Learning and Joint Deliberation through Argumentation in MAS Santi - - PowerPoint PPT Presentation

▶
learning and joint deliberation through argumentation in
SMART_READER_LITE
LIVE PREVIEW

Learning and Joint Deliberation through Argumentation in MAS Santi - - PowerPoint PPT Presentation

Dec 12, 2022 355 likes •824 views

Learning and Joint Deliberation through Argumentation in MAS Santi Ontan (GeorgiaTech) Enric Plaza (IIIA-CSIC) dijous 12 de novembre de 2009 1 Outline Introduction Justified Predictions in MAS Arguments and Counterexamples

slide-1
SLIDE 1

Learning and Joint Deliberation through Argumentation in MAS

Santi Ontañón (GeorgiaTech) Enric Plaza (IIIA-CSIC)

1 dijous 12 de novembre de 2009

slide-2
SLIDE 2

Outline

  • Introduction
  • Justified Predictions in MAS
  • Arguments and Counterexamples
  • Argumentation-based MAL
  • Experimental Evaluation
  • Conclusions & Future Work

2 dijous 12 de novembre de 2009

slide-3
SLIDE 3

Committee

3 dijous 12 de novembre de 2009

slide-4
SLIDE 4

Committee

Input Deliberation Aggregation Output

Committee: (1)A group of people officially delegated (elected or appointed) to perform a function, such as investigating, considering, reporting, or acting on a matter. Team: A number of persons associated in some joint action. A group

  • rganized to work together.

Coalition: A combination or alliance, esp. a temporary one between persons, factions, states, etc. An alliance for combined action, especially a temporary alliance of political parties.

3 dijous 12 de novembre de 2009

slide-5
SLIDE 5

Ensemble Effect

Data Set Bagging / Boosting / other... P1 P2 Pn-1 Pn ... Aggregation (e.g. Voting) Ensemble effect Joint Prediction Predictors Joint prediction better than best Pi

4 dijous 12 de novembre de 2009

slide-6
SLIDE 6

Committee of agents

DS 1 A1 A2 An-1 An ... Voting Ensemble effect Joint Prediction Agents Joint prediction better than best Ai DS n-1 DS 2 DS n

5 dijous 12 de novembre de 2009

slide-7
SLIDE 7

Deliberation + Voting

A1 A2 An-1 An ... Agents Deliberation (Argumentation framework)

6 dijous 12 de novembre de 2009

slide-8
SLIDE 8

Deliberation + Voting

A1 A2 An-1 An ... Agents Deliberation (Argumentation framework) Joint Prediction Agreement Voting

6 dijous 12 de novembre de 2009

slide-9
SLIDE 9

Introduction

  • CBR agents
  • solve problems and learn from them
  • How could CBR agents collaborate?
  • solving and/or learning from cases
  • Argumentation process
  • to mediate agent collaboration
  • achieving a joint prediction

7 dijous 12 de novembre de 2009

slide-10
SLIDE 10

Learning vs Cooperating

  • Why to learn in problem solving?
  • to improve accuracy, range, etc
  • Why to cooperate in problem solving?
  • to improve accuracy, range, etc
  • Learning-cooperation continuum
  • learning agents that cooperate by arguing

8 dijous 12 de novembre de 2009

slide-11
SLIDE 11

CBR cycle

Retrieve Reuse

New Case Problem Solved Case

Revise R e t a i n

Revised Case Precedent Case Domain Knowledge New Case Retrieved Case Retrieved Case 9 dijous 12 de novembre de 2009

slide-12
SLIDE 12

Argumentation in Multi-Agent Learning

  • An argumentation framework for learning agents
  • Justified Predictions as arguments
  • Individual policies for agents to
  • generate arguments and
  • generate counterarguments
  • select counterexamples

10 dijous 12 de novembre de 2009

slide-13
SLIDE 13

Justified Prediction

Justification: A symbolic description with the information relevant to determine a specific prediction

Problem Traffic_light: red Cars_passing: no Case 1 Traffic_light: red Cars_passing: no Solution: wait Case 3 Traffic_light: red Cars_passing: yes Solution: wait Case 4 Traffic_light: green Cars_passing: yes Solution: wait Case 2 Traffic_light: green Cars_passing: no Solution: cross Retrieved cases Solution: wait Justification Traffic_light: red

11 dijous 12 de novembre de 2009

slide-14
SLIDE 14

Justification example

Solution: hadromerida Justification: D1 Sponge Spikulate skeleton External features External features Gemmules: no Spikulate Skeleton Megascleres Uniform length: no Megascleres Smooth form: tylostyle

Case Base

  • f A1

LID

New sponge

P

!!

" ! !!"" #" $%&'()*'+&%" ,""

12 dijous 12 de novembre de 2009

slide-15
SLIDE 15

Justification example

Solution: hadromerida Justification: D1 Sponge Spikulate skeleton External features External features Gemmules: no Spikulate Skeleton Megascleres Uniform length: no Megascleres Smooth form: tylostyle

Case Base

  • f A1

LID

New sponge

P

!!

" ! !!"" #" $%&'()*'+&%" ,""

The predicted solution is hadromerida because the smooth form of the megascleres of the spiculate skeleton of the sponge is of type tylostyle, the spikulate skeleton of the sponge has not uniform length, and there are no gemmules in the external features of the sponge.

12 dijous 12 de novembre de 2009

slide-16
SLIDE 16

Counterargument generation

Sponge Spikulate skeleton External features External features Gemmules: no Growing: Spikulate Skeleton Megascleres Uniform length: no Megascleres Smooth form: tylostyle Growing Grow: massive

Case Base

  • f A2

LID

!!

" ! !!"" #" $%&'()*'+&%" ,""

Solution: astrophorida Justification: D2

""

# ! !!#" #" %-.'(/$('+&%" ,#"

13 dijous 12 de novembre de 2009

slide-17
SLIDE 17
  • Justified Prediction: An argument α

endorsing a individual prediction

  • Counterargument: An argument β
  • ffered in opposition to an

argument α

  • Counterexample: A case c

contradicting an argument α

Argument types

α = Ai,P,+,D

β = A2,P,−,D2

c = P1,−

14 dijous 12 de novembre de 2009

slide-18
SLIDE 18

Case-based Confidence

+ + + + + +

  • +

+

! ! !!!" #" "" $" #$%& '$%& () $*&+, -!

#"!.!/ !

! !"#"# ! 0%1

15 dijous 12 de novembre de 2009

slide-19
SLIDE 19

Preference on Arguments

+

?

+

  • +

+ -

α = Ai,P,+,D

Y N

C(α) = Y Y + N

Confidence on an argument based on cases

16 dijous 12 de novembre de 2009

slide-20
SLIDE 20

Preference on Arguments(2)

P

+

  • ?

α β

+ - + +

  • +

C(α) = 4 5 = 0.8 C(β) = 2 3 = 0.66

17 dijous 12 de novembre de 2009

slide-21
SLIDE 21

Preference on Arguments(2)

P

+

  • ?

α β

+ - + +

  • +

C(α) = 4 5 = 0.8 C(β) = 2 3 = 0.66

Preferred

17 dijous 12 de novembre de 2009

slide-22
SLIDE 22

Preference on Arguments(2)

P

+

  • ?

α β

+ - + +

  • +

C(α) = 4 5 = 0.8 C(β) = 2 3 = 0.66 C(α) = Y A1

α

+ Y A2

α

+ 1 Y A1

α

+ Y A2

α

+ N A1

α

+ N A2

α

+ 2

Joint Confidence Preferred

17 dijous 12 de novembre de 2009

slide-23
SLIDE 23

Relations between arguments

P P

P

a) b) c)

+ + + +

  • ?

? ?

α = A1,P,+,D1 β = A2,P,+,D2 β = A2,P,−,D2 α = A1,P,+,D1 α = A1,P,+,D1 β = A2,P,−,D2

α α α

β β β

18 dijous 12 de novembre de 2009

slide-24
SLIDE 24

Relations between arguments

P P

P

a) b) c)

+ + + +

  • ?

? ?

α = A1,P,+,D1 β = A2,P,+,D2 β = A2,P,−,D2 α = A1,P,+,D1 α = A1,P,+,D1 β = A2,P,−,D2

α α α

β β β

Consistent Incomparable Counterargument

18 dijous 12 de novembre de 2009

slide-25
SLIDE 25

Relations between cases and justified predictions

+

?

c = P1,+ c = P1,− P P P c = P1,−

a) b) c)

+ + +

  • ?

?

α = Ai,P,+,D α = Ai,P,+,D α = Ai,P,+,D

19 dijous 12 de novembre de 2009

slide-26
SLIDE 26

Relations between cases and justified predictions

+

?

c = P1,+ c = P1,− P P P c = P1,−

a) b) c)

+ + +

  • ?

?

α = Ai,P,+,D α = Ai,P,+,D α = Ai,P,+,D

Endorsing case Irrelevant case Counterexample

19 dijous 12 de novembre de 2009

slide-27
SLIDE 27

Argument Generation

  • Generation of a Justified Prediction
  • LID generates a description α.D subsuming P
  • Generation of a Counterargument
  • if no Counterargument can be generated then:
  • Selection of a Counterexample

20 dijous 12 de novembre de 2009

slide-28
SLIDE 28

Counterargument Generation

  • Counterarguments are generated based
  • n the specificity criterion
  • LID generates a description β.D

subsuming P and subsumed by α.D

P

+

  • ?

α β

21 dijous 12 de novembre de 2009

slide-29
SLIDE 29

Selection of a Counterexample

c = P1,− P

+

  • ?

α = Ai,P,+,D

  • Select a case c subsumed by α.D and

endorsing a different solution class.

22 dijous 12 de novembre de 2009

slide-30
SLIDE 30

AMAL protocol

assert(α) Ht = αt

1, ..., αt n

Justified prediction asserted in the next round Assertions of n agents at round t Agent states a counterargument ß

rebut(β, α) contradict(αt

i) = {α ∈ Ht|α.S = αt i.S}

Set of contradicting arguments for agent Ai at round t (those predicting a different solution)

23 dijous 12 de novembre de 2009

slide-31
SLIDE 31

Agents assert t All Agree? Joint Solution YES Generates CA

NO

Agent+token Generates CE Agent'+CE New arg DELIBERATION at Round t & Agent owning the token Nobody new args Voting α Better CA than ? α Agent asserts CA

YES

Rebut CA to Agent'

NO

Better CA than ? α Agent' asserts CA

YES NO

24 dijous 12 de novembre de 2009

slide-32
SLIDE 32

Argument Generation

contradict(αt

i) = {α ∈ Ht|α.S = αt i.S}

Generate CA for each β1...βk Select argument with a generated CA that has lowest confidence May not found a CA for each If empty generates CE βi Most likely to "convince" the

  • ther agent to change assertion

αi Select CA for that argument

25 dijous 12 de novembre de 2009

slide-33
SLIDE 33

Confidence-weighted Voting

  • Each argument in Ht is a vote for an alternative
  • Each vote is weighted by the joint confidence

measure of that argument

  • Weighted voting: alternative with higher

aggregated confidence wins

S = arg max

Sk∈S

  • αi∈Ht|αi.S=Sk

C(αi) in order to avoid infinite iterations, if an agent

26 dijous 12 de novembre de 2009

slide-34
SLIDE 34

Experiments

  • Designed to validate 2 hypotheses
  • average of 5 10-fold cross validation runs
  • (H1) that argumentation is a useful framework for

joint deliberation and can improve over other typical methods such as voting; and

  • (H2) that learning from communication improves

the individual performance of a learning agent participating in an argumentation process.

27 dijous 12 de novembre de 2009

slide-35
SLIDE 35

77 80 83 86 89 92 2 3 4 5

81,28 81,28 81,28 81,28 89,79 87,64 87,21 82,21 91,21 90,14 88,60 88,42 91,43 90,50 90,43 88,64

AMAL JV Voting Individual

#Agents Accuracy Sponges Data set

Accuracy after Deliberation

28 dijous 12 de novembre de 2009

slide-36
SLIDE 36

Accuracy after Deliberation

52 58 64 70 76 82 2 3 4 5

60,59 60,59 60,59 60,59 75,18 71,47 68,99 61,04 77,85 75,70 70,88 66,77 81,43 80,66 73,68 70,62

AMAL JV Voting Individual

#Agents Accuracy Soybean Data set

29 dijous 12 de novembre de 2009

slide-37
SLIDE 37

Individual Learning from Communication

60 66 72 78 84 25% 40% 55% 70% 85% 100%

63,57 63,57 63,57 63,57 63,57 63,57 63,57 68,43 73,00 76,36 77,86 79,21 63,57 72,79 76,71 78,93 80,57 83,14

LFC L NL

%cases Accuracy Sponges Data set

30 dijous 12 de novembre de 2009

slide-38
SLIDE 38

Individual Learning from Communication

32 40 48 56 64 72 25% 40% 55% 70% 85% 100%

33,75 33,75 33,75 33,75 33,75 33,75 33,75 38,50 45,93 52,96 55,63 59,93 33,75 49,38 58,44 65,34 69,06 70,62

LFC L NL

%cases Accuracy Soybean Data set

31 dijous 12 de novembre de 2009

slide-39
SLIDE 39

Case Base Size

30 60 90 120 150 180 210 240 270 300 Sponges Soybean

276,00 252,00 55,26 50,40 87,15 58,96

LFC L Total

#cases

23.4% 20% 31.58% 20.02%

32 dijous 12 de novembre de 2009

slide-40
SLIDE 40

30 60 90 120 150 180 210 240 270 300 Sponges Soybean

276,00 252,00 55,26 50,40 87,15 58,96

LFC L Total

#cases

23.4% 20% 31.58% 20.02%

33 dijous 12 de novembre de 2009

slide-41
SLIDE 41

Learning from a few good cases while arguing

32 40 48 56 64 72 25% 40% 55% 70% 85% 100%

33,75 33,75 33,75 33,75 33,75 33,75 33,75 38,50 45,93 52,96 55,63 59,93 33,75 49,38 58,44 65,34 69,06 70,62

LFC L NL

%cases Accuracy Soybean Data set

34 dijous 12 de novembre de 2009

slide-42
SLIDE 42

Learning from a few good cases while arguing

32 40 48 56 64 72 25% 40% 55% 70% 85% 100%

33,75 33,75 33,75 33,75 33,75 33,75 33,75 38,50 45,93 52,96 55,63 59,93 33,75 49,38 58,44 65,34 69,06 70,62

LFC L NL

%cases Accuracy Soybean Data set

34 dijous 12 de novembre de 2009

slide-43
SLIDE 43

Conclusions

  • An argumentation framework for learning agents
  • a case-based preference relation over arguments,
  • by computing a confidence estimation of arguments
  • a case-based policy to generate counter-

arguments and select counterexamples

  • an argumentation-based approach for learning

from communication

35 dijous 12 de novembre de 2009

slide-44
SLIDE 44

Future Work

  • Framework for agents using induction
  • better policies for generating CA
  • collaborative search in generalization space
  • Deliberative Agreement
  • for social choice and collective judgment models
  • aggregation procedures of sets of interconnected

judgments

  • deliberation over sets of interconnected judgments

36 dijous 12 de novembre de 2009

slide-45
SLIDE 45

AMAL vs Centralized

75,0000 79,0175 83,0350 87,0525 91,0700 1 2 3 4 5

AMAL Centralized

#Agents Accuracy Sponges Data set

37 dijous 12 de novembre de 2009