Utility of Communication

This section describes the role of communication and the

mea.sur<' of necessity of

conlmnnicatiou

in

the con text.

In distributed sy stems in 'vhich communication costs cau

not be ig

norC'd

,

agents or 1 ro­

grammers should se

l

ect

a

communica.tion structure. But there is uncertai

n

ty about the statr of other agents.

Thi.

is cau ed by both

a limited view of ag

C'

n

t.s and the iuhC'rent JH'OJH'rtirs of problems. Le. ser c

a

lled

it

c

oo

per

a

tive control uncerta.intic>s in

[

Les91

]

.

lost. distributed

programs do not suffer

from it. because a

communication structure

is

fixC'cl by

a programmer

to assure the correctness of execution. As ^a. result,

communication is

r

<

'qu

is

i

t

.C' ^<Ul l its

patt<'nt is

fixed. A DAI system, however. has another kind of

communication,

n

a

Juely optionaJ

('0111-munication. The purpose of the optional

communication is to share

k

no

wlrdg(

'

that is acquir<'d during execution, but is also not required for finishing

a job.

Thus ;.dl optical cO!lllllllllinLtious do not. require multicast. The

reasons

are the

following:

• Some pieces of knowledge are found in some agents si

m

ul

ta

nr

o

usly. T

o

av

o

id dup

l

iⁿt

t<

•d communication, the range of communication should

h<' controll<'d to au

appropriate'

size'.

• In some case ch

a

ng

i

ng the range

of communicatiou docs not a

ffect the cotTc'cturss of the program. It ch

a

nges only the pe

r

f

or

m

a

nce

of the program. Evc'u if t h<'

rcwgr

of

communication changes the quality of the answer, it d

o

es

not mean

that broadca.st is the best communication m thod since there would be

two agents that cU'<' ind p<'udeut

of each other.

•

ew knowledge woulcl makr another Oil<' out of datr.

If

Hew knowlrclgr

is

fonud

fn'qtH'lltl,\\

waiting for new knowleclgr makes anothrr pieces of information ohsolrt<' and n'dncrs

the

communication co.-t without performance detrriora.tion.

Therefore, if it can he allowed. redncing the communication

^SII.c

to an appropriate

stZ<'

IS

desirable on distributed sy'trms. This is a kind of di

s

tr

i l^mt

<

^'d consistency uutiutc'tl<l.ll<'<'

[FL77, GBH87].

Now it is required to balanrr the desire to share new knowledge and to

k<'<'P commtmicatiou

costs smalL in other words, to maximize the . atisfiabili ty

of

communication.

\Y<'

can

arhic'Y<'

it by introducing the

utility of communication.

Here thr utility of communication

is

ddi11Pd

as

the relative benefit of commuuicatioH ag<:unst. it co t. Figurr

2.5

illustrates the r

ela

.

^t

io

ll hC'twl�<'ll

communication den ity and the utility. The clotted line shows thr amount of tltr information that each agent has after communication. By communicating frequently or

widrly,

agC'nts can hold

^a.

more coherent and global view of other agents. But we can assunH' the amount of information is bounded in the domain of distributed problem solving. The solid line shows the cost of communication, which is monotonically increasing. Thus the utility clrcreasrs. To achieve the ma.ximum communication effect, agent. should he in tbC' shadrd

n'p;ioll. T

h

<' l><•st

balanced point i. th point that maximizes the utility.

A

n'ason that I\itcuuuras'

algoritl1111

does not work well with a large cost

might

be that thry did not introducC' the' idc'a of ntility ( see section _2.1.4).

Of course, with this definition, calculating the utility of communication is

difficult.

IH•cansc·

it depends on the global state of the system. Agents on distributed sy stc'ms can uot gC't the current status of the environment; th

y

can only make au incomplete'. partial

modd

of execution. Thus the utility of communication that determines how agents exchang<>

infonnatiou

should be calculated from

^a.

model of local

computation.

The local model is a

history of

a part of the system that an agent can investigate.

Now

we give uch a model for cooperative search. In

^a co

o

p

erati

Yc

search, agru ts usC' and

exchange some pieces of informatiou about the problem in order to n•cluc<' the sra.rch space'.

c 0

� E

� 0

(I) 0 (.)

i olated

cost information

connectivity between agents

broadcast blackboard

Figure 2.5: balancing utility and cost

Since they are acquired dynamically, an agent can not forecast the current valn<' of infonnat.iou1 that other agents have without communication. But under the assumption of homog<'H<'ity of the agents and the continuity of local computation, we can cou.·ickr the distribution of the renewal of information of an agent, which is calculated from its history. 'i!i that of anothrr agent. Thus the current values of information that other agents have cau be estimated from its history locally. This simplification makes building a modc'l Y<'ry rasy. In tbis case', botlt the communication costs and the amount of reduced ta.sks due to th<' acqnir<'d iHfonmtt ion determine the utility. Therefore in order to decide how to exchange information or uusolY<'d sub-problems between agents. an agent can use its partial history as a model of cx<'cutiou.

Precise knowledge with a poor cooperativr strategy sonwtinws lc'ads to selfish behavior.

Huberman et al. described such a phrnomcnon on shared object management iu

[HH ].

^For

example, let us con ider about a group of searchers. If agents share complet<' knowlc'd�<>, silH"<' they have the same knowledge, we can not expect suprr-linrar sperd np t.ltat emerges froru the diversity of heterogeneous agents. NK-rnodel

[

^I\:au93

]

^{a.s a} sdf C'Yolntional system mi�ht i><' another example^. In the NK-modcl, agents arc automata. At each time, c V<'ry agl'nt change's its state corresponding to the input (its environment) that is the states of the fixNl neighbor

1What we call information here is a object Star called ideal object' or 'map' in [Sta 0].

agents. The change of its sta.tr afft>cts th<' neighborc· at the lH'Xt tick. Thus itnactions amoug agents make complex feed-back loop�. This is a model of the interaction among <Utimals. Au<l the quality of the. ystem is dt'trrmiuccl by the sum of each agrnts' adaptability. If ag<'llts arr very sen itive to thC' changC' of their C'nviroumrnts, thC' cascade of changing h('haviors nut Hot top. The. ystcm in a positin' feedback loop could not hr adapted to tlH' situation. Oil thr otlH'r hand. the group of very robust agents can not be adaptrd to thr important chang('S iu tltrir environments. 1\:auffman claimed that most cvolntional systrms would sPlcct an appropriat(' connectivity between agents by evolution itself. Since th<' evolutionality of t h(' I\:-modd is corresponding to the adaptability of agents in eli tributed environmC'nts, we think that th(' same

result would be held in designing the organization of problC'm-solviug agents. Thrrdorr the nd utility of communication should not be measured by the amount of information to srucl, but the amount of contribution to solving thr problem of the informatiou. A prohkm-dep('Ud<'nt measure will be required.

ドキュメント内 協調探索における通信戦略の研究 (ページ 34-37)