (DPLL) + series of improvements

Stochastic local search: WalkSAT (issue?)

See R&N 7.6. “Ironically,” we are back to semantic model

checking, but way more clever than basic truth
assignment enumeration (exponentially faster)! 67
 DPLL improvements
Backtracking + …

1) Component analysis (disjoint sets of constraints? Problem

decomposition?)
2) Clever variable and value ordering (e.g. degree heuristics)
3) Intelligent backtracking and clause learning (conflict learning)
4) Random restarts (heavy tails in search spaces…)
5) Clever data structures

1+ Million Boolean vars & 10+ Million clause/constraints

are feasible nowadays. (e.g. Minisat solver)

Has changed the world of verification (hardware/software)

over the last decade (incl. Turing award for Clarke).
Widely used in industry, Intel, Microsoft, IBM etc. 68
 Satisfiability (SAT/CSP) or Inference?

Really solving the same problem but SAT/CSP view appears more effective.

KB ² ® iff M(KB) µ M(®) iff (KB Æ :®) is unsat

Assume KB and ® is in CNF (clausal form).

Note that: KB ² (¯ Æ °) iff
(KB ² ¯) and (KB ² °)
(consider defn. in terms of models)
So, can break conjunction in several queries, one for
each disjunction (clause). Each a call to a SAT solver to check
KB Æ : (l1 Ç l2 … ) for consist. or equivalently
check KB Æ : l1 Æ : l2 … (i.e. CNF form) for consistency.

69
 KB ² (l1 Ç l2 …) iff KB Æ : l1 Æ : l2 …
The SAT solvers essentially views the KB as a set of constraints that
defines a subset of the set of all 2N possible worlds (N prop. vars.).

The query ® also defines a subset of the 2N possible worlds. (Above

we used a single clause (l1 Ç l2 …) where li is a var or its negation.)

The SAT/CSP solvers figures out whether there are any worlds
consistent with the KB constraints that do not satisfy the constraints
of the query. If so, than the query does not follow from the KB.
Aside: In verification, such a world exposes a bug.
If such worlds do not exist, then the query must be entailed by the KB.
(M(KB) µ M(®) Search starts from query!)

Viewing logical reasoning as reasoning about constraints on the

way the “world can be” is quite actually quite natural! It’s definitely
a computationally effective strategy.
70
Some examples of SAT solving

71
 Application: I --- Diagnosis
• Problem: diagnosis a malfunctioning device
– Car
– Computer system
– Spacecraft
• where
– Design of the device is known
– We can observe the state of only certain parts of
the device – much is hidden
 Model-Based, Consistency-Based Diagnosis

• Idea: create a logical formula that describes how

the device should work
– Associated with each “breakable” component C is a
proposition that states “C is okay”
– Sub-formulas about component C are all conditioned on
C being okay
• A diagnosis is a smallest of “not okay” assumptions
that are consistent with what is actually observed
 Consistency-Based Diagnosis
1. Make some Observations O.
2. Initialize the Assumption Set A to assert that
all components are working properly.
3. Check if the KB, A, O together are inconsistent
(can deduce false).
4. If so, delete propositions from A until
consistency is restored (cannot deduce false).
The deleted propositions are a diagnosis.
There may be many possible diagnoses
Example: Automobile Diagnosis
• Observable Propositions:
EngineRuns, GasInTank, ClockRuns
• Assumable Propositions:
FuelLineOK, BatteryOK, CablesOK, ClockOK
• Hidden (non-Assumable) Propositions:
GasInEngine, PowerToPlugs
• Device Description F:
(GasInTank  FuelLineOK)  GasInEngine
(GasInEngine  PowerToPlugs)  EngineRuns
(BatteryOK  CablesOK)  PowerToPlugs
(BatteryOK  ClockOK)  ClockRuns
• Observations:
 EngineRuns, GasInTank, ClockRuns

Note: of course a highly simplified set of axioms.

 (GasInTank  FuelLineOK)  GasInEngine
Example (GasInEngine  PowerToPlugs)  EngineRuns

• Is F  Observations  Assumptions consistent?

• F  {EngineRuns, GasInTank, ClockRuns}

 { FuelLineOK, BatteryOK, CablesOK, ClockOK }  false
– Must restore consistency!
• F  {EngineRuns, GasInTank, ClockRuns}
 { BatteryOK, CablesOK, ClockOK }  satisfiable
–  FuelLineOK is a diagnosis
• F  {EngineRuns, GasInTank, ClockRuns}
 {FuelLineOK, CablesOK, ClockOK }  false
–  BatteryOK is not a diagnosis
 Complexity of Diagnosis
• If F is Horn, then each consistency test takes linear O(n)
time – unit propagation is complete for Horn clauses.
• Complexity = ways to delete propositions from
Assumption Set that are considered.
– Single fault diagnosis – O(n2)
– Double fault diagnosis – O(n3)
– Triple fault diagnosis – O(n4)

In practice, for non-Horn use SAT solver

for consistency check.

Horn clause: at most one positive literal. Also, consider )

 NASA Deep Space One
• Autonomous diagnosis & repair “Remote
Agent”
• Compiled systems schematic to 7,000 var
SAT problem

Started: January 1996

Launch: October 15th, 1998
Experiment: May 17-21
 Deep Space One
• a failed electronics unit
– Remote Agent fixed by reactivating the unit.

• a failed sensor providing false information

– Remote Agent recognized as unreliable and therefore correctly ignored.

• an altitude control thruster (a small engine for controlling the

spacecraft's orientation) stuck in the "off" position
– Remote Agent detected and compensated for by switching to a mode that
did not rely on that thruster.
 II --- Testing Circuit Equivalence
C C’
• Do two circuits compute
the same function?
• Circuit optimization
nand • Is there input for which
the two circuits compute
different values?
(Satisfying assignment
will tell us. Reveals bug!)

A B A B

Possible implementation
Formal spec. OR
using hardware gates (invertor & nand)
Note: if consitent,
model reveals “bug”
in hardware design. C   A  B Descr. of OR

C C’ C   ( D  E )
Descr. of
D  A Hardware.
nand E  B
Our query:
(C  C ) Does this hold?
D E
Yes iff negation with
KB is inconsistent.
What happens if you add
A B A B instead of ? Still OK?
Informally: Given the same inputs values for A and B,
logic specifies outputs (C and C’). Are there inputs for which
C and C’ differ?
 III --- SAT Translation of
• At least one queen each row:
(Q11 v Q12 v Q13 v ... v Q18)
N-Queens
(Q21 v Q22 v Q23 v ... v Q28)
...
• No attacks (columns):
(~Q11 v ~Q21)
(~Q11 v ~Q31)
(~Q11 v ~Q41)
...
• No attacks (diag; need left and right):
(~Q11 v ~Q22)
How about: No attacks (rows)
(~Q11 v ~Q33) (~Q11 v ~Q12)
(~Q11 v ~Q44) (~Q11 v ~Q13)
... (~Q11 v ~Q14)
…
Redundant! Why? Sometimes slows solver.
• At least one queen each row: More compactly
(Qi1 v Qi2 ... v QiN) for 1 <= i <= N

• No attacks (columns; “look in ith column”):

(~Qji v ~Qj’i) for 1 <= i,j,j’ <= N and j =/= j’

• No attacks (diag; need left and right):

(~Qij v ~Qi’j’) for 1 <= i, i’, j, j’ s.t. |i – i’| = |j – j’| & i =/= i’ & j =/= j’

Or: in first order logic syntax, e.g., second constraint set:

8 i 8 j 8 j’ (j =/= j’ ) (: Qj,i Ç : Qj’,i))
with bounded type quantifier 1 <= i, j, j’ <= N
Really a compact “propositional schema.”
First-order logic for finite domains is equiv. to prop. logic.
For SAT solver, always “ground to propositional.”
 IV --- SAT Translation of
At least one color per node i : Graph Coloring
(C_i1 v Q_i2 v Q_i3 v ... v Q_iK)

At most one color per node i :

(~C_ik v ~Q_ik’) for all k =/= k’

If node i and node j (=/= i) share an edge,

need to have different colors: Note: Translation from
(~C_il v ~Q_jl) for all 1 <= I <= K “problem” into SAT.
Reverse of usual
translation to show
NP-completeness.
C_ik for node i has color k
Total # colors: K. Total # nodes: N. Works also for (easy)
polytime problems!
 V --- Symbolic Model Checking
• Any finite state machine is characterized by a transition function
– CPU
– Networking protocol
• Wish to prove some invariant holds for any possible inputs
• Bounded model checking: formula is sat iff invariant fails k steps
in the future
St  vector of Booleans representing
1) Can go to CNF.
state of machine at time t
2) Equivalent to planning
as propositional inference  : State  Input  State
SATPLAN  : State  {0,1}
(Kautz & Selman 1996)
 k 1 
7.7.4 R&N 
  i 1
 i 0
S   ( S i , Ii   So    Sk 

The k-step plan
is satisfying assignment. Note: ½ is just like our “move” earlier.
Axioms says what happens “next.”
A real-world example
Bounded Model Checking instance

i.e. ((not x1) or x7)

and ((not x1) or x6)
and … etc.
 10 pages later:

…
(x177 or x169 or x161 or x153 …
or x17 or x9 or x1 or (not x185))

clauses / constraints are getting more interesting…

 4000 pages later:

!!!
a 59-cnf
clause…

…
 Finally, 15,000 pages later:

What makes this possible?

Note that: … !!!

The Chaff SAT solver (Princeton) solves

this instance in less than one minute.
 Progress in Last 20 years
• Significant progress since the 1990’s. How much?
• Problem size: We went from 100 variables, 200 constraints (early 90’s)
to 1,000,000+ variables and 5,000,000+ constraints in 20 years

• Search space: from 10^30 to 10^300,000.

[Aside: “one can encode quite a bit in 1M variables.”]

• Is this just Moore’s Law? It helped, but not much…

– 2x faster computers does not mean can solve 2x larger instances
– search difficulty does *not* scale linearly with problem size!
In fact, for O(2^n), 2x faster, handle 1 more variable!!
Mainly algorithmic progress. Memory growth also key.

• Tools: 50+ competitive SAT solvers available (e.g. Minisat solver)

• See http://www.satcompetition.org/
 Forces Driving Faster, Better SAT Solvers
Inference engines
• From academically interesting to practically relevant “Real”
benchmarks, with real interest in solving them

• Regular SAT Solver Competitions (Germany-89, Dimacs-93, China-96

SAT-02, SAT-03, …, SAT-07, SAT-09, SAT-2011)
– “Industrial-instances-only” SAT Races (2008, 2010)
– A tremendous resource! E.g., SAT Competition 2006 (Seattle):
• 35+ solvers submitted, downloadable, mostly open source
• 500+ industrial benchmarks, 1000+ other benchmarks
• 50,000+ benchmark instances available on the Internet

• Constant improvement in SAT solvers is the key to the success of, e.g
SAT-based planning, verification, and KB inference.