Sie sind auf Seite 1von 18

Can Agents Cover All the World?

Kai-H. Brassel1 , Michael Möhring2, Elke Schumacher2, and Klaus G. Troitzsch2


1 Institut für Soziologie, Technische Hochschule Darmstadt, D–64283 Darmstadt,
Germany
2 Institut für Sozialwissenschaftliche Informatik, Fachbereich Informatik, Univer-
sität Koblenz-Landau, Rheinau 1, D–56075 Koblenz, Germany

Abstract. This paper tries to answer the question whether the multi-agent ap-
proach to social science simulation embeds all other types of social simulation, and
in which cases it is sensible to use multi-agent methods for modeling. To find this
answer, we classify earlier modeling approaches and discuss which kinds of agents
must be used to replace a traditional model by a multi-agent model. We use a test case
to analyze problems of compatibility between agent kinds and interaction modes.

1 Is there a multi-agent paradigm?


Since the early sixties, social scientists try to simulate human social behavior. In
many papers, state or behavioral changes of human individuals, groups, organiza-
tions, or nations, or of the world as a whole have been represented (and simplified!)
in deterministic or stochastic models (for an overview and, at the same time, a taxon-
omy, see section 2). Beside the classical approaches, there were very early attempts
to rule-based modeling and simulation in the social sciences (see e.g. Sola-Pool &
Abelson 1962, Abelson & Bernstein 1963, Abelson & Carroll 1965, Abelson 1968)
which were not very successful, mostly due to hardware restrictions, and which had
no direct successors.
In the last few years, the focus of social science modeling and simulation has
moved to (or at least widened to) the application of methods developed in Dis-
tributed Artificial Intelligence. Generally speaking, DAI is concerned with sets of
units (agents) existing in an environment which they have in common. They com-
municate among each other and cooperate in the solution of complex problems by
adding their particular capabilities.
The main difference between multi-agent simulation and all earlier approaches
to social science simulation comes from the fact that multi-agent simulation makes
heavy use of artificial intelligence methods. Its kinship to classical mathematics and
stochastics is rather loose; it is less quantitative and more qualitative than most other
simulation approaches, invariants do not come in the form of equations, but in the
form of rules, and this makes it the most advanced (and the most appropriate) simu-
lation approach in social science both in a technical and methodological perspective
because models of this type “can embed all the other types, although it is not, of
course, necessarily sensible to do such a thing.” (Doran & Gilbert 1994, 10)
Multi-agent modeling “sees the world” as consisting of agents which may come
in several kinds. There does not seem to be a definition of “agent” which is agreed
upon. For the moment, we keep to the distinction into three kinds of agents (see
Wooldridge & Jennings 1995, Moulin & Chaib-Draa 1996, 8ff.):

Reactive agents react to messages from their surroundings by sending other mes-
sages to other agents and by actualizing the inner representation of their sur-
roundings. All this happens according to fixed rules or plans which cannot be
changed by these agents.
Intentional agents have the same capabilities as reactive agents. Applying
“metarules”, they are moreover capable of defining goals, e.g. depending on
their motivation or their needs. They can detect conflicts between goals, set
priorities, and design plans to achieve their goals, and they can be informed
about each other’s goals, assumptions, and actions.
Social agents additionally have explicit models of other agents. This is why they
are capable to reason about other agents’ goals, expectations, motives, and ca-
pabilities, and to include them into their action plans.

Kinds of agents are, of course, part of the model, not of the target system (i.e.
the part of reality which is being modeled). At the modeler’s discretion, “real-world
agents” may be modeled by model agents of whatever kind (see the discussion in
subsection 3.2), and it is at the modeler’s risk whether a human being is adequately
modeled by a reactive agent (or a falling stone by an intentional agent).
Especially social science models require a further distinction in indivisible agents
and aggregate (or systemic) agents. Indivisible agents do not consist of other agents,
aggregate agents are made up of other agents, but may interact with their surround-
ings in the same manner as indivisible agents.
The main question of our paper — and of our current work in our massif 1 project
— is just this: Can multi-agent models embed all the other types, and is it sensible
to do such a thing? Is multi-agent simulation a new paradigm?
Among the many meanings the word “paradigm” has in the philosophy of sci-
ence literature, at least a few are appropriate in the case of multi-agent modeling
and simulation: it is (1) a source of tools, or (2) a tool-producing factory, and (3)
it might open up a new “way of seeing”. Thus it fulfills Masterman’s (1970) defi-
nition of an “artefact paradigm” or constructed paradigm as something that makes
tools available. In so far as it opens a new way of seeing, it can be subsumed un-
der the term “metaphysical paradigm”. Whether it is also subsumable under Mas-
terman’s category of “sociological paradigms” is subject to further discussion — is
it really a widely accepted scientific achievement? Although the term “paradigm”
1 massif is the abbreviation of the German project title “Multi-Agenten-Simulation in der

sozialwissenschaftlich-interdisziplinären Forschung” (multi-agent simulation in social science in-


terdisciplinary research).
has been given a very formal definition (see, e.g., Balzer et al. 1987, 223) its use is
still ambiguous. Computer scientists, for instance, seem to use the term “paradigm”
for a new method(ology) which is more effective, i.e. solves a greater number of
problems, which is more directive, i.e. helps to formulate new problems, and which
is more elegant, i.e. allows to combine old solutions and to reformulate them in a
more compact manner.
We may expect that the multi-agent modeling and simulation approach will be
more effective because distributed systems of intelligent units should be more effec-
tive than isolated intelligent systems or distributed systems of non-intelligent units.
We may also expect that the “new way of seeing” multi-agent modeling offers will
help to formulate new problems (and, eventually, help to solve them). But so far, this
is only expectation. We will concentrate on the third requirement: that the multi-
agent approach will help to unify old solutions and to reformulate them in a more
compact manner.
The last requirement can be said to be fulfilled with respect to social science the-
ory building and methodology if the multi-agent approach is able to cover all the
earlier approaches to modeling in the social sciences. This is why we first try and
give an overview of earlier approaches and analyze the capacity of the multi-agent
approach to achieve the same as and more than the respective classical approach.

2 Catching the diversity of modeling and simulation


approaches
Modeling approaches may be distinguished according to several different character-
istics:

Levels and object types: One of the most important of these characteristics is the
number of different object types representing real world phenomena. On the one
hand, we have macro models in which one single object represents all the phenom-
ena at stake; whatever has to be modeled will be represented by this single object’s
attributes and by invariants between these attributes. On the other hand we have mi-
cro (or, better, multilevel) models in which at least two types of objects represent
different kinds of elements of the real world; here relations between objects have to
be defined, and invariants hold for attributes of different objects. Objects may be
grouped to different levels, with typically one instance of one object type at the top
of the level “hierarchy”, representing the target system as a whole, and several in-
stances on lower levels, representing, e.g., groups, or different types of individuals.
The introduction of new object types and new levels, respectively, is also a ques-
tion of model refinement: a group object may be dissolved into a set of individual
interacting objects or agents forming this group, and on the other hand, a number
of sufficiently similar objects may be combined, and consequently abstracted, into a
group object which may sufficiently be described at the group level, thus explicitly
disregarding any effects which are due to the interactions on the individual level.
Attribute domains and topography: Two other important characteristics are con-
cerned with the domains of the attributes and with (physical) space.
Attributes may be considered as continuous (which, in a strict sense, applies only
to mathematical models, though computer simulation is able to approximate con-
tinuous attributes to any desired degree), or they may be considered as integers or
as ordinally or nominally scaled, i.e. as discrete. Topographical space has seldom
been used in social simulation; it may be modeled as discrete like the squares on a
checkerboard, or continuously with real valued coordinates, or as a network reducing
the topography to a graph of reachabilities.

Time and Synchronicity: Time may be considered as continuous (which, in a


strict sense, applies only to mathematical models, though computer simulation is
able to approximate continuous time to any desired degree), or it may be considered
as a series of distinct points of time, equidistant or not. In the latter case we have
event oriented models in which time proceeds from event to event. Another impor-
tant characteristic is whether a (multi-object) model is evaluated in a synchronous or
asynchronous manner. In the first case, all objects are driven by the same clock, and
each clock tic makes each object change its state (including the case that the state is
not changed). In the asynchronous case, only selected objects perform a state change
at a given time. In the case of asynchronous models there is also the question whether
all types of objects use the same or their respective time scales (a complete econo-
metric model might be event driven as far as consumers’ purchases are concerned,
proceed by months as far as consumers’ incomes are concerned, and by years as far
as their tax payment is concerned).

Stochasticity: It makes a difference whether a model is deterministic or stochas-


tic. In the former case, both random effects and measurement errors are explicitly
neglected, while in the latter both can be included, either in one term or separately.
Birth and death processes may be explicitly considered in stochastic models, while in
deterministic models populations grow or decay continuously. The choice between
stochastic and deterministic elements of a model is also a question of abstraction and
refinement: in the process of model refinement (abstraction), e.g. birth and death
processes may replace (be replaced by) a differential equation describing growth in
a continuous manner.

Linearity: A less important distinction is between linear and nonlinear models.


This should be considered last because it is about the exact description of the laws
governing reality by mathematical invariants — though it is often considered first
(or rather: not considered at all) since a decision in favour of linear models is always
tempting because of their analytical solvability and ease of parameter estimation.

Tools and purposes: Two further distinctions used later on are concerned with the
availability of simulation tools and with the purpose (explanatory, predictive, deci-
sion support, didactic, etc.) for which modeling and simulation are done.
Computer simulation models thus will nearly cover the full range of conceivable
models; compared to other kinds of formal models, there are very few restrictions
— although all the existing traditions of social science computer simulation impose
very strict restrictions on the models they make possible, and this is why we now
have to discuss some well known approaches to social science modeling along the
lines of the above characteristics (omitting the question of linearity and nonlinearity,
since it is not discriminating — see table 1). We begin with Systems Dynamics which
obviously has its roots in systems of differential equations (Forrester 1968, Sec. 3.3)
from which it seems to differ mostly in two technical aspects: discrete time is used
as a coarse approximation for continuous time to achieve numerical solutions, and
functions of all kinds, including table functions, can be used with the help of the
available tools like DYNAMO or STELLA. Systems Dynamics is restricted to the
macro level. This is why it might be replaced by “multi-agent models” which consist
of only one agent — which obviously is of the “reactive agent” type. This agent
would be rather complex, having lots of attributes.

Characteristic Approaches
Systems Microanalytical Discrete Event Multilevel Cellular
Dynamics Simulation Models Models Models Automata

levels micro and several,


macro macro, no micro feedback micro
feedback

attributes continuous both both both discrete

time approximately discrete event continuous discrete


continuous oriented or discrete

topography n.a. n.a. network n.a. grid

evaluation n.a. synchronous asynchronous synchronous both

state change deterministic stochastic stochastic both both

tools DYNAMO, STELLA none many MIMOSE none

purpose prediction prediction prediction explanation explanation

Table 1: Taxonomy and classification of common social science modeling ap-


proaches (entries are prevailing characteristics)

For chronological reasons we turn next to microanalytical simulation models


(MSM) (Orcutt et al. 1986) whose mathematical kin seems to be the Markov chain.
They were “originally devised to facilitate effective use of the output of microana-
lytical research” (Orcutt 1986, 13), and they aim at predicting effects of (and thereby
supporting) social and financial policy. The main difference between Systems Dy-
namics and microanalytic simulation models is the same as the difference between
deterministic macro models and stochastic micro models. In Henize’s (1984, 571)
classification, microanalytic simulation models show up as static — which seems to
be due to the fact that there is usually no feedback from the reactions of microentities
on the macro policies. Since microanalytic simulation models usually use detailed
empirical information about the initial state of microunits such as persons and fam-
ilies (these data are often, but not always, qualitative), they are not only extremely
data based but they are extremely expensive, too, both for the data collection ex-
penses and for the necessary computing and data storing capacity. Their extreme
expensiveness, however, is not only due to data collection and to the peculiarities
of the micro level, but it also originates in the manner in which (and the tools with
which) microanalytic simulation models are designed and implemented. “There is
no universal general-purpose microsimulation software available ..., most MSMs are
developed in a conventional way from scratch” (Klösgen 1986, 485–486), i.e. they
have to be programmed in a general-purpose language. In terms of multi-agent mod-
eling, an MSM would be a model consisting of a large number of reactive agents: in
classical MSMs there are no individual objects defining or changing their goals and
developing plans, nor is there a need for agents to inform each other about their goals
or plans. A great progress in modeling could, however, be made if classical MSMs
could be amplified by giving their individual objects new, realistic capabilities of in-
tentional or even social agents (see, e.g., Gilbert & Chattoe 1995). MSMs, moreover,
raise the problem of aggregate agents (like households, e.g.).
Discrete event models have found wide use in economics and management sci-
ence. Outside social science, simulation in most cases means just queuing simula-
tion, since in most areas of applied science queuing models proved to be an adequate
modeling paradigm. Discrete event or queuing models consist of several types of
objects, namely static objects like servers and queues, and dynamic objects like cus-
tomers (which in turn are generated by sources and removed in sinks). Objects of
the static kind are linked together in networks, and customers move between servers
and queues on these links. Processes modeled include arrival, wait, and serve pro-
cesses. Time does not pass in equidistant time steps, but from event to event, so next
events have to be collected in an agenda and executed by a scheduler (which has no
counterpart in the modeled target system). In terms of multi-agent models, sched-
uler, servers and queues would be reactive agents (queues, moreover, of the aggre-
gate version), customers would be intentional agents. An amplification of queuing
models in the multi-agent paradigm could make both servers and customers social
agents, thus for instance enabling them to negotiate during the serving process.
The next approach is somewhat heterogeneous. What keeps it together is the fact
that all models merged under the roof of this approach are multilevel (which sepa-
rates them from Systems Dynamics) and incorporate feedbacks between levels and
in some cases also between objects of the same level (which separates them from
the classical microanalytical simulation models). Their next mathematical kin is the
mathematics of stochastic processes and of synergetics (Haken 1978, 1988) (master
equation, Fokker-Planck equation), although network modeling and dynamical game
theory have quite different roots. All the approaches in this group have in common
an emphasis not on prediction, but on explanation. In terms of multi-agent model-
ing, they consist of a large number of one type or very few types of homogeneous
reactive agents (which, of course, may be aggregated on a number of levels, thus ne-
cessitating a small number of additional types of aggregate reactive agents). Models
of this kind may be easily conceived of as enlarged by intentional and social agents’
capabilities, which would make these models far more realistic.
Although cellular automata can be seen as a special variant of multilevel mod-
els, it is in order to treat them separately in this overview. This is due to two special
features: unlike nearly all other social simulation models, they have a clear topog-
raphy — in most cases a two-dimensional rectangular grid —, and the interactions
modeled within the cellular automata approach is strictly local. Cells are reactive
agents, again, which act and interact due to fixed rules, and all of them apply the
same rules. Cells’ state spaces are identical, and in most cases contain only two or
very few states. Some cellular automata allow migration of cell “inhabitants” from
their cell into another cell in its neighborhood. In terms of multi-agent modeling, cel-
lular automata can be described in the same manner as multilevel models, restricted
to exactly one type of reactive agents, — with one exception, which is due to the
grid topology which must explicitly be modeled as an agent with quite poor capabil-
ities to notify the cell agents which cell they inhabit and which cells belong to their
neighborhood.
This short overview may have shown that all common modeling and simula-
tion approaches to social phenomena can be embedded in the multi-agent paradigm.
Moreover, it should have become clear that the multi-agent paradigm can add some
new insights that the earlier approaches, due to their respective restrictions, could
not.

3 Test case
The aim of introducing the following example is twofold: First, it will show that
describing models of social phenomena as close to reality as possible necessarily re-
quires the use of different modeling approaches. Of course, the social phenomenon
referred to here is far more complex than the following model description — it is just
as complex as necessary to show that no earlier approach is sufficient by itself. Sec-
ond, it will show that embedding models in terms of the the multi-agent paradigm
provides an integrated view and can add some new insights that singular approaches
could not.
What we do not aim at with the following example is a realistic model of a town’s
banking system with its environment, but we felt that readers will be familiar enough
with bank branches, their clients, and the relations and interactions between them to
accept this example as a test case for a simulation tool to be developed.

3.1 An informal description


The model describes the development of a bank system, depending mainly on the
competition between banks providing services for business clients. These services
are restricted to loans and investments. Each business client chooses the bank with
the “best” conditions, according to a negotiation process. A bank invests its profit in
different ways to improve its position in the bank system.
Base elements of the model are a set of banks and business clients, both located in
a two-dimensional geographical environment town. A second environment element
is the federal reserve bank, which determines general loan and investment conditions
(i.e. rate of discount) and publishes them to all model elements periodically. Each
bank consists of a main branch and branches. All branches get the concrete loan and
investment conditions (i.e. margins of interest rates) from their main branch, and
send back reports of the transacted loan and investment contracts. At the beginning,
each branch consists of two counters, one for loan services and one for investment
services.
To raise a loan or to invest money a business client requests an offer from each
of its most favored banks. According to these offers, the business client chooses the
bank with the “best” conditions and starts a negotiation process with a bank clerk,
responsible for the corresponding service.
The income of banks and business clients are changed by loans or investments in
a different way. For example, raising a loan by a business client increases his income
once but decreases it periodically by paying back the loan rates. In contrast to this,
granting a loan by a bank decreases its income once but increases it periodically.
Each bank can invest its profit in a different way: founding new branches or im-
proving services (i.e. reducing queuing time) by opening new loan or investment
service counters.

3.2 A graphical representation of the test case


Picture 1 shows a graphical representation of this example, based on the object mod-
eling technique (OMT) by Rumbaugh (1991).
The vertices represent agent types (in OMT: object classes), while the solid edges
describe structural relationships between agents2 . Additionally, the dashed edges de-
scribe interaction relationships, annotated by concrete actions. According to the kind
of agents presented in chapter 1, the agent types, which can be seen only as abstrac-
tions of entities in reality, are characterized as follows:

Reactive Agents

town  3
Aggregation of banks and business clients (“ ” in Picture 1)
 Provides a two-dimensional geographical space, in which all
system agents are located.
federal reserve bank  Calculates general loan and investment conditions by
fixed rules.
 Sends actual loan and investment conditions to all main
branches and business clients.

2 Forexample, a
2+-u b means “each agent of an agent class a has a relation to two
or more agents of class b”.
federal
town reserve
bank

environment
(bank) system level
statistics
1+

bank
1+

business
statistics
client
money request
money grant
1+

bank level

1+
rates loan/investment
main 1+ offer
branch
branch results,statistics 1+ loan/investment
request

rates
results
2+

counter

counter level

2+

change location
clerk queue

loan/investment
consulting

loan investment
services services

agent types structural relations interaction relations

Figure 1: Object model of the bank system


 Sends requested money to the main branches.
banks  Aggregation of a main branch and branches (“ ”).3
counters  3
Aggregation of a queue and a clerk (“ ”).
queues  Contain business clients, waiting for consultation.
 Deliver business clients to clerks.
Intentional Agents
main branches  Calculate their financial needs and their own conditions (i.e.
interest rate margins) by analyzing the general conditions, the
results of their branches, and the bank specific goals and plans.
 Send money requests to the federal reserve bank.
 Send the conditions to all of their branches.
 Create/Delete branches according to their goals.
branches  Send loan and investment offers to business clients.
 Distribute general bank conditions to their clerks.
 Collect clerk results and send statistics to the main branch.
 Create/Delete counter according to bank goals (i.e. improving
customer service by reducing queuing time).
Social Agents
business clients  Determine loan and investment needs and individual inter-
est rate margins by analyzing parameters like current account,
plans, etc.
 Select banks by parameters like geographical distance, service
in history, published conditions etc.
 Send loan and investment requests to all of them.
 Select the most favored banks by comparing the received of-
fers with the individual interest rate.
 Negotiate with the clerk in charge of the selected banks for
raising a loan or investing money.
clerks  Generalization of clerks responsible for loan or investment
services (“4”).
 Receive general interest rate margins from the branch.
 Negotiate with business clients for granting a loan or investing
money by reasoning about their goals, expectations etc.
 Send results to the branch.

3.3 Requirements for the design of the model and of a tool


At the end of this section, the main characteristics (see the first part of section 2) of
the model example, which describe some important requirements for the design of
multi-agent modeling and simulation tools as well, are summarized.
Multilevel Modeling The integration of different (aggregation) levels allows the
analysis of relations between these levels:

aggregation/emergence: individual branch results ;! bank success,


feedback: interest rate margins of the bank ;! individual branch success.

Additionally, this may have consequences for the characterization of agent types on
different levels (e.g. populations of social agents that panic like reactive agents).

Synchronization of Different Time Models Usually the activities of different


agent types take place by using different time-advance mechanisms and different
time scales which must be synchronized to ensure a consistent model behavior. For
example, the federal reserve bank publishes the rate of discount at distinct equidis-
tant points of time (e.g. monthly), whereas a business client invests money or raises
a loan event-oriented, according to his individual state. Additional examples are
the foundation of branches (event-oriented), consulting clients by clerks (event-
oriented), or branch reports (e.g. weekly).

Birth/Death Processes Modeling social phenomena usually includes the dynamic


generation and removal of agents. In the model example, this is true for most of the
agents. Banks and business clients may be set up and go bankrupt, branches and
counters/clerks may be established and removed, depending on bank strategies.

Abstraction/Refinement Providing abstraction and refinement mechanisms sup-


ports the modeler in choosing an appropriate model granularity. For example, if the
waiting time until a business client is served by a clerk is not important, the detailed
description of a counter (counter level) may not be necessary. However, a more de-
tailed analysis of business clients suggests the modeling of small, medium, and big
sized business clients, and, perhaps, of private clients, too. Abstraction and refine-
ment plays also an important role in modeling interactions. For example, “negotia-
tion” describes an interaction and cooperation process between two agents (i.e. clerk,
business client). The inner structure of this process can be modeled more or less in
detail, depending mainly on the capabilities of the participants: Negotiation between
social agents is usually more complex than between intentional agents.

4 Some real problems and their massif solutions


Whereas the requirements mentioned in the previous subsection are more or less re-
lated to tool-building in general, the following two subsections are devoted to two
problems which arise only in the framework of multi-agent models: In classical sim-
ulation approaches there are only very few modes of interactions between objects
(usually by state change functions), the overall structure of the model system is con-
stant, and the environment of the target system is represented by parameters, while
in multi-agent models interactions should be designable in a wide variety of modes,
structures should be variable, and it is reasonable to conceive of common and special
environments.

4.1 Kinds of agents and modes of interaction in heterogeneous


multi-agent systems
A main requirement for a toolbox for multi-agent modeling is that it should provide
the user with the flexibility to choose between different kinds of agent models to rep-
resent real-world agents. It should support the design of heterogeneous multi-agent
systems, i.e. systems with agents differing in behavior and capabilities (following
for example the distinction into three kinds of agents taken in chapter 1). In this
subsection we show some of the difficulties arising out of this task and a promising
approach to cope with it.
A lot of theoretical and practical work has been done on agent interaction capa-
bilities (see e.g. Moulin & Chaib-Draa 1996 for an overview). The spectrum ranges
from very simple ones, such as avoiding collision, to high-level coordination tech-
niques. The modes of interaction an agent is capable of relates to the purpose he is
designed for and the tasks he has to achieve. For agents in complex social systems
comprising a more or less dense network of inter-agent relations one sort of capa-
bilities seems to be most essential: to be able to take part in communicative actions.
Without communication the ability to interact is obviously very limited (see, e.g.,
Werner 1996). What kind of messages a specific kind of agent is able to send and to
understand depends on his internal structure and control mechanisms. As mentioned
above, heterogeneous multi-agent systems comprise more than one kind of agents.
Thus, for the agents to understand each other, a framework is necessary that serves as
a common communication basis. For this purpose, agent communication languages
like KQML (Finin & Fritzon 1994) have been developed.
But the ability to exchange messages is just not enough for successfully perform-
ing interactions. Designing interaction patterns between different kinds of agents
necessitates constraints to be considered: The bandwidth of interaction capabilities
of simple reactive agents, which act upon their environment by stimulus-response
mechanisms and triggering rules, differs very much from the case of social agents,
as their plan-based behavior takes into account both their own mental states and those
of other agents. Yet the analysis of pre- and post-conditions holding for interacting
agents (mainly based on speech act theory) covers only one specific kind of agents
(see, e.g., Dignum & van Linder 1997). The task of specifying constraints between
different agent categories seems so far unsolved.
A tool just putting together agent models stemming from different paradigms (as
the models of a reactive, an intentional, and a social agent) and offering them to a
modeler suffers from an important shortcoming. The modeler is faced with the de-
cision which kind of proposed agent model best represents the real-world agent to
be described. Thus, the modeler is forced to take a somewhat artificial perspective.
Instead of solely regarding the properties of the part of the target-system to be mod-
eled, he has to judge whether the characteristic overall behavior of the chosen agent
model is appropriate. In some cases this may lead to undesired compromises.
An approach to overcome this shortcoming is to reconcile the different function-
alities of distinct agent classes in a single agent model. Thus, it should combine re-
activity with goal-directed behavior and facilities for reasoning about other agents
and for interacting with them. The so called hybrid agents (see Wooldridge & Jen-
nings 1995 for an overview) have been developed to fulfill this requirement. Their
architecture is organized in two or more interacting layers. Each of them provides
the agent with various capabilities: In the INTERRAP architecture, e.g., (see Müller
1996) there are layers for reactive, intentional, and social properties, and a control
framework links these layers together to achieve an overall coherent behavior of the
agent.
In our mind the hybrid approach could serve as a foundation for specifying a
broad spectrum of different agent types. This includes several types of hybrid agents
as well as of pure agents with a variety of interaction modes. Thus, a tool should
support a modeler with methods and techniques allowing for refinement, associa-
tion and aggregation. It should follow the object-oriented approach and extend it to
agent based systems (Kinny & Georgeff 1997). This seems to be a promising way
to exhaust the modeler’s flexibility designing complex social systems.

4.2 Representing Agents’ Environments as Agents


Now we consider how the environment of agents should be represented. In this con-
text, we think of both conceptual and technical representation at the same time, be-
ing well aware that there might be a trade-off between both of them. We first point
out some environments in our test case, then try to get rid of some ambiguities con-
cerning the term “environment”, and at least motivate our point of view that agent’s
environments should be represented as agents, too.
The attentive reader will have noticed that every entity in our test case comes
as a kind of agent. But which of them should be seen as (part of) an environment?
Obviously, the town in which branches and their business clients are located acts like
an environment for them, and the federal reserve bank functions as an environment
for the banks (see section 3, and notice that it only depends on the model’s focus,
which agents count as “main agents” and which form their environment). On the
other hand, there is good reason to call, for example, all other competing banks of a
particular bank part of its environment.
We lay particular emphasis on the distinction between the former and the latter
use of the term “environment”: A so called common environment is an environment
that all agents in the model, or at least all agents at a given level of aggregation, have
in common. It is fixed over time, i.e. its behavior cannot be influenced by agents and
no agent must leave it. Thus, common environments define constraints for their in-
habitants. Further, we can assume that common environments tend to be much sim-
pler than the agents embedded. On the contrary, the specific environment of a partic-
ular agent consists of those comparably complex agents that interact with it. Which
agents belong to a specific environment may change over time, and their behavior
can be influenced more or less by the agent in focus.
The specific environment of an agent plus the common environment shape the
overall environment of that agent. Specific environments are best represented by
agents because, by definition, they consist of agents. Nevertheless the modeler may
either represent a specific environment by a number of individual agents, or by one
aggregate agent (the group, e.g.) who acts on (each) individual agent as a whole.
But how should we represent common environments? First of all, we have to
be aware of the fact that there are many of common environments one can think
about: two-dimensional planes (bordered or closed), three-dimensional physical
spaces (continuous or discrete), geographical spaces (e.g. given by maps), differ-
ent kinds of network structures, islands with specific locations like drinkwater-holes,
beaches, mountains and so on. All those common environments can be represented

 implicitly in the agents (e.g. by storing coordinates of the agent’s location),


 explicitly as specific data structures in the model, or by modules or classes
in the simulation system’s library, like the grids in SWARM (Langton et al.
1995), or
 explicitly as agents.

As one might expect, current multi-agent models use different representations of


common environments. For example, Holland’s ECHO and the sugarscape model by
Epstein and Axtell, despite of their similarities, strongly differ in this respect. In both
models agents “live” on a rectangular 2-dimensional grid with renewable resources,
but in ECHO these resources have a lot in common with agents (e.g. the so called
defense tag, see Holland 1995, 105), whereas resources and agents in sugarscape are
very different things (Epstein & Axtell 1996, 19).
Before figuring out the advantages and disadvantages of the agent approach to
represent common environments, we have to convince ourselves that it is really pos-
sible to represent all the different kinds of common environments by means of agents.
Take a bordered euclidean plane as an extreme example. The standard approach to
represent this very simple common environment would be to provide the agents with
coordinates of their position, and to constrain their moving capabilities appropriately,
e.g. to prevent the agents from leaving the plane. Imagine now, the plane would be
an agent itself. Then moving around the plane would have to be done by commu-
nication between agents instead of “anonymous” manipulation of some agent’s at-
tributes. Moving agents could ask the plane agent for their actual position and they
would have to inform it about their intention to change position. The plane agent in
turn, would confirm the movement if it is allowed. It would be easy to enhance the
plane agent’s behavior, for example, to automatically generate messages to agents
that have collided on their way.
The very simplicity of this example elucidates two disadvantages of the agents
approach representing common environments. The first, merely a matter of taste,
says that the agent metaphor will be overstressed if it is used to describe things as
simple as euclidean planes or physical spaces. The second objection concerns the
simulation’s time efficiency: a linear worsening due to increasing communication
costs has to be assumed here.
On the other hand we can identify a lot of advantages. The first and most abstract
one could be called soundness of concepts. Since both, common and specific envi-
ronments are represented by the same mechanism we expect that they can handled
in a more unified manner. Besides, the principal interchangeability of the model’s
“main agents” and their common environment becomes more visible.
Avoiding the “representational gap” between agents and environment allows
smooth transitions between them. For example, a model with a common environ-
ment consisting of a static network in which agents travel can easily be enhanced to
one with a flexible network built up of agents of another type. Flexibility will also
be improved because the creation of new common environments or the adoption of
existing ones can much easier be performed than if they were represented by specific
modules, or even worse, if they were hard-coded into the agents. Likewise, reusing
agents within another common environment should be no big deal.
However, we hope for a simpler implementation of common environments by
exploiting all the mechanisms that have to be provided by implementing agents any-
way. Because of all these advantages, in the framework of the massif project, we
intend to represent common environments as agents on their own.

5 Some massif Ideas: A Look Ahead


The multiplicity of approaches to be integrated as well as the different requirements
of potential user groups in the social sciences discourage the development of a uni-
fied modeling language. Instead, only a unified modeling framework will be created
in our new massif project by structuring the modeling process in
modeling agents: which structure do agents have, in which way will they be able
to act / react, which capabilities will they need etc.?
constructing societies: in which way will agents interact, in which manner is inter-
action controlled for problem solving, which kinds of organization are needed
to facilitate coordination and/or cooperation etc.?
The core of the modeling tool massif will consist in four levels of realization,
stacked on top of each other and transferable between each other. They will support
the construction of simulation models on different levels of abstraction:

the programming level: programming simulation models in a general purpose


object-oriented programming language, using interfaces to existing agent def-
initions,
the module level: combining agent descriptions and interaction structures into sim-
ulation models,
the scheme level: defining simulation models for classes of models, based on
scheme models,
the application level: developing model class specific simulators and at the same
time making problem oriented user interfaces available.

Modeling with massif will consist in (see the example in subsection 3.2)
 identifying types of agents and of interactions in the target system and its en-
vironment,
 associating kinds of agents and modes of interactions to the identified types of
agents and of interactions,
 defining attributes,
 defining rules for action, communication, and decision within the framework
of compatible interaction modes.
massif should and will also allow for initializing models, i.e. specifying values
of constants, and for preparing the collection of data during typical experiments, as
well as the description of particular experiments with their specific parameterization,
data collection and analysis.

References
Abelson, Robert P., and A Bernstein. A computer simulation of community referen-
dum controversies. Public Opinion Quarterly, 27(1):93–122, 1963.
Abelson, Robert P. Simulation of social behavior. In G. Lindzey and Elliot Aron-
son, editors, Handbook of Social Psychology, volume 2, pages 274–356. Addison-
Wesley, Reading MA, 1968.
Abelson, Robert P., and J. Douglas Carroll. Computer simulation of individual belief
systems. American Behavioral Scientist, 8:24–30, 1965.
Balzer, Wolfgang, C. Ulises Moulines, and Joseph D. Sneed. An Architectonic for
Science. The Structuralist Program, volume 186 of Synthese Library. Reidel,
Dordrecht, 1987.
Chattoe, Edmund, and Nigel Gilbert. A simulation of budgetary decision-making
based on interview data. Paper presented at SimSoc’95 (Boca Raton, FL)., August
1995.
Doran, Jim, and G. Nigel Gilbert. Simulating societies: an introduction. In Jim Do-
ran and G. Nigel Gilbert, editors, Simulating Societies: the Computer Simulation
of Social Phenomena, pages 1–18. University of London College Press, London,
1994.
Dignum, Frank, and Bernd van Linder. Modelling social agents: Communication
as action. In Jörg P. Müller, Michael J. Wooldridge, and Nicholas R. Jennings,
editors, Intelligent Agents III, volume 1193 of LNAI, pages 205–218. Springer,
Berlin, 1997.
Epstein, Joshua M., and Robert Axtell. Growing Artificial Societies – Social Science
from the Bottom Up. MIT Press, Cambridge, MA, 1996.
Finin, T., and R. Fritzon. KQML — a language and protocol for knowledge and
information exchange. In Proceedings of the 13th International Distributed Arti-
ficial Intelligence Workshop, pages 127–136, Seattle, WA, 1994.
Forrester, Jay W. Principles of Systems. MIT Press, Cambridge, Mass., London,
1968, 2nd preliminary edition 1980.
Haken, Hermann. Synergetics. An Introduction. Nonequilibrium Phase Transitions
and Self-Organization in Physics, Chemistry and Biology. Springer Series in Syn-
ergetics, vol. 1. Springer, Berlin, Heidelberg, New York, 2nd enlarged edition,
1978.
Haken, Hermann. Information and Self-Organization. A Macroscopic Approach to
Complex Systems. Springer Series in Synergetics, vol. 40. Springer, Berlin, Hei-
delberg, New York, 1988.
Henize, John. Critical issues in evaluating socio-economic models. In Tuncer I.
Ören, Bernard P. Zeigler, and Maurice S. Elzas, editors, Simulation and Model-
Based Methodologies: An Integrative View, NATO Advanced Science Insti-
tutes Series, Series F: Computer and Systems Science, vol. 10, pages 557–590.
Springer, Berlin, Heidelberg, New York, Tokyo, 1984.
Holland, John H. Hidden Order – How Adaption Builds Complexity. Addison-
Wesley, Reading, MA, 1995.
Kinny, David, and Michael Georgeff. Modelling and design of multi-agent systems.
In Jörg P. Müller, Michael J. Wooldridge, and Nicholas R. Jennings, editors, Intel-
ligent Agents III, volume 1193 of LNAI, pages 1–20, Berlin, 1997. Springer.
Klösgen, Willy. Software implementation of microanalytic simulation models —
state of the art and outlook. In Guy H. Orcutt, Joachim Merz, and Hermann
Quinke, editors, Microanalytic simulation models to support social and financial
policy, Information Research and Resource Reports, vol. 7, pages 475–491. North
Holland, Amsterdam, New York, Oxford, 1986.
Langton, Chris, Nelson Minar, and Roger Burkhart. The swarm simulation system. a
tool for studying complex systems. http://www.santafe.edu/projects/swarm, April
1995.
Masterman, Margaret. The nature of paradigm. In Imre Lakatos and Alan Musgrave,
editors, Criticism and the Growth of Knowledge. Cambridge University Press,
London, 1970.
Moulin, Bernhard, and Brahim Chaib-Draa. An overview of distributed artificial in-
telligence. In G.M.P. O’Hare and Nick R. Jennings, editors, Foundations of Dis-
tributed Artificial Intelligence, pages 3–55. Wiley, New York etc., 1996.
Müller, Jörg P. The Design of Intelligent Agents: a Layered Approach, volume 1177
of LNAI. Springer, Berlin, 1996.
Orcutt, Guy H., Joachim Merz, and Hermann Quinke, editors. Microanalytic sim-
ulation models to support social and financial policy. Information Research and
Resource Reports, vol. 7. North Holland, Amsterdam, New York, Oxford, 1986.
Orcutt, Guy H. Views on microanalytic simulation modeling. In Guy H. Orcutt,
Joachim Merz, and Hermann Quinke, editors, Microanalytic simulation models to
support social and financial policy, Information Research and Resource Reports,
vol. 7, pages 9–26. North Holland, Amsterdam, New York, Oxford, 1986.
Rumbaugh, James, Michael Blaha, William Premerlani, Frederick Eddy, and
William Lorensen. Object-Oriented Modeling and design. prentice Hall, Engle-
wood Cliffs, 1991.
Sola Pool, Ithiel de, and Robert Abelson. The simulmatics project. In Harold Guet-
zkow, editor, Simulation in Social Science: Readings, pages 70–81. Prentice Hall,
Englewood Cliffs, 1962. originally in: Public Opinion Quarterly 25, 1961, 167-
183.
Werner, Eric. Logical foundations of distributed artificial intelligence. In G.M.P.
O’Hare and Nick R. Jennings, editors, Foundations of Distributed Artificial Intel-
ligence, pages 57–117. Wiley, New York etc., 1996.
Wooldridge, Michael J., and Nicolas R. Jennings. Agent theories, architectures, and
languages: A survey. In Intelligent Agents: ECAI-94 Workshop on Agent The-
ories, Architectures, and Languages, Amsterdam, The Netherlands, August 8–9,
1994, volume 890 of LNAI, pages 1–39. Springer, Berlin, 1995.

Das könnte Ihnen auch gefallen