Meaning, Evolution and the Structure of Society
Roland M ¨uhlenbernd
November 7, 2014
O
VERVIEWGame Theory and Linguistics Pragm. Reasoning Language Evolution
Signaling Games
Replicator Dyn. Imitation Dyn. Learning Dyn.
Reinforcement BL & BR GT in Lang. Use
I Accounts differ inpsychologyof language users
I What role does thesociologyof language users play?
E
MERGENCE OFH
UMANL
ANGUAGE”[w]e can hardly suppose a parliament of hitherto speechless elders meeting together and agreeing to call a cow a cow and a wolf a wolf.”
Russell, The Analysis of Mind
”Two savages, who had never been taught to speak, but had been bred up remote from the societies of men, would naturally begin to form that language by which they would endeavor to make their mutual wants intelligible to each other, by uttering certain sounds, whenever they meant to denote certain objects.”
Smith, Considerations Concerning the First Formation of Languages
D
EFINITION OF AS
IGNALINGG
AME(Lewis game)
Signaling GameSG=h{S,R},T,M,A,Pr,C,Ui
I set of players{S,R}
I set of statesT(T={t1,t2})
I state probability functionPr∈∆(T)(Pr(t1) =Pr(t2) =.5)
I set of messagesM(M={m1,m2})
I cost functionC:M→R(C(m1) =C(m2) =0)
I set of actionsA(A={a1,a2})
I utility functionU(ti,mj,ak) =
1−C(mj) ifi=k 0−C(mj) else a1 a2
t1 1,1 0,0 t2 0,0 1,1
P
URES
TRATEGIESI the players’ ”actions” can be represented as pure strategies
I set of pure sender strategiesS={s|s∈[T→M]}
I set of pure receiver strategiesR={r|r∈[M→A]}
I for the Lewis game there are 4 strategies for each player
I S={s1,s2,s3,s4}
I R={r1,r2,r3,r4}
s1: t1 m1
t2 m2
s2: t1
m2 t2
m1
s3: t1 m1
t2 m2
s4: t1
m2 t2
m1
r1: m1 a1
m2 a2
r2: m1
a2 m2
a1
r3: m1 a1
m2 a2
r4: m1
a2 m2
a1
S
IGNALINGS
YSTEMS...
I are combinations of pure strategies. The Lewis game has two:L1 =hs1,r1iandL2=hs2,r2i
L1: tL
tS
m1 m2
aL aS L2:
tL
tS
m1 m2
aL aS
I are strictNash equilibriaover expected utilities
I associate messages to states in an unique way
I areevolutionary stable states
E
XPECTEDU
TILITIES OF THEL
EWISG
AMEThe expected utility for a combination of strategies is given as:
EU(s,r) =X
t∈T
Pr(t)×U(t,s(t),r(s(t))) (1)
r1 r2 r3 r4 s1 1 0 .5 .5 s2 0 1 .5 .5 s3 .5 .5 .5 .5 s4 .5 .5 .5 .5
Table: The static Lewis game: Expected utilities for all strategy combinations of the dynamic Lewis game
T
HEH
ORNG
AMESG=h{S,R},T,M,A,Pr,C,Ui
I T={tf,tr},M={mu,mm},A={af,ar}
I Pr(tf)>Pr(tr)
I C(mu)<C(mm)
I S={sh,sa,ss,sy}
I R={rh,ra,rs,ry}
sh:
tf mu
tr mm
sa:
tf mm tr
mu
ss:
tf mu
tr mm
sy: tf
mm tr
mu
rh: mu
af
mm ar
ra: mu
ar mm
af
rs: mu
af
mm ar
ry: mu
ar mm
af
E
XPECTEDU
TILITIES OF THEH
ORNG
AMEExemplary parameters for the Horn Game:
I Pr(tf) =.7
I C(mu) =.1,C(mm) =.2
rh ra rs ry sh .87 -.13 .57 .17 sa -.17 .83 .53 .13
ss .6 .2 .6 .2
sy .1 .5 .5 .1
Table: Payoff table of the asymmetric static Horn game with Pr(tf) =.75,C(mu) =.1 andC(mm) =.2
T
HEP
ROGRAM OFT
ODAYI Which strategy emerges under evolutionary (population) dynamics like the replicator dynamics?
I How might the outcome differ for agent-based dynamics like imitation or learning?
I How does the outcome vary for different typologies?
I What role might structural aspects play for the emergence of specific strategies?
R
EPLICATORD
YNAMICSGiven a very large (effectively infinite) population of agents playing a symmetric static gameh{P1,P2},S,U:S×S→Ri randomly against each other. Then we can define
I p(si): proportion of agents in the population playing strategysi I U(si) =P
sj∈Sp(sj)U(si,sj): expected utility for agents playingsi I U=P
si∈Sp(si)U(si)the average fitness of the whole population
Definition (Replicator Dynamics)
The RD is defined by the following differential equation:
dp(si)
dt =p(si)[U(si)−U]
E
VOLUTIONARYS
TABLES
TRATEGYDefinition (Evolutionary stable strategy)
For a symmetric static gameh{P1,P2},S,U:S×S→Ria strategy si ∈Sis said to be a evolutionary stable strategy if and only if one of the following two conditions holds:
1. U(si,si)>U(si,sj)for all sj 6=si
2. if U(si,si) =U(si,sj)for some sj6=si, then U(si,sj)>U(sj,sj) A ESSsihas the following properties:
I sihas an invasion barrier: resistant against the invasion of a small proportion of mutants withsj 6=si
I siis a attractor under the replicator dynamics
I Strict NE ⊂ESS ⊂NE
H
ORNG
AME UNDERR
EPLICATORD
YNAMICSStrategies:•Horn,•anti-Horn,•Smolensky,•anti-Smolensky Start population:
[.25, .25, .25, .25]
B
ASIN OFA
TTRACTION ANDG
AMEP
ARAMETERSHorn game 1:
I Pr: [.6, .4]
I C: [.05, .1]
Lh: 55%,La: 43%,Ls: 2%
Horn game 2:
I Pr: [.7, .3]
I C: [.1, .2]
Lh: 59%,La: 35%,Ls: 6%
Horn game 3:
I Pr: [.9, .1]
I C: [.1, .3]
Lh: 64%,La: 22%,Ls: 14%
A
ND NOW?
”...the replicator dynamics is a natural place to begin investigations of dynamical models of cultural evolution, but I do not believe that it is the whole story.”
Skyrms, Evolution of the Social Contract
I
MITATIOND
YNAMICSThe replicator dynamics describes the most likely path of strategy distribution in a virtually infinite and
homogeneous population, when every agent updates her behavior by conditional imitation.
shown by e.g. Helbing 1996, Schlag 1998
Conditional Imitation: the probability of an agent playing strategysto switch to a randomly chosen participant’s strategys0is given as:
P(s→s0) =
( 0 ifU(s,s0)≥U(s0,s0) α×U(sU0,s0)−U(s,s0)
max−Umin else (2)
with:
I α: scaling factor
I Umax: maximal value of the utility table
I Umin: minimal value of the utility table
I
MITATIONVS R
EPLICATORD
YNAMICSHorn
Anti-Horn Smolensky
Macro .5 .374 .122 .003
(a)Replicator Dynamics
Horn
Anti-Horn Smolensky
Micro 100 .459 .365 .160
(b)Conditional Imitation Figure: Basins of attraction for the replicator dynamics and the conditional imitation rule by numerical simulation for a population of 100 randomly interacting agents.
→the Correspondence of basins of attraction is93.2%
I
MITATION OND
IFFERENTT
YPOLOGIESSimulations with 100 agents on different typologies:
I β-graph (Watts-Strogatz algorithm, starting ring) I scale-free networks (Holme-Kim algorithm) I 10×10 toroid lattice
percentage
0 0.2 0.4 0.6
complete (93.2%) scale-free (91%) β-graph (89.1%) grid (88.2 %) Smolensky
anti-Horn Horn
Figure: Comparing distributions of basin of attraction and
RD-correspondence (percentage in brackets) for different network topologies
T
HEE
VOLUTION OFR
EGIONALM
EANINGZollman (2005) made experi- ments with agents
I playing the Lewis game
I updating by Imitation
I placed on a 100×100 toroid lattice
Result: the emergence of re- gional meaning
Source: Kevin J. S. Zollman (2005): Talking to Neighbors:
The Evolution of Regional Meaning. Philosophy of Science 72.1, 69–85.
R
EINFORCEMENTL
EARNINGAgents play the dynamic gameSG=h(S,R),T,M,A,Pr,C,Ui
I as probabilistic strategiess:T→∆(M),r:M→∆(A)
I communicative outcome shifts probabilities in appropriate directions
I successful communication viaht∈T,m∈M,a∈Aimakes the choicess(m|t)andr(a|m)more probable in subsequent rounds
I agents might learn a pure strategy profile that constitutes a signaling system
T
HEE
VOLUTION OFR
EGIONALM
EANINGExperiments on a 30×30 toroid lattice:
The Lewis Game
L1 L2 other
The Horn Game
Horn Anti-Horn other Regional Convention evolve not always for the Horn Game.
E
XPERIMENTS ON AS
OCIALM
API the closer the distance on the lattice, the more probable the interaction I relationship controlled bydegree of localityγ
γ prob. for neighborhood
communication
0 1 2 3 4 5 6 7 8
0 0.2 0.4 0.6 0.8 1
percentage
0%
20%
40%
60%
80%
100%
Horn game Lewis game
Local interaction supports the emergence of local conventions.
S
MALLW
ORLDN
ETWORKSβ-graph network (Watts & Strogatz 1998)
scale-free network (Holme & Kin 2002)
E
XPERIMENTALS
ETUPSimulation runs for populations of 300 agents
I placed on small world network interaction structures
I ’interact’ via Lewis game and Horn game
I decide and update by learning dynamics: reinforcement learning or belief learning
I interact until 90%of all agents are language learners Analyzis of
I properties of language regions
I interrelationships between agents’ network position and learning behavior
I specific roles of agent in the emerging process of conventions/signaling languages
R
ESULTS: L
ANGUAGER
EGIONSdensity
n
0 50 100 150 200 250 300
0 0.2 0.4 0.6 0.8 1
language region (β−graph) language region (scale-free network)
average clustering
n
0 50 100 150 200 250 300
0.2 0.4 0.6 0.8 1
language region (β−graph) language region (scale-free network)
Cliquishness supports the emergence of language regions
C
ONCLUSIONSocial structure matters to the evolution and establishment of conventional meaning!
I different interaction structures lead to shifts in the basin of attraction of communication strategies
I locality of structure supports regional variety
I particular network properties support the emergence of regional conventions