Cse@buffalo S.C. Shapiro Knowledge Representation and Reasoning Stuart C. Shapiro Professor, CSE Director, SNePS Research Group Member, Center for Cognitive.

S.C. Shapiro

cse@buff

alo

Knowledge Representation and Reasoning

Stuart C. ShapiroProfessor, CSE

Director, SNePS Research Group

Member, Center for Cognitive Science

S.C. Shapiro

cse@buff

alo

Introduction

S.C. Shapiro

cse@buff

alo

Long-Term Goal

• Theory and Implementation of

Natural-Language-Competent

Computerized Cognitive Agent

• and Supporting Research in

Artificial Intelligence

Cognitive Science

Computational Linguistics.

S.C. Shapiro

cse@buff

alo

Research Areas

• Knowledge Representation and Reasoning

• Cognitive Robotics

• Natural-Language Understanding

• Natural-Language Generation.

S.C. Shapiro

cse@buff

alo

Goal

• A computational cognitive agent that can:– Understand and communicate in English; – Discuss specific, generic, and “rule-like” information;– Reason;– Discuss acts and plans;– Sense;– Act;– Remember and report what it has sensed and done.

S.C. Shapiro

cse@buff

alo

Cassie

• A computational cognitive agent– Embodied in hardware– or Software-Simulated– Based on SNePS and GLAIR.

S.C. Shapiro

cse@buff

alo

GLAIR Architecture

Knowledge Level

Perceptuo-Motor Level

Sensory-Actuator Level NL

Vision

Sonar

MotionProprioception

Grounded Layered Architecture with Integrated Reasoning

SNePS

S.C. Shapiro

cse@buff

alo

SNePS• Knowledge Representation and Reasoning

– Propositions as Terms

• SNIP: SNePS Inference Package– Specialized connectives and quantifiers

• SNeBR: SNePS Belief Revision

• SNeRE: SNePS Rational Engine

• Interface Languages– SNePSUL: Lisp-Like– SNePSLOG: Logic-Like– GATN for Fragments of English.

S.C. Shapiro

cse@buff

alo

Example Cassies& Worlds

S.C. Shapiro

cse@buff

alo

BlocksWorld

S.C. Shapiro

cse@buff

alo

FEVAHR

S.C. Shapiro

cse@buff

alo

FEVAHRWorld Simulation

S.C. Shapiro

cse@buff

alo

UXO Remediation

CassieCorner flag

NonUXO object

Corner flag

UXO

Batterymeter

Corner flag

Drop-off zone

Field

Safe zone

RechargingStation

S.C. Shapiro

cse@buff

alo

Crystal Space Environment

S.C. Shapiro

cse@buff

alo

Sample Research Issues:Complex Categories

S.C. Shapiro

cse@buff

alo

Complex Categories 1

• Noun Phrases:

<Det> {N | Adj}* N

Understanding of the modification must be left to reasoning.

Example:

orange juice seat

Representation must be left vague.

S.C. Shapiro

cse@buff

alo

: Kevin went to the orange juice seat.

I understand that Kevin went to the orange juice seat.

: Did Kevin go to a seat?

Yes, Kevin went to the orange juice seat.

Complex Categories 2

S.C. Shapiro

cse@buff

alo

: Pat is an excellent teacher. I understand that Pat is an excellent teacher.

: Is Pat a teacher? Yes, Pat is a teacher.

: Lucy is a former teacher. I understand that Lucy is a former teacher.

Complex Categories 3

S.C. Shapiro

cse@buff

alo

: `former' is a negative adjective.

I understand that `former' is a negative adjective.

: Is Lucy a teacher?

No, Lucy is not a teacher.

Complex Categories 4

Also note representation and use of knowledge about words.

S.C. Shapiro

cse@buff

alo

Sample Research Issues:Indexicals

S.C. Shapiro

cse@buff

alo

Representation and Use of Indexicals

• Words whose meanings are determined by occasion of use

• E.g. I, you, now, then, here, there

• Deictic Center <*I, *YOU, *NOW>

• *I: SNePS term representing Cassie

• *YOU: person Cassie is talking with

• *NOW: current time.

S.C. Shapiro

cse@buff

alo

Analysis of Indexicals(in input)

• First person pronouns: *YOU• Second person pronouns: *I• “here”: location of *YOU• Present/Past relative to *NOW.

S.C. Shapiro

cse@buff

alo

Generation of Indexicals

• *I: First person pronouns

• *YOU: Second person pronouns

• *NOW: used to determine tense and aspect.

S.C. Shapiro

cse@buff

alo

Come here.

Use of Indexicals 1

S.C. Shapiro

cse@buff

alo

Come here.I came to you, Stu.I am near you.

Use of Indexicals 2

S.C. Shapiro

cse@buff

alo

Who am I?Your name is ‘Stu’and you are a person.

Who have you talked to?

I am talking to you.Talk to Bill.

I am talking to you, Bill.Come here.

Use of Indexicals 3

S.C. Shapiro

cse@buff

alo

Come here.

I found you.I am looking at you.

Use of Indexicals 4

S.C. Shapiro

cse@buff

alo

Come here.

I came to you.I am near you.

I found you.I am looking at you.

Use of Indexicals 5

S.C. Shapiro

cse@buff

alo

Who am I?

I talked to Stuand I am talking to you.

Your name is ‘Bill’and you are a person.

Who are you?I am the FEVAHRand my name is ‘Cassie’.

Who have you talked to?

Use of Indexicals 6

S.C. Shapiro

cse@buff

alo

Current Research Issues: Distinguishing Perceptually

Indistinguishable ObjectsPh.D. Dissertation, John F. Santore

S.C. Shapiro

cse@buff

alo

Some robots in a suite of rooms.

S.C. Shapiro

cse@buff

alo

• Are these the same two robots?• Why do you think so/not?

S.C. Shapiro

cse@buff

alo

Next Steps

• How do people do this?– Currently doing protocol experiments

• Getting Cassie to do it.

S.C. Shapiro

cse@buff

alo

Current Research Issues: Belief Revision

in aDeductively Open Belief SpacePh.D. Dissertation, Frances L. Johnson

S.C. Shapiro

cse@buff

alo

Belief Revision in a Deductively Open Belief Space

• Beliefs in a knowledge base must be able to be

changed (belief revision)

– Add & remove beliefs

– Detect and correct errors/conflicts/inconsistencies

• BUT …

– Guaranteeing consistency is an ideal concept

– Real world systems are not ideal

S.C. Shapiro

cse@buff

alo

Belief Revision in a DOBS Ideal Theories vs. Real World

• Ideal Belief Revision theories assume:– No reasoning limits (time or storage)

• All derivable beliefs are acquirable (deductive closure)

– All belief credibilities are known and fixed

• Real world– Reasoning takes time, storage space is finite

• Some implicit beliefs might be currently inaccessible

– Source/belief credibilities can change

S.C. Shapiro

cse@buff

alo

Belief Revision in a DOBS A Real World KR System

• Must recognize its limitations

– Some knowledge remains implicit

– Inconsistencies might be missed

– A source turns out to be unreliable

– Revision choices might be poor in hindsight

• After further deduction or knowledge acquisition

• Must repair itself

– Catch and correct poor revision choices

S.C. Shapiro

cse@buff

alo

Belief Revision in a DOBS Theory Example – Reconsideration

College A is better than College B. (Source: Ranking 1)

College B is better than College A. (Source: Ranking 2)

Ranking 1 is more credible that Ranking 2.

Ranking 1 was flawed, soRanking 2 is more credible than Ranking 1.Need to reconsider!

Ranking 1 is more credible that Ranking 2.

College B is better than College A. (Source: Ranking 2)

S.C. Shapiro

cse@buff

alo

Next Steps

• Implement reconsideration

• Develop benchmarks for implemented krr systems.

S.C. Shapiro

cse@buff

alo

Current Research Issues: Default Reasoning

byPreferential Ordering of Beliefs

M.S. Thesis, Bharat Bhushan

S.C. Shapiro

cse@buff

alo

Small Knowledge Base

• Birds have wings.

• Birds fly.

• Penguins are birds.

• Penguins don’t fly.

S.C. Shapiro

cse@buff

alo

KB Using Default Logic

x(Bird(x) Has(x, wings))

x(Penguin(x) Bird(x))

x(Penguin(x) Flies(x))

• Bird(x): Flies(x)

Flies(x)

S.C. Shapiro

cse@buff

alo

KB Using Preferential Ordering

x(Bird(x) Has(x, wings))

x(Penguin(x) Bird(x))

x(Penguin(x) Flies(x))

x(Bird(x) Flies(x))

• Precludes(x(Penguin(x) Flies(x)),

x(Bird(x) Flies(x)))

S.C. Shapiro

cse@buff

alo

Next Steps

• Finish theory and implementation.

S.C. Shapiro

cse@buff

alo

Current Research Issues: Representation & Reasoning

with Arbitrary ObjectsStuart C. Shapiro

S.C. Shapiro

cse@buff

alo

Classical Representation

• Clyde is gray.– Gray(Clyde)

• All elephants are gray. x(Elephant(x) Gray(x))

• Some elephants are albino. x(Elephant(x) & Albino(x))

• Why the difference?

S.C. Shapiro

cse@buff

alo

Representation Using Arbitrary & Indefinite Objects

• Clyde is gray.– Gray(Clyde)

• Elephants are gray.– Gray(any x Elephant(x))

• Some elephants are albino.– Albino(some x Elephant(x))

S.C. Shapiro

cse@buff

alo

Subsumption Among Arbitrary & Indefinite Objects

(any x Elephant(x))

(any x Albino(x) & Elephant(x))

(some x Albino(x) & Elephant(x))

(some x Elephant(x))If x subsumes y, then P(x) P(y)

S.C. Shapiro

cse@buff

alo

Example (Runs in SNePS 3)Hungry(any x Elephant(x)

& Eats(x, any y Tall(y)

& Grass(y)

& On(y, Savanna)))

Hungry(any u Albino(u)

& Elephant(u)

& Eats(u, any v Grass(v)

& On(v, Savanna)))

S.C. Shapiro

cse@buff

alo

Next Steps

• Finish theory and implementation of arbitrary and indefinite objects.

• Extend to other generalized quantifiers– Such as most, many, few, no, both, 3 of, …

S.C. Shapiro

cse@buff

alo

For More Information

• Shapiro: http://www.cse.buffalo.edu/~shapiro/

• SNePS Research Group: http://www.cse.buffalo.edu/sneps/