Efficient Sharing of Conflicting Opinions with Minimal Communication in Large Decentralised Teams

Introduction Model AAT Experiments and Results Conclusions

Efficient Sharing of Conflicting Opinions withMinimal Communication

in Large Decentralised Teams

Oleksandr Pryymak, Alex Rogers and Nicholas R. Jennings

University of Southampton

{op08r,acr,nrj}@ecs.soton.ac.uk

July 20, 2011

0 / 19


Disaster response and large decentralised teams

2010, Haiti earthquake – Citizen and publicnews reporting, plotted on an online map(Ushahidi).

2010, Chile earthquake – Twitter is one ofthe speediest, albeit not the most accurate,sources of real-time information (France24).

Large teams of individuals

Decentralised

Not every individual can make anobservation

Observations are uncertain and conflicting

Individuals share opinions withoutsupporting information

How opinions are shared and how to improve their accuracy?

1 / 19



2010, Haiti earthquake – Citizen and publicnews reporting, plotted on an online map(Ushahidi).2010, Chile earthquake – Twitter is one ofthe speediest, albeit not the most accurate,sources of real-time information (France24).


Decentralised





1 / 19





Decentralised





1 / 19





Decentralised





1 / 19





Decentralised





1 / 19





Decentralised





1 / 19





Decentralised





1 / 19





Decentralised




How opinions are shared and how to improve their accuracy?1 / 19


How opinions are shared – Can we trust what we share?

Can we trust what we share?

Chile’10 : yes / no (Mendoza et al.2010)

Santiago airport is closed

Fire at the University ofConceptcion

Looting in Conceptcion

Looting in Santiago

Tsunami warning

Active volcano

Opinions are shared in cascades (avalanches)

Even in cooperative settings opinions might be incorrect

2 / 19








Looting in Santiago

Tsunami warning

Active volcano



2 / 19








Looting in Santiago

Tsunami warning

Active volcano



2 / 19



Can we trust what we share?Chile’10 : yes / no (Mendoza et al.2010)




Looting in Santiago

Tsunami warning

Active volcano



2 / 19







Looting in Santiago

Tsunami warning

Active volcano



2 / 19







Looting in Santiago

Tsunami warning

Active volcano



2 / 19







Looting in Santiago

Tsunami warning

Active volcano



2 / 19


Problem of Forming a Correct Opinion

How do agents make a decision which opinion is correct?based on own priors, observationsbased on information from others

by analysing communicated informationreaching agreements interactivity with others

The Problem

However, if:

agents’ processing abilities are limited

communication is strictly limited to opinion sharing

The Solution

Agents have to exploit properties of opinion sharing dynamics, andfilter out incorrect opinions in the sharing process

How to find such settings by independent actions of the agents?

3 / 19





The Problem

However, if:



The Solution



3 / 19





The Problem

However, if:



The Solution



3 / 19





The Problem

However, if:



The Solution



3 / 19





The Problem

However, if:



The Solution


How to find such settings by independent actions of the agents?3 / 19


Outline

Remaining sections:

1 Model of opinion sharing

2 Existing message-passing algorithm

3 Our algorithm based on independent actions

4 Evaluation

4 / 19


Model – an Agent

Agent

5 / 19


Model – an Agent

Will it rain tonight?

Subject ofinterest

Agent

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief1Prior0

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief

Own observations

1Prior0

Updated with:sensors

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief

Own observations

1Prior0


5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief

Own observations

Opinions' of othersYes No? No...

1Prior0


network neighbours

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief

Own observations


1Prior0


network neighbours

5 / 19


Model – an Agent


YesNo Don't know

Subject ofinterest

Opinion

Agent

Belief

Own observations


1Prior0


network neighbours

5 / 19


Model – Sample Dynamics

red nodes are agents with sensors;

green nodes are agents with undeter. opinion;

white and black are agents that support thecorresponding opinions. (b = white)

opinions are shared incascades

cascades might bewrong and fragile

cascades depend ontrust levels

double counting fallacy

6 / 19










6 / 19










6 / 19










6 / 19










6 / 19


Settings for Improved Reliability – Metrics

0.55 0.6 0.65 0.7 0.750

20

40

60

80

100Agentsholdingopinion,%

tcritical

correctincorrectundetermined

0.55 0.6 0.65 0.7 0.750

0.2

0.4

0.6

0.8

1

Reliability

Trust level (common for all agents)

tcritical

Reliability

Awareness

Scale-Invariant dynamicsStable dynamics Unstable dynamics

7 / 19


Cascades Distribution

100 101 102 103100

101

102

103

104t=0.6

100 101 102 103100

101

102

103

104t=0.63

100 101 102 103100

101

102t=0.66

Cas

cade

Fre

quen

cy

Size of Opinion Cascade

Cas

cade

Fre

quen

cy


Cas

cade

Fre

quen

cy


Scale-Invariant DynamicsStable Dynamics Unstable Dynamics

Branching factor of opinion sharing

αimproved reliability = 1

R. Glinton, P. Scerri, and K. Sycara. (2010)

Exploiting scale invariant dynamics for efficient information propagation in large teams.In Proceedings of 9th International Conference on Autonomous Agents and Multiagent Systems(AAMAS’10), pages 21-28, Toronto, Canada.

8 / 19


Cascades Distribution

100 101 102 103100

101

102

103

104t=0.6

100 101 102 103100

101

102

103

104t=0.63

100 101 102 103100

101

102t=0.66

Cas

cade

Fre

quen

cy


Cas

cade

Fre

quen

cy


Cas

cade

Fre

quen

cy


Scale-Invariant DynamicsStable Dynamics Unstable Dynamics

Branching factor of opinion sharing

αimproved reliability = 1

R. Glinton, P. Scerri, and K. Sycara. (2010)

Exploiting scale invariant dynamics for efficient information propagation in large teams.In Proceedings of 9th International Conference on Autonomous Agents and Multiagent Systems(AAMAS’10), pages 21-28, Toronto, Canada.

8 / 19


DACORYes

Yes

?

?

?

?

α

introduces additional communication〈NumberOfNeighbours〉2 additional messages for a singleopinion change

exhibits low adaptivityrequires tuning of its parameters

9 / 19


DACORYes

Yes

?

?

?

?

α

introduces additional communication〈NumberOfNeighbours〉2 additional messages for a singleopinion change

exhibits low adaptivityrequires tuning of its parameters

9 / 19


Autonomous Adaptive Tuning of Trust Levels

How to find the settings for improved reliability based on localobservations only?

Intuition

An agent must use the minimal trust level that still enables it toform its opinion

However, the agent’s choice influences others in the team

10 / 19




0.55 0.6 0.65 0.7 0.750

0.2

0.4

0.6

0.8

1

Rel

iabi

lity


tcritical

Reliability

Awareness


Intuition



10 / 19




0.55 0.6 0.65 0.7 0.750

0.2

0.4

0.6

0.8

1

Rel

iabi

lity


tcritical

Reliability

Awareness


Intuition



10 / 19




0.55 0.6 0.65 0.7 0.750

0.2

0.4

0.6

0.8

1

Rel

iabi

lity


tcritical

Reliability

Awareness


Intuition


However, the agent’s choice influences others in the team10 / 19



Agent i has to select minimal trust level t li from the candidates.

The agent with t li has to achieve the target awareness rate, hbest

ti = arg mint li

|hi (t li )− hbest|

1 How to select candidate trust levels?

2 How to estimate their awareness rates?

3 How to choose the trust level to use?

11 / 19



Agent i has to select minimal trust level t li from the candidates.The agent with t li has to achieve the target awareness rate, hbest

ti = arg mint li





11 / 19




ti = arg mint li





11 / 19




ti = arg mint li





11 / 19


AAT – Candidate Trust Levels

1P'i0 σ1-σ

Pki

oi =bl

ack

k=3

o i =white

k=12

To form the most accurate opinion the agent must form its opinionwhen it observes the strongest support.

Since the number of neighbours |Ni | is limited, the set of thecandidate trust levels is:

Ti = {t l−i , t l+i : l = 1 . . . |Ni |}

In the settings of dynamic topology and agent may use arbitrary Ti

12 / 19



1P'i0 σ1-σ 0.5

Pki

oi =bl

ack

k=1

ti1− ti

1+o i

=white

k=1

1P'i0 σ1-σ 0.5

Pki

ti2− ti

2+



Ti = {t l−i , t l+i : l = 1 . . . |Ni |}


12 / 19



1P'i0 σ1-σ 0.5

Pki

oi =bl

ack

k=1

ti1− ti

1+o i

=white

k=1

1P'i0 σ1-σ 0.5

Pki

ti2− ti

2+



Ti = {t l−i , t l+i : l = 1 . . . |Ni |}


12 / 19



1P'i0 σ1-σ 0.5

Pki

oi =bl

ack

k=1

ti1− ti

1+o i

=white

k=1

1P'i0 σ1-σ 0.5

Pki

ti2− ti

2+



Ti = {t l−i , t l+i : l = 1 . . . |Ni |}


12 / 19


AAT – Estimation of the Awareness Rates

The awareness rates of the candidate trust levels cannot becalculated.

There are two evidences that indicate that agent could haveformed an opinion with t li actually using ti :

1 Ev1: If an opinion was formed, then all higher trust levels(t li ≥ ti ) would have led to opinion formation as well.

2 Ev2: Otherwise, if t li requires less updates to form an opinionthen the observed strongest support.

hi (tli ) ≈ hi (t

li )

13 / 19



The awareness rates of the candidate trust levels cannot becalculated.There are two evidences that indicate that agent could haveformed an opinion with t li actually using ti :



hi (tli ) ≈ hi (t

li )

13 / 19






hi (tli ) ≈ hi (t

li )

13 / 19






hi (tli ) ≈ hi (t

li )

13 / 19






hi (tli ) ≈ hi (t

li )

13 / 19


AAT – Strategies to Select a Trust Level

The problem of selecting t li ∈ Ti , accordingly their h(t li ), resemblesthe standard multi-armed bandit (MAB) model.

The agent can apply MABstrategies:

Greedy

ε-greedy

ε-N-greedy

Soft-max

– assume that rewarddistribution is unknown.

However, for ascendantly ordered Ti :

0.55 0.6 0.65 0.70

0.2

0.4

0.6

0.8

1

Aw

aren

ess

Rat

e

Trust Leveltcritical

Hill-climbing: Select a trust level fromthe closest to the currently used

Since an agent’s choice influences others, strategies with lessdramatic changes to the dynamics are expected to perform better.

14 / 19





Greedy

ε-greedy

ε-N-greedy

Soft-max



0.55 0.6 0.65 0.70

0.2

0.4

0.6

0.8

1

Aw

aren

ess

Rat

e

Trust Leveltcritical



14 / 19





Greedy

ε-greedy

ε-N-greedy

Soft-max



0.55 0.6 0.65 0.70

0.2

0.4

0.6

0.8

1

Aw

aren

ess

Rat

eTrust Level

tcritical



14 / 19





Greedy

ε-greedy

ε-N-greedy

Soft-max



0.55 0.6 0.65 0.70

0.2

0.4

0.6

0.8

1

Aw

aren

ess

Rat

eTrust Level

tcritical



14 / 19


Selection of the Target Awareness Rate

0.8 0.85 0.9 0.95 10.6

0.7

0.8

0.9

1

Target awareness rate, hbest

Reliability

0.8 0.85 0.9 0.95 10.6

0.65

0.7

0.75


Averagetrustlevel,⟨ti⟩

random scalefree smallworld

The agents have to compromise their awareness rates to improveteam’s reliability.

With a high target awareness rate, hbest, a team exhibits unstabledynamics, thus the reliability drops.

15 / 19


Selection of the Target Awareness Rate

0.8 0.85 0.9 0.95 10.6

0.7

0.8

0.9

1


Reliability

0.8 0.85 0.9 0.95 10.6

0.65

0.7

0.75


Averagetrustlevel,⟨ti⟩

random scalefree smallworld

The agents have to compromise their awareness rates to improveteam’s reliability.With a high target awareness rate, hbest, a team exhibits unstabledynamics, thus the reliability drops. 15 / 19


Reliability of a Team

500 1000 1500 20000.5

0.6

0.7

0.8

0.9

1(a) Random Network

Rel

iabi

lity

Network Size

AATDACORPre-tuned Trust LevelsAverage Pre-tunedTrust Levels

AAT significantly outperforms prediction of the best parameters(average pre-tuned) and existing DACOR. Individually pre-tunedtrust levels indicate on the upper-bound that can be achieved.

16 / 19



500 1000 1500 20000.5

0.6

0.7

0.8

0.9

1(b) Scale−Free Network

Rel

iabi

lity

Network Size



16 / 19



500 1000 1500 20000.5

0.6

0.7

0.8

0.9

1(c) Small−World Network

Rel

iabi

lity

Network Size



16 / 19


Communication Expense

MinimalCommunication =∑

Agents

NumberOfNeighbours

AAT is communicationally efficient while DACOR requires 4-7times more messages to operate

17 / 19


Communication Expense

MinimalCommunication =∑

Agents

NumberOfNeighbours

500 1000 1500 20000

20

40

60

80

Mes

sage

s pe

r Age

nt

Network Size

DACOR

Minimal Communication

AAT

AAT is communicationally efficient while DACOR requires 4-7times more messages to operate

17 / 19


Performance in the Presence of Indifferent Agents

0 20 40 60 80 1000.5

0.6

0.7

0.8

0.9

1(a) Random Network

Rel

iabi

lity

% of Indifferent Agents


AAT installed on a half of a team delivers higher reliability than wecan predict by using the average pre-tuned trust-levels.

18 / 19



0 20 40 60 80 1000.4

0.5

0.6

0.7

0.8

0.9

1(b) Scale−Free Network

Rel

iabi

lity




18 / 19



0 20 40 60 80 1000.4

0.5

0.6

0.7

0.8

0.9

1(c) Small−World Network

Rel

iabi

lity




18 / 19


Conclusions

AAT exploits properties of social behaviour to improve accuracy ofagents’ opinions. Contributions:

improves Reliability

minimises Communication – the first to operate under thisrestriction

Computationally inexpensive

Adaptive, Scalable, Robust to the presence of indifferentagents

Future work:

Tuning an individual trust level for each neighbour

Attack-resistant solution

19 / 19


Conclusions

AAT exploits properties of social behaviour to improve accuracy ofagents’ opinions. Contributions:

improves Reliability

minimises Communication – the first to operate under thisrestriction

Computationally inexpensive

Adaptive, Scalable, Robust to the presence of indifferentagents

Future work:

Tuning an individual trust level for each neighbour

Attack-resistant solution

19 / 19

Efficient Sharing of Conflicting Opinions with Minimal Communication in Large Decentralised Teams

Technology