From: AAAI Technical Report SS-94-07. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.
DESIGNING
SUPER-AGENTS
Sarosh Talukdar
Engineering Design ResearchCenter
CarnegieMellon University
Pittsburgh, PA15213
INTRODUCTION
A super-agent is an aggregation of lesser
agents, just as a person is an aggregationof
cells, a corporationis an aggregation
of people,a
flock is an aggregationof birds, and a complex
program is an aggregation of routines. The
aggregation schemes (organizations)
synthetic super-agents often use mechanisms
borrowedfrom other systems.For instance, the
hierarchical structures of moderncorporations
canbe traced backto thoseof ancienttribes, the
alternating cycles of creation anddestructionin
genetic algorithms havebeen adaptedfrom the
theory of evolution, and the arrangementsof
artificial neural nets havebeenborrowedfrom
models of the humanbrain. Wheremight the
designer of future super-agents look for
mechanismsto borrow and adapt? Howshould
such mechanisms
be combined?Moregenerally,
howshould the synthetic super-agentsof the
future be designed?
There is a vast body of literature
on
organizational design. It goes back to Adam
Smith’sworkon the division of labor ("An Inquiry
into the Nature and Causesof the Wealth of
Nations," 1776), and stretches evenfurther, to
SunTzu’s the Art of War(circa 500BC)and the
Book of Exodus(circa 1500 BC). Scattered
through this literature are organizational
structures that have, in certain circumstances,
been found to work well. Hierarchies, free
markets, task forces, assemblylines, genetic
algorithms, simulatedannealingand artificial
neural nets are someexamples.But there is no
uniform treatment of such structures, nor any
indicators of wherenewand better structures
maybe found. In other words, there is no
comprehensive space of organizational
structures, nor anywayto explore sucha space,
were it to be assembled. In the succeeding
material,I will attemptto make
gooda part of this
187
deficit. Specifically, I will describe a fairly
comprehensive space of organizations and
outline a procedure, basedon this space, for
designingorganizationsfor softwareagents.
SKILLS AS COMMODITIES
Thinkof a task (problem)as a demand
for skills
and an agent (problem-solver) as a supply
skills. Thenthe successfullycompletionof a task
is contingent on meetingits demand
for skills
fromone or moresupplies (agents).
TERMINOLOGY
Certain terms that are essential to the
succeedingdiscussionsare defined below.
Skill: anability to transform
a given(initial) object
or condition into a desired (goal) object
condition.
Operator:a computationalmodelof a skill. More
specifically, an operatoris a transformationfrom
one set of data-objects (called an in-set)
another (called an out-set). Thesesets may
arbitrarily large and the mappingarbitrarily
complex. In other words, an operator can be
usedto modelarbitrarily complex
skills.
Task:a demand
for skills. As such,a task canbe
modeled
as a needto mapan initial state (suchas
a set of specificationsfor a house)to a goalstate
( such as a set of blueprints that meet the
specifications).
Agent:a bundle of skills together with some
meansfor managingtheir deployment.As such,
an agent can be modeledas a set of controlled
operators. Thecontrols can be divided into two
categories: selectors and schedulers. The
selectors pick objects for the operators to
transform. The schedulers decide whenthese
transformations
will occur.
Autonomous
agent: one that makes all its
selection andschedulingdecisionsinternally. An
agent is non-autonomous(supervised) if
receives some selection or scheduling
instructions fromother agents.
Super-agent: an aggregation of agents.
(Super-agentsexist becausethe skills of any
single agent are limited and becausethere are
important tasks whose demandsfar exceed
theselimits.)
Organization:a structural modelof a superagent,just as circuit diagram
is a structural model
of a microelectronic chip and a blueprint is a
structural modelof a building. Morespecifically,
an organization identifies a set agents and
describeshowthey interact.
Cooperation: any exchange of data among
agents,whetherproductiveor not.
Openness:
a super-agent is open if it can be
assembled and disassembled easily. More
specifically, a super-agentis "opento a set of
agents, A," if members
of A caneasily be added
to, or removed
from, the super-agent.
Task-cover:
the net skills of a super-agent.In
other wordsthe set of skills it can synthesize
from those providedby its agents.Theextent of
the task-coverdependson howthis synthesis is
done and is often less than the union of the
agents’skills, but can,occasionally,be greater.
Effectiveness:
a super-agentis effective if it
performs its assigned tasks well. More
specifically, a super-agent
is effective overa task
if its coverincludes the skills demanded
by the
task.
Theintent is to design super-agentsthat are
specialized for the tasks they are to perform.
Havingcompletedthese tasks, the super-agents
could be disassembled, making their agents
available for reusein other super-agents.
Or they
could be expandedwith newagents to become
effective overevenmoredifficult tasks.
Notethat the extent of A is limited only by the
meansavailable for its members
to communicate.
If the forecasts for growth in communication
networks prove to be right, organizationdesignerscould, in the nearfuture, extendA to
include virtually every person, robot andother
computer-based
agent in the world.
REPRESENTING ORGANIZATIONS
This section arguesthat any super-agentcan be
thoughtof as a net of memories
andagents. This
net can be representedby a pair of directed
graphs, one to showthe paths by whichcontrol
flows in the super-agent,the other to showthe
paths taken by data. The focus is on
organizations for computer-based
agents. Other
sorts of agents, suchas insects, fish and even
people, are included
provided they
communicate
in waysno moresubtle than those
usedby computers.
Modeling Assumptions
1. The communication processes used by
agents can be simulated by shared memory
processes,that is, by agentsreadingfrom, and
writing to, memoriesthey share. (All the
communication processes used by computerbasedagentssatisfy this assumption.I do not
knowthe extent, if any, to which other agents
may
violateit.)
Manytypes of organizations sacrifice openness 2. The contents of each shared memoryare
for effectiveness, or vice-versa. Theproblem stored in the form of data-objects, that is,
consideredhere is howto design organizations
encapsulated packages of data. (This
that makesuchsacrifices unnecessary.In other
assumptionhasno effect on modelfidelity. Its
words:
only purpose is to establish a convenient
visualization device, called a state space, for
Given: T, a set of tasks,
eachmemory.This spaceis the set of all the
A, a space(set) of existing and
data-objectsthat canbe stored in the memory.)
anticipated agents,and
C, a meansby whichthe agentsin
Data And Control Flows
The definition of an agent together with the
A can communicate;
Design: k*, anorganizationthat is effective over modeling assumptions allow one think of an
T andopento A.
organization as a net , k, of memoriesand
TASK-SPECIFIC ORGANIZATIONS: THE
DESIGN PROBLEM
The effectiveness and opennessof a superagentdepend
on its organization,that is, on its
agentsand howthey interact. Thesameagents,
interacting in different wayscan have very
different levels of effectivenessandopenness.
188
agents, andfurther, to represent this net by a
double:
k(a) = {Df(a),Cf(a)}
wherea is the set of agentsin the organization,
andDf andCf are graphs.Morespecifically, Cf is
a directed graph, called a control flow, whose
nodes denote agents, and whosearcs denote
the supervisory relations amongthese agents.
Df is a special type of directed graph, called a
data flow, whosenodes are Venndiagrams.
Eachof these diagramsshowsthe intersections
betweenthe state spaceof a memory
andthe inand out-spacesof the agents that share (can
read from or write to) that memory.Arcs in Df
identify the in- andout-spaces
of eachagent.
Notethat:
¯ If the control flow of an organizationis null (a
graph with no edges), then its agents are
autonomous.
¯ Agentsin disconnected
parts of a data flow can
cooperate(exchangeinformation) in only very
limited ways--by issuing commands
that must
travel over the routes, if any, providedby the
controlflow.
¯ Most organizations have a few memoriesto
which agents write infrequently, if at all.
Henceforth, such memories will be called
referencememories
to distinguish themfrom the
other memorieswhich will be called working
memories.The purpose of a reference memory
is to store informationof long-termvalue, suchas
reusable knowledge, learned knowledge, and
the invariant parts of task descriptions. The
purposeof the working memoriesis store the
transient results of an organization’sactivity. As
weshall see, workingmemories
reflect the task
decomposition
scheme used by the
organization. In fact, eachworking memory
is
dedicatedto somesub-task.
Organization Spaces
Let A be the given set of agentsandS(A), called
the organizationspaceof A, be the set of all the
different waysin whichthe members
of A canbe
organized.In other words,S(A)is the set of all
k(a) for all a, where
a is a subsetof
A TAXONOMY OF ORGANIZATIONS
Even whenA is small and homogeneous,
S(A)
can be large and surprisingly diverse. In the
189
following material we will partition S(A) into
smaller morehomogeneous
regions.
Feedback and Autonomy
Suppose
that S(A) is partitioned alongits dataflow-axis into organizationswith strongly cyclic
data flows on one side and organizations with
essentially acyclic data flows on the other;
supposealso that S(A) is partitioned along its
control-flow-axis into organizationswith mostly
autonomous agents on one side and mostly
non-autonomous
agents on the other (precise
membership
functions for these fuzzy partitions
are not importanthere). Asa result, S(A)is
into four qualitativelydifferentregions(Fig. 1).
Organizationsfrom the two upperregionsof Fig.
1 rely on the use of supervisors, and usually,
hierarchical control flows. Consequently,they
tend to be moreefficient in their useof agents
but less robust. (Agentsthat are told whatto do
by their supervisorsare less at risk of workingat
cross purposes than autonomousagents. On
the downside, supervisoryerrors are amplified
by hierarchicalcontrol flows, andthe loss of a key
supervisorcanbe crippling to the organization.)
Organizationsfrom the two regions on the right
of Fig. 1 useacyclic data flows. This makesthem
well suitedto repetitive androutine tasks,but ill
suited to tasks that are uncertain or have
conflicting requirements, becausesuch tasks
invariably requirefeedback
anditeration, that is,
cyclic dataflows.
Differential,
Integral and Hybrid TaskCovers
Tobe effective over a task, the cover(set of net
skills) of a super-agentmustinclude the skills
demandedby the task. This cover can have
threeforms:differential, integral andhybrid.
In a differential cover, the overall task is
decomposedinto sub tasks so each can be
assigned to a single agent. (Note: these sub
tasks do not have to be disjoint, somemay
overlap in the skills they demand,
and somemay
combineor otherwise use the results of other
tasks.) An integral organization does not
decompose
the task; rather, it integrates the
skills of its agentsto meetthe demands
of the
overall task. A hybrid organization is part
differential,part integral.
Processes by which differential covers are
formed have been carefully studied for many
centuries(see, for instance,the writings of Adam
Smith,1776)undertitles suchas "the division of
labor" and "the delegation of responsibility."
They are nowwell understood and intuitive.
Integral coversare muchless familiar. Bywayof
an illustration, considerthe social insects(ants,
termites, certain beesandwasps).For them, the
construction of a nest is a massiveengineering
project requiring the cooperation of many
workers.Invariably, the nest is sophisticatedin
architectureandcustomized
for its surroundings:
features that are remarkablefor four reasons.
First, there are no leadersto decidewhatis to be
done,nor anycentralizedcoordinationof effort.
Second,the workershave accessonly to local
information(in spaceas well as in time). Third,the
individual insect hasonly modest
capabilities for
learning, reasoning,andabstracting knowledge.
Fourth, there are no blueprints to showwhatthe
final result should be nor any centralized
planning.In other words,there are no provisions
for decomposing
the overall task into subtasks,
nor any assignmentof sub tasks to individual
workers. Rather, the insects act as autonomous
agents and attack the overall task en masse.
Plans, designs and other products of task
decomposition, if they exist at all, are
unobservable.In still other words, the insects
useintegral covers:their organizationsintegrate
their skills to coverthe demands
of large tasks
and in the process, allow them to do so much
withsolittle (in thewayof individualskills.
Notethat the form of a super-agent’stask-cover
is determined
by the formof its data flow: if this
flow has only one workingmemory,the cover is
integral, if it hasseveral workingmemories
but
only oneoperator terminating at eachof these,
then the cover is differential, otherwise, the
coveris a hybrid. Also, in a differential or hybrid
cover, eachworkingmemory
is dedicatedto one
task.
It follows that eachof the regionsof Fig. 1 canbe
divided into three sub-regionsby partitioning it
along the data-flow-axis corresponding to
differential,
integral and hybrid covers.
190
Organizations
that usedifferential coversare the
most familiar, and therefore, the easiest to
design, but hybrid and integral covers have
somevery desirable properties that we that we
do not havethe spaceto discusshere.
A DESIGN PROTOCOL
The taxonomy above suggests a design
protocol with three stagesthat maybe performed
once,but moreoften, require iteration:
1. Basedon the characteristicsof the tasks to be
performed(whether they are routine or not,
whether centralized planning and control
would seem to be indicated or not, and
whetherthe tasks are decomposable
or not),
select one of the 12 sub regions of the
organizationspace. This selection determines
the grosscharacteristics of the organization’s
data andcontrol flows.
2. Complete
the data flow. Thesteps are:
¯ Decompose
the tasks into sub tasks. Designa
working memoryfor each sub task (choose
howmanyestimatesto the sub task-solution
will be held by the memoryand how these
estimatesare to be represented).
¯ Designandstock reference memories
for each
subtask.
¯ Choosethe mix of agents to write to each
workingmemory
payingparticular attention to
the balance betweentheir construction and
destructionskills.
3. Complete
the control flow. Thesteps are:
¯ Designthe selection controls for eachagent.
¯ Design the scheduling controls for each
agent.
¯ Distribute thesecontrols, that is, designthe
supervisorylinks among
the agents.
ASYNCHRONOUS TEAMS
Oneof the least explored and mostinteresting
parts of organization
space(Fig. 1) is in its SouthWestcorner. Specifically, it is the subsubregion
of stronglycyclic dataflows, virtually null control
flows (most, if not all, the agents are
autonomous),hybrid covers and asynchronous
cooperation. (Recall that agents are said to
cooperateif they exchange
data. Wewill say that
they cooperateasynchronouslyif the exchanges
occurso no agentis requiredto pauseor wait for
data being producedby another. In other words,
asynchronous
cooperationallows all the agents
to workin parallel all the time.) Wewill call
membersof this subsubregion asynchronous
teams(A-Teams,
for short).
Advantages and Disadvantages
Autonomous agents that cooperate
asynchronouslyhave considerable advantages.
Theyyield super-agentsthat are extremelyopen
and that can mountmassively parallel efforts
without the needfor centralized schedulersor
any fear of deadlocks. But how is all this
computationaleffort to be put to gooduse?Any
agent,able to decidefor itself what,if anything,
to do could chooseto sit idle or evenundothe
work of its fellow agents. When
computercosts
were high, such individual freedomsand the
resultant wastage of computer cycles, were
unthinkable. But now, with costs decreasingby
=,o
tO
O
¯
¯
¯
¯
¯
NONAUTONOMOUSAGENTS
CYCLIC DATA FLOW
Blackboards
GeneticAlgorithms
SubsumptionArchitectures
SimulatedAnnealing
MostCorporations
BB BB BB BB BB BII
¯
¯
¯
¯
¯
¯
¯
BB BB lib
AUTONOMOUSAGENTS
CYCLIC DATA FLOW
Cellular Communities
Insect Societies
Schoolsof Fish
RecurrentNeural Nets
Free Markets
Scientific Communities
AsynchronousTeams
the month, large networks of computers,each
dedicated to one agent, can be assembled.
Individual productivity is of little consequence.
Someagents may be unproductive or even
counterproductive as long as the collection of
agents,as a whole,is effective in performingits
assigned tasks. How may this overall
effectiveness be achieved?Wehave developed
a set of case studies that provide some
guidance. Spaceconstraints allow meto only
partially list these case studies here. They
include: traveling salesmenproblems,high-rise
building design, task-specific robot design,
network diagnosis, contingency planning for
electric networks, train scheduling, job-shop
schedulingandsteel mill scheduling.
m
m
II
m
m
NONAUTONOMOUSAGENTS
ACYCLIC DATA FLOW
¯ AssemblyLines
¯ MostReal-TimeControl
¯ Most CADPackages
II
BB imB m
m
II
I
I
I
m
m
m
m
m
m
m
mm
m
AUTONOMOUSAGENTS
ACYCLIC DATA FLOW
¯ ForwardNeural Nets
¯ Systolic Arrays
¯ Bucket Brigades
I
l
data flow
Fig. 1. A taxonomy
of super-agents
obtainedby partitioning S(k), the spaceof all super-agents,into four
fuzzyregions. All the regionscontain natural andsynthetic super-agents
but the distribution is far from
uniform.Specifically, there are fewexamples
of syntheticsuper-agents
in the lowerleft region. In contrast,
manynatural super-agentshavefound it advantageous
to reside there. Eachof the abovefour regions
canbe dividedinto threesubregionsby the formof its task-cover(differential, integral or hybrid).
191