p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy,...

21
Optimal control of stochastic multi-agent systems in continuous space and time Wim Wiegerinck, Bart van den Broek, Bert Kappen SNN, Radboud University Nijmegen Presented at UAI 2006 The research reported here is part of the Interactive Collab- orative Information Systems (ICIS) project, supported by the Dutch Ministry of Economic Affairs, grant BSIK03024. W. Wiegerinck - UvA, 9 October, 2006 – p.1

Transcript of p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy,...

Page 1: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Opt

imal

cont

rolo

fst

ocha

stic

mul

ti-a

gent

syst

ems

inco

ntin

uous

spac

ean

dti

me

Wim

Wie

geri

nck,

Bar

tva

nde

nB

roek

,B

ert

Kap

pen

SNN

,Rad

boud

Uni

vers

ityN

ijmeg

en

Pres

ente

dat

UA

I20

06

The

rese

arch

repo

rted

here

ispa

rtof

the

Inte

ract

ive

Col

lab-

orat

ive

Info

rmat

ion

Syst

ems

(IC

IS)

proj

ect,

supp

orte

dby

the

Dut

chM

inis

try

ofE

cono

mic

Aff

airs

,gra

ntB

SIK

0302

4.

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

1

Page 2: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Stoc

hast

icm

ulti

-age

ntsy

stem

san

dop

tim

alco

ntro

l

Mul

ti-ag

ents

yste

ms

(e.g

.fir

emen

-se

efig

ure)

that

have

todi

stri

bute

them

selv

esov

era

num

ber

ofta

rget

s(e

.g.

fires

)no

isy,

non-

linea

rdy

nam

ics

inco

ntin

uous

spac

e-tim

ead

ditiv

eco

ntro

lof

the

dyna

mic

s

Opt

imal

cont

rol:

min

imiz

eto

talj

oint

cost

(=

effo

rtco

st+

end

cost

)to

whi

chfir

esh

ould

afir

eman

go?

whe

nto

deci

de?

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

2

Page 3: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Con

tent

s

Stoc

hast

icop

timal

cont

roli

nco

ntin

uous

spac

etim

e,H

amilt

on-J

acob

i-B

ellm

aneq

uatio

n

Tra

nsfo

rmin

toa

linea

rPD

EFr

omsi

ngle

-age

ntsi

ngle

-tar

gets

yste

ms

tom

ulti-

agen

tmul

ti-ta

rget

syst

ems

MA

Sco

ntro

land

grap

hica

lmod

els

Sim

ulat

ions

Sum

mar

y,di

scus

sion

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

3

Page 4: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Stoc

hast

icdy

nam

ics

We

cons

ider

asy

stem

inco

ntin

uous

spac

ean

dtim

e.It

sst

ate

x∈

Rk

obey

sth

est

ocha

stic

dyna

mic

s

dx

=(b

(x,t

)+

u)d

t+

b:

drif

tter

mm

odel

ing

the

dyna

mic

sdu

eto

the

envi

ronm

ent,

u:

cont

rolt

oin

fluen

ceth

edy

nam

ics,

aW

iene

rpr

oces

s(i

.e.

nois

e)w

ith〈d

ξ idξ j〉

ijdt.

00.

20.

40.

60.

81

1.2

1.4

−0.

8

−0.

6

−0.

4

−0.

20

0.2

0.4

0.6

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

4

Page 5: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Con

trol

prob

lem

Find

the

cont

rolu

(.)

that

min

imiz

esth

eex

pect

edco

st-t

o-go

C(x

i,t i

,u(.

))=

φ(x

(T))

+

T

t i

dt(

1 2u

(x,t

)>R

u(x

,t)+

V(x

(t),

t))

inw

hich x

i:in

itial

stat

e,

t i:

initi

altim

e,

T:

Fixe

den

d-tim

e,

φ(x

(T))

:co

stof

bein

gin

stat

ex

aten

d-tim

eT

,

V(x

(t),

t)dt:

cost

ofbe

ing

inst

ate

xdu

ring

time

inte

rval

[t,t

+dt]

,

u>

Ru

dt:

cost

ofco

ntro

ldur

ing

the

sam

etim

ein

terv

al,

Ris

aco

nsta

ntk×

km

atri

x(p

aram

etri

zing

the

cost

ofco

ntro

l).

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

5

Page 6: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Ham

ilton

-Jac

obi-

Bel

lman

equa

tion

and

op-

tim

alco

ntro

l

Opt

imal

(exp

ecte

d)co

st-t

o-go

J(x

,t)=

min

u(.

)C

(x,t

,u(.

)).

Itsa

tisfie

sth

est

ocha

stic

Ham

ilton

-Jac

obi-

Bel

lman

(HJB

)eq

uatio

n

−∂

tJ

=m

inu

(.)

(

1 2u>

Ru

+V

+(

b+

u)

>∇

J+

1 2T

r(

ν∇

2J)

)

=−

1 2∇

J>

R−

1∇

J+

V+

b>∇

J+

1 2T

r(

ν∇

2J)

with

boun

dary

cond

ition

J(x

,T)=

φ(x

).

The

min

imiz

atio

nw

ithre

spec

tto

uyi

elds

u=

−R

−1∇

J,

whi

chde

fines

the

optim

alco

ntro

l.

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

6

Page 7: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Log

tran

sfor

mat

ion

and

linea

rev

olut

ion

Ass

ume

ν=

λR

−1,t

hen

the

non-

linea

rPD

Eof

Jca

ntr

ansf

orm

edin

toa

linea

ron

eby

the

log

tran

sfor

m(W

.Fle

mm

ing,

1978

).Se

t

J(x

,t)=

−λ

log

Z(x

,t)

with

“par

titio

nfu

nctio

n”

Z(x

,t)=

Z

dkyρ(y

,T|x

,t)ex

p(−

φ(y

)/λ)

inw

hich

the

‘pro

babi

lity

dens

ity’ρ

satis

fies

the

Fokk

er-P

lanc

keq

uatio

n

∂ϑρ(y

,ϑ|x

,t)=

−V λ

ρ−

∇> y

bρ+

1 2T

r`

ν∇

2 yρ

´

.

ρ(y

,T|x

,t):

prob

abili

tyof

getti

ngat

yat

time

Tgi

ven

initi

alst

ate

xat

time

t,

follo

win

gst

ocha

stic

syst

emdy

nam

ics

inab

senc

eof

cont

rol,

i.e.,

u=

0

with

anni

hila

tion

prob

abili

tyV

(x,t

)dt/

λ

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

7

Page 8: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Lin

ear

theo

ry

Exp

ecte

dop

timal

cost

togo J

(x,t

)=

λlo

gZ

(x,t

)

Zis

expr

esse

das Z

(x,t

)=

dkyρ(y

,T|x

,t)ex

p(−

φ(y

)/λ)

inw

hich

ρsa

tisfie

sth

eFo

kker

-Pla

nkeq

uatio

n.

The

optim

alco

ntro

lis

give

nby

u(x

,t)=

ν∇

log

Z(x

,t).

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

8

Page 9: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Exa

mpl

e:Q

uadr

atic

end-

cost

b=

0,V

=0

Ran

scal

ars,

Solu

tion

ofth

edi

ffus

ion

equa

tion

ρ(y

,T|x

,t)

=(2

πν(T

−t)

)k/2ex

p(−

|y−

x|2

2ν(T

−t)

)

Ass

ume

end-

cost

:φ(x

)=

α|x

−µ|2

The

nZ

follo

ws

from

conv

olut

ion

with

exp(−

φ(y

)/λ),

Z(x

,t)∝

exp

(

−|x

−µ|2

2ν(T

−t+

R/α

)

)

.

The

optim

alco

ntro

lfol

low

sfr

omu

=ν∇

log

Z,

u(x

,t)=

µ−

x

T−

t+

R/α

.W

.Wie

geri

nck

-U

vA,9

Oct

ober

,200

6–

p.9

Page 10: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Sing

leta

rget

Bac

kto

arbi

trar

yb

and

V.

Toen

forc

ean

end-

stat

eat

targ

etµ

,we

set

exp(−

φ(y

)/λ)∝

δ(y−

µ)

Thi

sim

plie

s

Z(x

,t;µ

)∝

ρ(µ

,T|x

,t),

u(x

,t;µ

)=

ν∇

log

Z(x

,t;µ

)

=ν∇

log

ρ(µ

,T|x

,t)

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

10

Page 11: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Run

ning

exam

ple

Ass

ume

b=

0,V

=0,

Ran

scal

ars.

Z(x

,t;µ

)∝

exp

[

−|x

−µ|2

2ν(T

−t)

]

,

u(x

,t;µ

)=

µ−

x

T−

t.

For

any

blin

ear

inx

,and

V=

0,in

the

Fokk

er-P

lanc

keq

uatio

nca

nbe

solv

edan

alyt

ical

ly.

Its

solu

tion

isa

Gau

ssia

n.Z

and

uar

eof

esse

ntia

llyth

esa

me

form

asin

the

runn

ing

exam

ple.

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

11

Page 12: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Mul

tipl

eta

rget

s

Toen

forc

ean

end-

stat

eat

one

ofm

targ

etsµ

s,w

ithta

rget

pref

eren

ces

expr

esse

dby

rela

tive

cost

E(s

),

exp(−

φ(y

)/λ)∝

m∑ s=

1

exp(−

E(s

)/λ)δ

(y−

µs)

.

Part

ition

func

tion

and

optim

alco

ntro

lcan

beex

pres

sed

assu

mof

sing

le-t

arge

tqu

antit

ies,

Z(x

,t)

∝m

∑ s=

1

exp(−

E(s

)/λ)Z

(x,t

;µs)

,

u(x

,t)

=m

∑ s=

1

p(s|x

,t)u

(x,t

;µs),

inw

hich

p(s|x

,t)

isth

epr

obab

ility

p(s|x

,t)=

exp(−

E(s

)/λ)Z

(x,t

;µs)

m s′=

1ex

p(−

E(s

′ )/λ

)Z(x

,t;µ

s′).

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

12

Page 13: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Exa

mpl

e:M

ulti

ple

targ

ets

Inth

eru

nnin

gex

ampl

e,op

timal

cont

rolw

ithm

ultip

leta

rget

sis

u(x

,t)

=∑

s

p(s|x

,t)u

(x,t

;µs)

=∑

s

p(s|x

,t)

(

µs−

x

T−

t

)

=µ̄−

x

T−

t

with

µ̄th

e‘e

xpec

ted

targ

et’

µ̄=

m∑ s=

1

p(s|x

,t)µ

s

whi

chis

the

expe

cted

valu

eof

the

targ

etac

cord

ing

toth

epr

obab

ility

p(s|x

,t)=

exp(−

E(s

))ex

p[

−|x

−µ

s|2

2ν(T

−t)

]

m s=

1ex

p(−

E(s

))ex

p[

−|x

−µ

s|2

2ν(T

−t)

]

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

13

Page 14: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Mul

ti-a

gent

s,m

ulti

ple

targ

ets

MA

Sst

ate

~ x=

(x1,.

..,x

n),

sing

leag

ents

tate

xa.

Dyn

amic

sno

n-in

tera

ctiv

e:b

a(~ x

,t)=

ba(x

a,t

)

V(~ x

,t)

=∑

aV

a(x

a,t

).

Furt

herm

ore,

ν=

λR

−1

glob

ally

.T

here

fore

ρfa

ctor

izes

,

ρ(~ y

,T|~ x

,t)

=∏

a

ρa(y

a,T

|xa,t

).

Cou

plin

gvi

ajo

intt

ask:

dist

ribu

teov

erta

rget

s

exp(−

φ(~ y

)/λ)

=∑

s

exp(−

E(~s

)/λ)∏

a

δ(y

a−

µs

a).

s ala

belo

fta

rget

reac

hed

byag

enta

.E

(~s)

=E

(s1,.

..,s

n)

isth

eco

stw

hen

agen

t1re

ache

s1,

agen

t2re

ache

s2

etc.

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

14

Page 15: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Mul

ti-a

gent

s,m

ulti

ple

targ

ets

(con

t)

Con

trol

for

agen

ta(d

epen

dson

join

tsta

te)

ua(~ x

,t)

=

m∑ sa=

1

p(s

a|~ x

,t)u

a(~ x

a,t

;µs

a)

.

Con

trol

invo

lves

mar

gina

ldis

trib

utio

nfo

rag

enta

p(s

a|~ x

,t)=

~s\s

aex

p(−

E(~s

)/λ)∏

bZ

b(~ x

b,t

;sb)

~sex

p(−

E(~s

)/λ)∏

cZ

c(~ x

c,t

;sb)

,

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

15

Page 16: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Exa

mpl

e:M

AS,

mul

tipl

eta

rget

s

Inth

eru

nnin

gex

ampl

e,op

timal

cont

rolw

ithm

ultip

leta

rget

sis

ua(~ x

,t)

=µ̄

a−

xa

T−

tw

ithµ̄

the

‘exp

ecte

dta

rget

’fo

rag

enta

µ̄a

=m

∑ s=

1

p(s

a|~ x

,t)µ

sa

whi

chis

the

expe

cted

valu

eof

the

targ

etac

cord

ing

toth

epr

obab

ility

p(s

a|~ x

,t)∝

w(s

a|~ x

\a,t

)ex

p

[

−|x

a−

µs

a|2

2ν(T

−t)

]

with

w(s

a|~ x

\a,t

)=

∑ ~s\s

a

exp(−

E(~s

))ex

p

[

b6=

a|x

b−

µs

b|2

2ν(T

−t)

]

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

16

Page 17: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

MA

Sco

ntro

land

grap

hica

lmod

els

MA

Sco

ntro

lreq

uire

sin

fere

nce

ofp(s

a|~ x

,t).

Gra

phic

alm

odel

met

hods

can

beex

ploi

ted

ifp(~s

)is

afa

ctor

grap

h,i.e

.if

the

cost

can

bew

ritte

n

E(~s

)=

α

Eα(s

α)

with

αgr

oups

ofag

ents

.E

.g.

pair

wis

eco

sts

Eab(s

a,s

b)

=−

c abδ s

as

b.

Bol

tzm

ann

mac

hine

anal

ogy

Ea,b(s

a,s

b)

play

sro

leof

coup

lings

ina

Bol

tzm

ann

mac

hine

,con

stan

tin

the

syst

emZ

a(x

a,t

;µs

a),

(i.e

.,ρ(µ

sa,T

|xa,t

))pl

ays

role

ofbi

asin

aB

M,

chan

ges

over

time

Inge

nera

l:gr

aphi

cals

truc

ture

ispr

eser

ved

over

time

(unl

ike

disc

rete

time

fact

ored

MD

Ps).

Spar

segr

aphs

⇒ju

nctio

ntr

eeal

gori

thm

.

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

17

Page 18: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Sim

ulat

ions

b=

0,V

=0,

pair

wis

ere

latio

nsc a

b.

1-d

‘Pos

ition

s’x

a.

for

plot

ting

purp

oses

.k-

dim

ensi

onal

wou

ldbe

feas

ible

asw

ell.

‘Exp

ecte

dta

rget

s’µ

a=

sap

a(s

a|x

)µs

a.

Onl

yfo

ril

lust

rati

on:u

a=

µa−

xa

ν(T

−t)

ism

athe

mat

ical

lyop

tim

al!

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

18

Page 19: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Fir

emen

prob

lem

6ag

ents

(fire

men

),th

ree

targ

ets

(fire

s).

Fully

conn

ecte

dgr

aph

with

c ab

=c

nega

tive⇒

aim

todi

stri

bute

even

ly.

01

−101

time

positions

01

−101

time

expected targetsSy

mm

etry

brea

king

asde

laye

dch

oice

(Kap

pen,

2005

)

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

19

Page 20: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Hol

iday

reso

rtpr

oble

m

42ag

ents

,3ta

rget

s(r

esor

ts).

Ere

pres

ente

dby

spar

segr

aph:

each

agen

thas

pair

wis

ere

latio

ns,

c ab

1,w

ithth

ree

othe

rag

ents

.A

gent

only

care

sfo

rre

late

dag

ents

whe

ther

orno

tto

have

holid

ayin

the

sam

ere

sort

.

Join

ttas

kis

toop

timal

lydi

stri

bute

MA

Sov

erth

ere

sort

s.

Cliq

ue-s

ize

=7

01

−2

−1012

time

positions

01

−101

time

expected targets

W.W

iege

rinc

k-

UvA

,9O

ctob

er,2

006

–p.

20

Page 21: p · 2006. 10. 9. · distrib ute themselv es o v er a number of tar gets (e.g. fires) noisy, non-linear dynamics in continuous space-time additi v e control of the ... 0.4 0.6 0.8

Sum

mar

yan

ddi

scus

sion

We

stud

ied

optim

alco

ntro

lin

stoc

hast

icM

AS

inco

ntin

uous

spac

e-tim

e

Opt

imal

cont

rolc

anbe

deri

ved

from

the

solu

tion

ofH

amilt

on-J

acob

i-B

ellm

aneq

uatio

ns

Und

erso

me

cond

ition

s,th

elo

g-tr

ansf

orm

atio

ntr

ansf

orm

sth

eno

n-lin

ear

HJB

equa

tions

(ano

n-lin

ear

PDE

)in

toa

linea

rPD

E

unde

rth

ese

cond

ition

s,a

supe

rpos

ition

prin

cipl

eho

lds.

Thi

sen

able

sus

toge

nera

teM

AS

mul

ti-ta

rget

solu

tions

from

sing

leag

ents

ingl

e-ta

rget

solu

tions

Add

ition

alco

mpu

tatio

nalc

ostf

orM

AS

invo

lves

prob

abili

stic

infe

renc

e.w

hich

istr

acta

ble

insp

arse

lyco

nnec

ted

syst

ems.

Inde

nse

MA

Ss,e

xact

infe

renc

eis

intr

acta

ble;

ana

tura

lapp

roac

hw

ould

beap

prox

imat

ein

fere

nce

usin

gm

essa

gepa

ssin

gal

gori

thm

s.(c

urre

ntst

udy)

.

Inlin

ear

mod

els,

with

V=

0an

dno

agen

tint

erac

tions

,opt

imal

cont

roli

nM

AS

mul

ti-ta

rget

syst

ems

isso

lved

anal

ytic

ally

.

Inge

nera

l,ho

wev

er,e

ven

the

sing

leag

entp

robl

emre

quir

esco

mpu

tatio

nal

inte

nsiv

eap

prox

imat

ions

(Kap

pen,

2005

).W

.Wie

geri

nck

-U

vA,9

Oct

ober

,200

6–

p.21