Interacting agents with public and private information...

20
Interacting agents with public and private information (statistical mechanics perspective) A De Martino (SMC, Roma 1) mostly joint work with M Marsili (ICTP Trieste) http://chimera.roma1.infn.it/ANDREA A De Martino and M Marsili, J. Physics A 39 R465 (2006)

Transcript of Interacting agents with public and private information...

Page 1: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

Interacting agents with public and private information(statistical mechanics perspective)

A De Martino (SMC, Roma 1)

mostly joint work with M Marsili (ICTP Trieste)

http://chimera.roma1.infn.it/ANDREA

A De Martino and M Marsili, J. Physics A 39 R465 (2006)

Page 2: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

Helbing+2 02Selten+8 04

1

2

3

4

5

6

7

8

9

0 100 200 300 400 500

Pla

yer

Iteration t

1

2

3

4

5

6

7

8

9

500 600 700 800 900 1000

Pla

yer

Iteration t

1

2

3

4

5

6

7

8

9

1000 1100 1200 1300 1400 1500

Pla

yer

Iteration t

0

1

2

3

4

0 500 1000 1500 2000 2500

Ch

an

gin

g R

ate

user-speccorrected

own payoff both similar

Page 3: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

P1(n1) = P0

1 ! P1

1 n1

P2(n2) = P0

2 ! P1

2 n2

! feq

1=

P 12

P 11

+ P 12

+1

N

P 01 " P 0

2

P 11

+ P 12

0

1

2

3

Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5

0 500 1000 1500 2000 2500

Ove

rre

actio

n

Iteration t

user-speccorrected

own payoff

both similar

“congestion game”

Page 4: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

in financial markets (see T Odean’s group work)

people systematically overweight some typesof information and underweight others ; overweight their successes and underweight their failures ; overestimate the precision of their information ; etc.

}overconfidence(increases trading volumes,decreases exp. utility, causesunderreaction to realinformation, etc.)

“hold losers/sell winners”

Page 5: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

a. when is convergence to an optimal state possible?b. when there are many, which one is selected?c. can one supply information so as to stabilize decision dynamics?d. how do information schemes affect the dynamics?e. is there an optimal information scheme?

some questions

1 agent , 1resourceN agents , 1 resourceN agents , O(N) resources

Page 6: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

learning? (backward-looking induction vs forward-looking deduction)

one agent , choices : s ! {+,"}

learning :

decision : P (s(t) = s) ! ![ys(t)]

ys(t + 1) ! ys(t) = us(t)

ys(t + 1) ! ys(t) = us(t)!s(t),s

(full)

(partial)

payo!s us(t) i.r.v. with Eus = vs (v+ > v!

)

learn the best choice?

full + !(y) = ey

!

(" > 0) ! yes

partial + !(y) = y! (" = 1) ! yes

Rustichini 99

Page 7: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

one agent , state a ! {H, L} (well/badly informed)

forecasts an event ! correctly with prob. a

can he learn a from his past success? s(t) =!

t!<t

!!(t!),f(t!)

learning : P [a = H|s(t) = s] ! !!(t|s)

=!sHs(1 ! H)t!s"0

!sHs(1 ! H)t!s"0 + Ls(1 ! L)t!s(1 ! "0)

! = 1 ! rational ; ! > 1 ! underreact to errors

! if a = L the updated posterior converges a.s. as

!!(t|s) !

!

"

#

"

$

1 for " > "!

!0 for " = "!

0 for " < "!

Gervais-Odean 01

correct priors,biased updates

Page 8: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

1 to N agents

choices ai ! {"1, 1} , A =!

j

aj

payo!s ui(ai, a!i) =N ! aiA

2

decision : P (ai(t) = a) ! eyia(t)

(full)

learning : yia(t + 1) ! yia(t) =!

N

N ! aA(t)

2

Ui = yi+ ! yi! " Ui(t + 1) ! Ui(t) = !

!

NA(t)

(optimal learning)

Page 9: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

0 2 4 6s0

5

10

15!c(s)

2 4 6 8 10!

0

0.5

1

"2 /#2

s=1/2s=1 ! < !c ! !

2" N

! > !c ! !2" N

2

Fluctuations : !2 =

!

i

(!a2

i " # !ai"2) =

!

i

(1 # !ai"2)

yia(0) randomly sampled from G(0, s2)

so the larger the spread of i.c.’s (heterogeneity), the smaller the fluctuations but the steady state is not optimal

Optimal state :

!

N!1

2do a

N+12

do !a

! !2

= 1 NNE " eN!

Page 10: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

Ui(t + 1) ! Ui(t) = !!A(t)/N , A(t) =!

j

aj(t) , ! > 0

i is in hereWhy can’t they get to Nash?

remove self-interaction (learn to respond to others, not to yourself)

! ! {0, 1}Ui(t + 1) ! Ui(t) = !

!

N[A(t) ! !ai(t)]

!Ui(t + 1)" # !Ui(t)" = #!

N[!

j

mj # !mi] mi = !ai"= !

!

N

!H

!mi

H =1

2

!

"

i

mi

#2

!

!

2

"

i

m2

i

! = 0 ! mi = 0

! = 1 : H is harmonic ! mi = ±1

! !2

= N

! !2 = 1 (odd N)

minima :

Page 11: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

0

1

2

3

Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5

0 500 1000 1500 2000 2500

Ove

rre

actio

n

Iteration t

user-speccorrected

own payoff

both similar

partial full NE

externally supplied, unrealistic(it is seen as a 1/N effect, “price-taking”)

Page 12: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

buy/sell difference

first gap

AZN, LSE 1999-2002

Farmer+2 04

price

old price price

Large price variation =“closing” of the first gap

Page 13: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

Agents areheterogeneousselfishinductive

Resources areequivalentscarce(distributed)

No direct interaction

N agents, M resources and a matrix of dependencies {!µi }

N ! "random {!µ

i }Typical properties :

Resources

Load

{Qµ}

Q

N = cM

Page 14: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

N drivers on a grid of M streetsone driver’s

feasible routes

another driver’sfeasible routes

A

BA

B

route for i : !"i = {"µi }

c = N/M = density of vehicles

Page 15: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

Aµ(t) = Qµ(t) ! Q

Resources

Load

{Qµ}

Q

how evenly are resources used?

how large are fluctuations?

!i(t) !!

µ

aµi (t)Qµ(t) "

!

i

!i(t) !!

µ

Qµ(t)2

!2 =1

M

!

µ

!(Qµ)2" # !Qµ"2

H =1

M

µ

!Aµ"2

Page 16: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

two feasible routes per agent : !"is , s ! {+,"}

decision : P (si(t) = s) ! e!Uis(t)

! ! 0

learning : Uis(t + 1) ! Uis(t) = !

1

M

!

µ

!µisQ

µ(t) +1

2(1 ! "s,si(t))zis(t)

Qµ(t) =!

i

!µis"s,si(t)

“congestion game”

information noise

! = 0

! > 0

! < 0

: unbiased information: overestimates unused routes: underestimates unused routes

!!ig(n)" = "

!!ig(n)!jh(m)" = !"ij"gh"nm

! = 0 : same information (biased or not) for all drivers

! > 0 : user-specific information

Page 17: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

! = 0

Not optimal butnot that bad

! = ! = 0

! = !

Disaster for c > cc

0.1

1

10

100

!2

/N

1/"=0

"=0

10!2

10!1

100

101

102

c

0.0

0.1

0.2

0.3

0.4

0.5

H/N

cc ! 3

theory (steady state) : the dynamics minimizes H (approx.)

Uncorrelated result : cc = 2.9638 . . .

noiseless case

(high density)

inductive perform worse than random at high densities!

problem : the quenched disorder is spatially correlated ...

Page 18: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

10!2

10!1

100

101

c

0.0

0.2

0.4

0.6

0.8

1.0

!2

/N

random drivers

adaptive drivers, "=0

adaptive drivers, "=#2

! = !2

! = 0 ! = 0

`pessimist’ information bias reduces fluctuations

theory (steady state) : the dynamics minimizes fluctuations

biased case

high car density (hard phase)

calibrated user-specific noise reduces fluctuations

10!1

100

101

102

103

104

!

10!1

100

101

"2/#

100

102

104

! teq

1/4

10!1

101

! = 0

noisy case

Page 19: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

0

1

2

3

Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5

0 500 1000 1500 2000 2500

Ove

rre

actio

n

Iteration t

user-speccorrected

own payoff

both similar

10!1

100

101

102

103

104

!

10!1

100

101

"2/#

100

102

104

! teq

1/4

10!1

101

! = 0

10!2

10!1

100

101

c

0.0

0.2

0.4

0.6

0.8

1.0

!2

/N

random drivers

adaptive drivers, "=0

adaptive drivers, "=#2

! = !2

! = 0

! = 0

noiseless case = minH

min!2

Page 20: Interacting agents with public and private information ...chimera.roma1.infn.it/ANDREA/aaland.pdf · Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5 0 500 1000 1500 2000

minH

min!2

0.01 0.1 1 10 100!

0

0.5

1

"

RS

RSB

ERGODIC

NON ERGODIC

1/c

!=

!"/2

min(1 ! !)H + !"2

no PT(freezing, exp. many SS, ...)

PT(no freezing)