70
CMU DESIGN GOALS { Kevin T. Kelly , Hanti Lin } Carnegie Mellon University Responsi ve-ness CMU

{ Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Embed Size (px)

Citation preview

Page 1: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

CMU DESIGN GOALS

{ Kevin T. Kelly , Hanti Lin }Carnegie Mellon University

Responsive-ness

CMU

Page 2: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Page 3: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Page 4: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Probabilistic

conditioning

Page 5: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Probabilistic

conditioning

Acceptance

Page 6: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Probabilistic

conditioning

AcceptancePropositional

belief revision

Page 7: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Probabilistic

conditioning

AcceptancePropositional

belief revision

Acceptance

Page 8: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Qualitative Reasoning that Tracks Conditioning

Probabilistic

conditioning

AcceptancePropositional

belief revision

Acceptance=

Conditioning + acceptance = acceptance + revision

Page 9: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Pre-established Harmony

Acceptance

Propositional

belief revisio

n

Probabilistic

conditioning

Page 10: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Cheap Bayes With Harmony

Acceptance

Tie shoes?

Probabilistic

conditioning

Eat breakfast?

Get out of bed?

Page 11: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

When You Need Bayes…

Acceptance

Probabilistic

conditioning

Help! Bayes!

Invest?

Tie shoes?

Eat breakfast?

Get out of bed?

Page 12: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Call Him Then

Acceptance

Condition only once

Tie shoes?

Eat breakfast?

Get out of bed?

Invest?

Page 13: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Call Him Then

Acceptance

Thanks.I’ll take it from here

Tie shoes?

Eat breakfast?

Get out of bed?

Invest?Condition only once

TV?

Page 14: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Expensive Bayes Without Harmony

Acceptance

Repeated conditioning

Tie shoes?

Eat breakfast?

Get out of bed?

Invest?

TV?

Page 15: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Cheap Bayes with Harmony

Acceptance

Tie shoes?

Eat breakfast?

Get out of bed?

Invest?Condition only once

TV?

Page 16: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

LMU Design Principle: Steadiness

Steadiness = “Just conjoin the new data with

your old propositions if the two are consistent”

EB

LMU

Page 17: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

AGM is Steady

B C

A

Page 18: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

AGM is Steady

C

A

Page 19: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Non-steady Revision Rule

A

B C

Yoav Shoham

Page 20: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Non-steady Revision Rule

A

C

Yoav Shoham

Page 21: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Non-steady Revision Rule

A C

Yoav Shoham

Page 22: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Some Shared Design Principles

LMU

CMU

Page 23: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Consistency

Inconsistency is accepted nowhere.

Page 24: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Non-skepticism

Every atom A is accepted over some open neighborhood.

Page 25: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Non-OpinionationThere is an open neighborhood over which you accept a non-atom and nothing stronger.

A v B

Page 26: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Corner-monotonicity

C

If an atom is accepted, it continues to be accepted along the straight line to the corresponding corner.

Page 27: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

C

CC

C

C

Corner-monotonicity

If an atom is accepted, it continues to be accepted along the straight line to the corresponding corner.

Page 28: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Sensible Rules

Sensible = all four properties.

C

CC

C

C

A v B

Page 29: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Both are Sensible!

A v C

A

CB v C

T

B

A v B A v C

A

CB v C

T

B

A v B

LMU CMU

Page 30: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Incompatibility Theorem

No sensible acceptance rule is both steady and tracks conditioning.

consumer designer

Sorry. You can’t have both.

Page 31: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

A New Paradox of AcceptanceA

B C

A

A v B

p

p(.|A v B)

Page 32: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

A New Paradox of AcceptanceA

B C

A

A v B

p

p(.|A v B)

Accept A.

Learn its consequence A v B.

If you track, you retract A!

Page 33: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

“Cautious” Monotonicity= Hypothetico-Deductive Monotonicity

If you accept a hypothesis, don’t retract it when you learn what it entails (i.e. predicts).

Page 34: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

A Better Idea?

A v B A v C

A

B CB v C

T

0.8

0.9

Page 35: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Another New Paradox of Acceptance

p

A

B

Page 36: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Another New Paradox of AcceptanceA

B

B

p

p(.|B)

Page 37: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Another New Paradox of AcceptanceA

B

A

p

p(.|B)

Page 38: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Another New Paradox of AcceptanceA

B

A

B

T

p

p(.|B)

p(.|B)

You will accept A v B no matter whether B or B is learned.

But if you track, you don’t accept A v B.

Page 39: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Case Reasoning

Accept a hypothesis, if you will accept it no matter whether E is learned or E is learned.

Page 40: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Theorem

The CMU rule + Shoham revision (non-steady) satisfies:

sensible tracks conditioning avoids both new paradoxes

Page 41: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Partial Converse

Shoham revision sensible tracks conditioningImpliesCMU rule + avoidance of the 2 new paradoxes.

Page 42: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Gettier Without False Lemmas

Nogot

Nobody

Somebody

Gettier case

Havit= the Truth

Page 43: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

CMU Rule Represents it

Havit= the Truth

Nogot

Nobody

Somebody

Page 44: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

CMU Rule is Unsteady!

HavitNogot

Nobody

Somebody

“Somebody”is retracted but not refuted.

Page 45: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Gettier/Unsteadiness Zones

HavitNogot

Nobody

Somebody

Page 46: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Shoham Revision vs. AGM Revision

Havit

Nogot Havit

Nobody

Nogot

Nobody

Page 47: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Shoham Revision vs. AGM Revision

“Trust what you accepted”

“Re-examine your reasons”Havit

Nogot Havit

Nobody

Nogot

Nobody

Page 48: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Structure Preservation

(0, 1, 0)

(0, 0, 1)(1, 0, 0)

(1/3, 1/3, 1/3)

LogicGeometry

A CB

Acpt

Page 49: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Some Clear CasesA

B C

Page 50: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

InterpolationA

B C

Page 51: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

What About Here?A

B C

Page 52: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Probability Lives in the Unit Cube111

100 010 001

000

011110 101

Page 53: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Classical Logic Lives on the Corners111

100 010 001

000

011110 101

Page 54: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

But What if Logic Filled the Cube?111

100 010 001

000

011110 101

Page 55: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Classical Negation111

100 010 001

000

011110 101

Page 56: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Partial Negation111

100 001

000

011110

010

101

Page 57: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

GeologicClose classical logic underPartial negation

Page 58: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Geological EntailmentLogical Closure =Sub-crystals

Page 59: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Probability is a Surface in Geologic

Page 60: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Classical Principle of Indifference

Page 61: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Principle of Indifference Completed

Page 62: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Probalogic = Projection of Geologic

Page 63: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Probalogic as Geologic in Perspective

Page 64: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Probalogic as Geologic in Perspective

Page 65: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Projection of Geological Consequences

Page 66: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

= Probalogic

Page 67: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Acceptance Should Preserve Logical Structure

Acpt

Page 68: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Representation TheoremThe CMU rule is the only rule that preserves logical structure (entailment, disjunction and consistent conjunction).

Acpt

Page 69: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

Feature Checklist for the CMU Rule

The CMU rule + Shoham revision satisfies

sensible tracks conditioning avoids both new paradoxes represents no-false-lemma Gettier cases unique geo-logical representation

Page 70: { Kevin T. Kelly, Hanti Lin } Carnegie Mellon University Responsive -ness CMU

THANK YOU!

The CMU rule + Shoham revision satisfies

sensible tracks conditioning avoids both new paradoxes represents no-false-lemma Gettier cases unique geo-logical representation