agreement dynamics on interaction networks: the naming game
DESCRIPTION
Agreement dynamics on interaction networks: the Naming game. A. Barrat LPT, Université Paris-Sud, France ISI Foundation, Turin, Italy. A. Baronchelli (La Sapienza, Rome, Italy) L. Dall’Asta (LPT, Orsay, France) V. Loreto (La Sapienza, Rome, Italy). http://www.th.u-psud.fr/. - PowerPoint PPT PresentationTRANSCRIPT
Agreement dynamics on interaction networks: the Naming game
A. Baronchelli (La Sapienza, Rome, Italy)L. Dall’Asta (LPT, Orsay, France)V. Loreto (La Sapienza, Rome, Italy)
http://www.th.u-psud.fr/Phys. Rev. E 73 (2006) 015102(R)Europhys. Lett. 73 (2006) 969 Phys. Rev. E 74 (2006) 036105
http://cxnets.googlepages.com
A. BarratLPT, Université Paris-Sud, France
ISI Foundation, Turin, Italy
Introduction
Statistical physics: study of the emergence of global complex properties from purely local rules
“Sociophysics”: Simple (simplistic?) models which mayallow to understand fundamental aspectsof social phenomena
=>Voter model, Axelrod model, Deffuant model….
Opinion formation models
Simplified models of interaction between N agents
Questions:● Convergence to consensus without global
external coordination?● How?● In how much time?
Opinion formation models
Most initial studies:● “mean-field”: each agent can interact with all the others● regular lattices
Recent progresses in network science:social networks: complex networks
small-world, large clustering, heterogeneousstructures, etc…
Studies of agents on complex networks
Naming game
Interactions of N agents who communicate on how to associate a name to a given object
=> Emergence of a communication system?Agents: -can keep in memory different words/names-can communicate with each other
Example of social dynamics/agreement dynamics
(Talking Heads experiment, Steels ’98)
Convergence? Convergence mechanism?Dependence on N of memory/time requirements?Dependence on the topology of interactions?
Naming game: dynamical rules
At each time step:-2 agents, a speaker and a hearer, are randomly selected-the speaker communicates a name to the hearer(if the speaker has nothing in memory –at the beginning- it invents a name)
-if the hearer already has the name in its memory: success-else: failure
Minimal naming game: dynamical rules
success => speaker and hearer retain the uttered word as the correct one and cancel all other words from their memory
failure => the hearer adds to its memory the word given by the speaker
(Baronchelli et al, JSTAT 2006)
Minimal naming game: dynamical rules
Speaker
Speaker Speaker
SpeakerHearerFAILURE
HearerHearer
Hearer
SUCCESS
ARBATIZORGAGRA
ARBATIZORGAGRA
ZORGA
ARBATIZORGAGRA
ZORGA
REFOTROGZEBU
REFOTROGZEBUZORGA
ZORGATROGZEBU
Naming game: other dynamical rules
Speaker
Speaker Speaker
SpeakerHearerFAILURE
HearerHearer
Hearer
SUCCESS
1.ARBATI2.ZORGA3.GRA
1.ARBATI2.ZORGA3.GRA
1.ZORGA2.ARBATI3.GRA
1.ARBATI2.GRA3.ZORGA
1.TROG2.ZORGA3.ZEBU
1.REFO2.TROG3.ZEBU
1.REFO2.TROG3.ZEBU4.ZORGA
1.TROG2.ZEBU3.ZORGA
Possibility of giving weights to words, etc...=> more complicate rules
Naming game:example of social dynamics
-no bounded confidence( Axelrod model, Deffuant model)
-possibility of memory/intermediate states( Voter model, cf also Castello et al 2006)
-no limit on the number of possible states(no parameter)
Simplest case: complete graph
interactions among individuals create complex networks: a population can be represented as a graph on which
interactions
agents nodes
edges
a node interacts equally with all the others, prototype of mean-field behavior
Naming game:example of social dynamics
Baronchelli et al. JSTAT 2006
Complete graph
Total number of words=totalmemory used
N=1024 agents
Number of different words
Success rate
Memory peak
Building of correlations
Convergence
Complete graph:Dependence on system size
● Memory peak: tmax / N1.5 ; Nmaxw / N1.5
average maximum memory per agent / N0.5
● Convergence time: tconv / N1.5
Baronchelli et al. JSTAT 2006
diverges asN 1
Another extreme case:agents on a regular lattice
N=1000 agents
MF=complete graph
1d, 2d: agents on a regularlattice
Nw=total number of words; Nd=number of distinct words; R=success rate
Baronchelli et al., PRE 73 (2006) 015102(R)
Local consensus is reached very quickly through repeated interactions. Then:-clusters of agents with the same unique word start to grow, -at the interfaces series of successful and unsuccessful interactions take place.
coarsening phenomena (slow!)
Few neighbors:
Another extreme case:agents on a regular lattice
Baronchelli et al., PRE 73 (2006) 015102(R)
The evolution of clusters is described as the diffusion of interfaces which remain localized i.e. of finite width
Diffusion equation for the probability P(x,t) that an interface is at the position x at time t:
Each interface diffuses with a diffusion coefficient D(N)» 0.2/N
The average cluster size grows as
Another extreme case:agents on a regular lattice
tconv » N3
Another extreme case:agents on a regular lattice
d=1tmax/ Ntconv/ N3
d=2tmax/ Ntconv/ N2
Regular lattice:Dependence on system size
● Memory peak: tmax / N ; Nmaxw / N
average maximum memory per agent: finite!
● Convergence by coarsening: power-law decrease of Nw/N towards 1
● Convergence time: tconv / N3 =>Slow process!(in d dimensions / N1+2/d)
Two extreme cases
Complete graph dimension 1maximummemory
/ N1.5 / N
convergence
time
/ N1.5 / N3
Naming Game on a small-world
Watts & Strogatz, Nature 393, 440 (1998)
N = 1000
•Large clustering coeff. •Short typical path
N nodes forms a regular lattice. With probability p, each edge is rewired randomly
=>Shortcuts
1D Random topologyp: shortcuts
(rewiring prob.)
(dynamical) crossover expected:
● short times: local 1D topology implies (slow) coarsening
● distance between two shortcuts is O(1/p), thus when a cluster is of order 1/p the mean-field behavior emerges.
Dall'Asta et al., EPL 73 (2006) 969
Naming Game on a small-world
Naming Game on a small-world
increasing p
p=0
p=0: linear chainp À 1/N : small-world
-slower at intermediate times (partial “pinning”)-faster convergence
Naming Game on a small-world
convergence time:/ N1.4
maximum memory:/ N
Complete graph
dimension 1 small-world
maximummemory
/ N1.5 / N / N
convergence time
/ N1.5 / N3 / N1.5
What about other types of networks ?
Better not to haveall-to-all communication,nor a too regular network structure
Definition of the Naming Game on heterogeneous networks
recall original definition of the model:
select a speaker and a hearer at random among all nodes
=>various interpretations once on a network:
-select first a speaker i and then a hearer among i’s neighbours
-select first a hearer i and then a speaker among i’s neighbours
-select a link at random and its 2 extremities at random as hearer and speaker
can be important in heterogeneous networks because: -a randomly chosen node has typically small degree-the neighbour of a randomly chosen node has typically large degree
Dall’Asta et al., PRE 74 (2006) 036105
(cf also Suchecki et al, 2005 and Castellano, 2005)
NG on heterogeneous networks
Different behaviours
shows the importanceof understanding the roleof the hubs!
Example: agents on a BA network:
NG on heterogeneous networks
Speaker first: hubs accumulate more words
Hearer first: hubs have less words and “polarize” the system,hence a faster dynamics
NG on homogeneous and heterogeneous networks
-Long reorganization phasewith creation of correlations, at almost constant Nw and decreasing Nd
-similar behaviour for BAand ER networks(except for single node dynamics),as also observed for Voter model
NG on complex networks:dependence on system size
● Memory peak: tmax / N ; Nmaxw / N
average maximum memory per agent: finite!
● Convergence time: tconv / N1.5
Effects of average degree
larger <k>
● larger memory, ● faster convergence
larger clustering
● smaller memory, ● slower convergence
Effects of enhanced clustering(more triangles, at constant number of edges)
C increases
Bad transmissions/errors?
Modified dynamical rules:in case of potential successful communication:
● With probability : success● With probability 1-: nothing happens (irresolute attitude)
1 : usual Naming Game => convergence0 : no elimination of names => no convergence
Expect a transition at some c
A. Baronchelli et al, cond-mat/0611717
Mean-field case
Stability of the consensus state ? consider a state with only 2 words A, B
Evolution equations for the densities: nA, nB, nAB
> 1/3 : states (nA=nAB=0, nB=1), (nB=nAB=0, nA=1)< 1/3 : state with nAB > 0 , nA=nB > 0
Mean-field case
At c = 1/3, •Consensus to Polarization transition•tconv / (-c)-1
The polarized state is active( Axelrod model, in which the polarized state is frozen)
Mean-field case:numerics
Usual NG NG with at most m different words
=>At least 2 different universality classes
Series of transitions
tm=time to reach a state with m different words
Transitions to more and more disordered active states
On networks
-Influence of strategy-Transition preserved on het. networks( Axelrod model)
On networks, as in MF
At c , Consensus to Polarization transition(c depends on strategy+network heterogeneity)
The polarized state is active
Other issues
● Community structures (slow down/stop convergence)(cf also Castello et al, arXiv:0705.2560)
● Other (more efficient) strategies (dynamical rules) (A. Baronchelli et al., physics/0511201; Q. Lu et al., cs.MA/0604075)
● Activity of single nodes(L. Dall’Asta and A. Baronchelli, J. Phys A 2006)
● Coupling the dynamics of the network with the dynamics on the network: transitions between consensus and polarized states, effect of intermediate states…
On networks
Possible to write evolution equations=> c ()