Autonomous Agents in Multiagent Organizations

Marc Premm and Stefan Kirn

University of Hohenheim, Department of Information Systems 2, Stuttgart, Germany

Keywords: Multiagent Organizations, Agent Autonomy, Organizational Modelling.

Abstract: Autonomous agents are constantly gaining relevance in economic applications. Autonomy as a characteris-

tic of agents enables flexible behavior in cases of unforeseeable conditions. Multiagent systems research has

analyzed various dimensions of autonomous behavior. However, the application of agents in an organiza-

tional context requires the actors to apply externally given rules that restrict agent autonomy. While multia-

gent systems aim at maximum flexibility, economical applications in organizations require stable structures.

Multiagent organizations in terms of structured and stable multiagent systems are necessary to successfully

link autonomous agents with organizations. Modelling autonomous agents in multiagent organizations re-

quires to include the organizational structure and the operational processes, but also needs to consider the

constitutive processes that enable the creation, adaption and dissolution of multiagent organizations. We

survey extant literature from distributed artificial intelligence and management science and propose models

for organizational structure and procedure of multiagent organizations. The models address new aspects for

including autonomous agents in organizations that result from the linkage between both perspectives.

1 INTRODUCTION

Agent autonomy is a widely discussed field of re-

search that continuously gained relevance in the last

decade (Vernon, 2014). Especially the emergence of

autonomous cars has contributed to spread the term

“autonomous” both in scientific literature and in

everyday life (Gerla et al., 2014). However, the def-

inition of autonomy remains unclear and, thus, cre-

ates difficulties to integrate autonomous agents in

organizations. In the economically orientated con-

text of organizations, a trade-off between flexibility

and stability is necessary to ensure processes that are

mainly stable but may be adapted due to modified

circumstances (Brenner, 2003).

The problem of connecting autonomously work-

ing machines with organizational concepts has al-

ready been addressed by Grochla (1966). Already in

1966, he analyzed whether machines might get intel-

ligent enough that the complexity of their work may

be classified on a similar level as the work of hu-

mans. Since then, this thesis has been controversially

discussed in management science. Technical sys-

tems continuously gain a higher level of autonomy

and multiagent literature assumes agents to be au-

tonomous. Autonomy, which mainly results from

learning capabilities, enables agents to consider

unpredictable environmental effects. Built-in incor-

rect or incomplete knowledge might be compensated

by learning capabilities, which contributes to agents’

decisions that are mainly independent from the work

of the developer (Russel and Norvig, 2009).

Multiagent systems research assumes autono-

mous agents to cooperate when necessary and there-

fore participate in multiagent systems (MAS). Thus,

MAS are created to follow a single task and dimin-

ish as soon as the task is solved, creating maximum

flexibility. However, the integration of agents and

MAS in organizations requires reliable structures.

We use the term multiagent (MA)-organization to

describe a collaboration of autonomous agents that is

Figure 1: Autonomy and participation in multiagent

organizations.

Premm M. and Kirn S.

Autonomous Agents in Multiagent Organizations.

DOI: 10.5220/0006094901210128

In Proceedings of the 9th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2017), pages 121-128

ISBN: 978-989-758-219-6

121

structured similar to organizations, creating stability

and means for economical applications.

Figure 1 illustrates the interconnections between

the initial influence of the developer, the learning

capabilities of the agent and its autonomous decision

on participation in MA-organizations. However, the

figure reveals an additional aspect: As autonomous

agents may participate in multiple MA-

organizations, goal conflicts between them may

occur and need to be resolved by the agent. Auton-

omy is also present in other domains and on another

level: Processes in hospitals are performed highly

autonomous by the involved departments. The cen-

tral hospital process management has only limited

influence on the process steps within the depart-

ments. The concept of fractal processes helps to

analyze and explain the interdependencies arising

with autonomous departments. We address this

problem by modelling the interdependencies be-

tween different organizational departments and frac-

tal business processes that involve autonomous

agents. This paper is based on previous work

(Premm and Kirn, 2015, Widmer et al., 2016) and its

aim is twofold: First, we survey literature from dis-

tributed artificial intelligence (DAI) and organiza-

tional theory to underline the importance of agent

autonomy in organizations. Second, we present

models for organizational structure and procedure to

analyze dependencies between them.

The remainder is organized as follows. In section

2, we discuss the state of the art. Section 3 presents a

modelling approach for MA-organizations. Section 4

discusses the linkage between organizational struc-

ture and procedure. Section 5 concludes.

2 STATE OF THE ART

This section surveys extant literature from DAI and

organizational literature, with a focus on agent au-

tonomy, basic problems of DAI and the fractal com-

pany approach introduced by Warnecke (1993).

2.1 Agent Autonomy

Autonomy has been attributed to software agents

widely in DAI literature. A definition, often quoted,

defines autonomous agents as agents that “have

control both over their internal state and over their

own behavior” (Jennings, 2000). However, this

definition lacks a specification, when an agents has

control over its internal state or its own behavior,

including its own reconfigurability (Dennis et al.,

2014): Software agents still have a developer who

has large influence on its behavior and, thus, also

over its internal state. Therefore, 4 levels of autono-

my can be distinguished in the context of DAI (Mül-

ler-Hengstenberg and Kirn, 2016): (i) Autonomy of

the developer: Besides the agents, the developer also

a certain kind of autonomy in developing the agent.

This includes choosing the software architecture and

programming language. (ii) Autonomy by design is

part of the product respectively agent definition and

insures that the agent is protected against external

influence. (iii) Technical autonomy addresses the

feasibility of intelligent and autonomous behavior

with respect to sensors and actuators as well as

available resources, e.g. energy, time, storage.

(iv) Autonomy of MAS describes the ability of MAS

to develop and maintain problem solving capabilities

partially independently from the participating soft-

ware agents. Here, we stick to the level of technical

autonomy, i.e. the technical feasibility of autono-

mous actions, and thus will first analyze the term

autonomy that is broadly used in literature.

In current technical development the term auton-

omous car is widely used to describe cars that are

capable of driving to a specified destination without

interaction of a human driver. But are these “auton-

omous” cars really autonomous? The goals of the

car are partly given by the developer, e.g. re-

strictions, rules, and partly given by human users,

e.g. destination, maybe even restrictions on the pos-

sible route. Thus, the car has no control over its own

goals as they are externally given and consequently

only partly over its internal state. So, we might ra-

ther term it a “fully automated” car as it is only au-

tomated in the sense of following directions. How-

ever, we might also think of an autonomous taxi that

may act on behalf of a goal function, e.g. earn mon-

ey, and may decide on its own with whom it is will-

ing to make a contract. This taxi would be classified

as “more autonomous” than a purely automated car.

Autonomy in general includes the ability to tem-

porarily give control away, e.g. an autonomous or

fully automated car may decide to join a group of

other cars to optimize travelling characteristics:

Safety distance may be reduced to decrease air re-

sistance and space occupied on the road. These par-

ticipation decisions are directly connected to organi-

zational structuring as each participant brings along

new competences and responsibilities for new tasks

(Widmer et al., 2016). Although, these organization-

al structures do not include humans, maximum flex-

ibility – as provided by MAS – would not be expe-

dient: Other cars of the group need to rely on fore-

seeable behavior of each participant. These MA-

organizations, however, inherently limit the autono-

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

122

my of the participating agents as organizational rules

limit their scope of action (Schillo and Fischer,

2004). Besides these self-controlled restrictions by

participation decisions, agents also face continuous-

ly restrictions by authorities: Not only do autono-

mous or automated cars need to follow traffic rules,

but these restrictions may be even increased in the

case of a police car or ambulance in action passing

their way. Consequently, autonomy is never a char-

acteristic that can be ultimately applied. Instead,

autonomy is in general restricted in some way and is

always a relation concept (Castelfranchi, 1995).

Agent autonomy is inevitably linked to the re-

sponsibility for actions. As an autonomous agent

needs the freedom to act autonomously, the question

arises whether the agent is able to be responsible for

its actions. However, we have to distinguish between

two types of responsibilities: (i) the implementation

responsibility for actions that lead to a successful

solution of the task and (ii) the juristic responsibility

that offers the possibility for contractual partners to

claim contractual performances (Widmer et al.,

2016). While autonomous agents may take the first

type of responsibility, they cannot be a contractual

partner and cannot be responsible in a legal sense.

MA-organizations primarily aim at structuring

the decisions of autonomous agents in a long-term

manner (Hübner et al., 2010). In contrast to MAS

that dissolve after the problem is solved, MA-

organizations have a long-term goal. Thus, the

members of a MA-organization have to restrict their

own freedom to act and give up some of their auton-

omy. On the other side the MA-organization needs

to address the still existing autonomy of the agents

by creating motivation to work towards the global

goal. However, the MA-organization is not able to

prevent agents from withdrawing from the organiza-

tion, in case their own goals run contrary to those of

the MA-organization. These problems are largely

analyzed in organizational theory (Pfeffer and Sa-

lancik, 1978), however, the involvement of autono-

mous software agents in organizations and the crea-

tion of MA-organizations is unattended in literature.

2.2 Basic Problems of DAI

Bond and Gasser define basic problems of DAI and

distinguish five basic questions that DAI needs to

address (Bond and Gasser, 1988). Agent autonomy

has major influence on these basic problem, but is

often neglected in literature. The influence of agent

autonomy on each of the five basic problems of DAI

is analyzed in the following paragraphs:

Problem Description and Solving. The first basic

problem of DAI is about “how to formulate, de-

scribe, decompose, and allocate problems and syn-

thesize results” (Bond and Gasser, 1988). However,

problem description is usually performed on a global

basis, neglecting the fact that agents act autono-

mously and might use their own models. Especially

in an organizational context, intelligent agents need

to align their own models, including problem de-

scriptions, with those of the MA-organization. On

the other side, due to single agents’ autonomy, Ma-

organizations have only restricted influence on in-

ternal models of their members.

Communication. In general, interaction and com-

munication between agents needs communication

languages or protocols. In a MA-organization,

communication languages and protocols are usually

defined by the MA-organization itself and may in-

tend strict rules. The purpose of strict rules are fore-

seeable behavior of the involved actors. However,

the autonomy of each agent may hinder communica-

tion by disregarding communication rules or the

sequence of a protocol, e.g. the initiation of an inter-

action. Analogously to other rules, motivation, e.g.

by incentives or sanctions, is necessary to ensure

agents following given rules.

Local Decisions. Autonomous agents are inevitably

associated with local decisions and the effects of

local decisions form the basis for decentralized con-

trol. In an MA-organization, decisions of each mem-

ber need to be aligned with the goals of the MA-

organization. However, the participating agents will

act on behalf of their own goals in the first place.

Thus, the regulations for their membership need to

include incentive mechanisms that ensure local deci-

sions in the MA-organization’s interest.

Coordination. Coordination in MAS is often char-

acterized and influenced by approaches from man-

agement science especially organizational theory.

Subsumed under the term coordination, this basic

question includes “how to enable individual agents

to represent and reason about the actions, plans,

and knowledge of other agents in order to coordi-

nate with them” (Bond and Gasser, 1988). This in-

cludes the information exchange of autonomous

agents, which is especially relevant in an organiza-

tional context and is directly connected to the local

decisions of each agent: Without the willingness to

share information the MA-organization might not

profit from additional knowledge generation, e.g. by

combining knowledge, or might suffer from incon-

sistencies in distributed knowledge bases.

Consistency. Consistency, e.g. of distributed

knowledge, in a group of agents depends on the

local decisions of single agents. Comparing MAS

Autonomous Agents in Multiagent Organizations

123

with agents in an organizational context, the prob-

lem of inconsistencies between the knowledge bases

of different agents even gains importance. As agents

in MA-organizations will work on a long-term basis,

inconsistencies might have severe impact on organi-

zational outcome. Influencing the coordination

mechanism and encouraging agents to resolve in-

consistencies in their knowledge base or conflicting

intensions is key to organizational performance.

The basic problems of DAI form a basis to dis-

cuss agent autonomy in an organizational context.

Addressing corresponding questions helps to under-

stand the impact of deploying agents in organiza-

tions and, hence, leads to an improvement of organi-

zational performance. While MAS evolve to solve a

problem and dissolve after a solution is found, MA-

organizations need to address the basic problems of

DAI from a long-term perspective.

2.3 Fractal Company

Enterprise modelling is constantly advancing by

including additional aspects of the real world. A

topic that has been addressed lately is the view on

enterprises as fractal companies (Witte, 2001; Sand-

kuhl and Kirikova, 2011). Ever-changing competi-

tion on globalized markets and the corresponding

complexity of decisions reveals new challenges for

the involved actors. Decentralization is one key for

enterprises to address these challenges and involves

decisions about organizational structures: Which

degree of autonomy is appropriate for which hierar-

chical level? Which resources and responsibilities do

the members need for decentralized decision (e.g.

Warnecke, 1993, Tapscott and Caston, 1993)? Or-

ganizations are forced to substitute strict hierarchical

structures with decentralized patterns of coordina-

tion. Conversely, organizational subunits that al-

ready show autonomous characteristics, increase

their degree of freedom and thus their level of au-

tonomy. To address the emerging challenges,

Warnecke (1993) introduces the term organizational

fractals that are characterized by the following four

major criteria: (i) Self-similarity are the structural

characteristics of organizations and modalities of

generating added value enabling resource sharing

between different organizational fractals. (ii) Self-

organization and self-optimization represents a de-

centralized approach addressing the strategic, the

tactical as well as the operational level of autono-

mous local decisions for solving tasks that have

previously unknown conditions. (iii) Goal-

orientation enables continuous measurement of each

organizational fractal’s performance, controlling

their autonomous behavior, e.g. by motivation. (iv)

Dynamic adaptation to unforeseeable changes of the

environment are enabled by autonomous behavior.

Organizational fractals in the sense of Warnecke

(1993) have a high degree of local autonomy, self-

control, and self-organization skills. The paradigm

has been transferred to multiagent literature under

the term holonic multiagent systems (Fischer et al.,

2003). Like autonomous agents in organizations,

organizational fractals aim to maximize their local

utility (e.g. in terms of workload or profit). Local

decisions as one basic problem of DAI is also rele-

vant in terms of organizational fractals: They auton-

omously decide whether to cooperate with other

organizational fractals. There are no means to force

organizational fractals to act in a specific way. Alt-

hough a direct influence is not possible, one way to

manage the behavior of organizational fractals or

groups of collaborating fractals is motivation. Ad-

dressing goal-orientation, the individual goal-

systems of organizational fractals needs to be

aligned to a globally consistent objective system by

incentives as well as sanctions (Warnecke, 1993). As

human members are usually only bounded rational,

establishing consistent goal hierarchies in organiza-

tions is nearly impossible (March and Simon, 1958).

Consequently, parallel existing goals within an or-

ganization are lacking consistency and, thus,

knowledge about internal objectives and dependen-

cies between them remains incomplete or even false.

In this context, consistency as one of the basic prob-

lem of DAI is directly affected by the autonomy of

the organizational members and their internal deci-

sion mechanisms. This inconstancies only address

organization internal goals. However, goal conflicts

also emerge on a higher level, as objectives between

the organization and its customers or other cooperat-

ing or competitive organizations may differ.

3 MODEL

Modelling MA-organizations has to involve soft-

ware agents as well as human agents and, thus,

needs to link the perspectives of information systems

with those of economics: First, constitutive process-

es are necessary to analyze the evolvement, adaption

and dissolution of MA-organizations. Second, the

organizational structure has to be mapped to the

model as organizations are usually divided into de-

partments with local decision competences, refer-

enced as organizational fractals. Third, organization-

al procedures involve multiple of these organiza-

tional fractals working together in multiple ways.

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

124

3.1 Constitutive Processes

The constitution of a MA-organization is necessary

to distinguish it from MAS that do not show long-

term stability. Therefore, the constitutional processes

help to scale up from sole non-binding interaction to

organizational structures with the corresponding

active processes (Cooren and Fairhurst, 2009).

Constitutive processes include the constitution,

adaption and dissolution of a MA-organization. The

processes that constitute organizations have already

been analyzed for their transferability to MA-

organizations: Kirn and Gasser (1998) provide a life-

cycle framework based on the Vienna Development

Method with three types of constitutive processes:

Constitution. The constitution of MA-organizations

requires agents willing to found and participate in a

new MA-organization. The constitution includes the

creation of organizational structures and processes

(see section 3.2 and 3.3). Hence, constituting a new

organization is dependent on the local decision of

several agents and is usually performed if all partici-

pants expect to have a positive return.

Adaption. While the initial constitution is designed

for a long-term period, changing environmental

circumstances and evolving goal definitions may

occur, but may also be initiated by a subset of the

organizational members. It includes the reorganiza-

tion of the organizational structure and procedures as

well as the admittance of new members or the exclu-

sion of existing ones.

Figure 2: Constitutive process of multiagent organizations.

Dissolution. If running the organization is not bene-

ficial for the remaining members anymore, dissolu-

tion processes are necessary. These processes have

to determine how the (non-)physical resources are

spread between the current set of members.

Constitutive processes aim at creating, maintain-

ing and if necessary dissolving the organizational

structure and procedure. Figure 2 shows the consti-

tutive processes for MA-organizations in an UML

activity diagram. The presented models are precon-

ditions for describing and applying the constitutive

processes creating and running a MA-organization.

3.2 Organizational Structure

As the agent’s autonomy forms the basis for its local

decisions, the organizational structure also influ-

ences the negotiations between the organization and

potential members. Term definitions for a common

language between both negotiators are necessary to

prevent inconsistency in alter stages. Therefore, we

adapt terms from literature of organizational theory.

The smallest autonomously acting organizational

unit is the position, which is in general constituted

independently of instantiations. The position and the

occupying agent are linked, while the unidirectional

cardinality is set to a maximum of 1. Thus, a posi-

tion can only be occupied by a single agent. Howev-

er, an agent may occupy multiple positions, possibly

in multiple organizations. A position is linked to one

or more tasks and thus to the corresponding compe-

tences and responsibilities: (i) A task is the target

performance linked to the position. Reaching the

target performance is one goal of the agent occupy-

ing the position. Handling tasks requires the agent in

charge to provide a set of capabilities and access to

resources. (ii) Competences denote the right to act in

a certain way in organizational literature (Hill et al.,

1994) and are the formal basis for position-specific

influence on the work of agents. (iii) The obligation

to act in a certain way (especially with respect to an

assigned task) is denoted responsibility. The respon-

sibility has an executive (performing an action) and

a legal dimension (bear the legal consequences).

Assigning a task to a position and consequently

to an agent implies expectations of other members

for task fulfillment. These expectations in MA-

organizations are called roles. Three types of roles

can be found in an organizational context (Hill et al.,

1994): (i) Task-specific roles specifying expectations

in relation to a task. (ii) Position-specific roles as

expectations related to the position that is occupied

by a member/agent. (iii) Individual-specific roles are

expectations that are founded in individual behavior

Autonomous Agents in Multiagent Organizations

125

or certain characteristics of a member, e.g. former

behavior of a member that is expected for future

actions. Figure 3 gives an overview on the main

UML classes for modelling hierarchical structures.

The link between a member of a MA-

organization and a position is the main representa-

tion of organizational membership. The position

itself is linked to several tasks, competences and the

corresponding responsibilities to use the competen-

cies to solve the tasks. The member provides a set of

capabilities and resources that are useful for the

MA-organization. “Usefulness” in the context of

organizational membership is represented by match-

ing capabilities used for solving specific tasks. Thus,

tasks are linked to capabilities and resources that the

agent provides and that are required for solving the

task. The class role is divided into three major sub-

classes representing expectations with respect to a

specific subject: (i) Individual-specific roles linked

to members, (ii) position-specific roles linked to

positions and (iii) task-specific roles linked to tasks.

Figure 3: UML class definition of organizational structure.

3.3 Organizational Procedure

Besides organizational fractals in the sense of auton-

omous departments, network-wide business process-

es consist of flexibly coordinated process fractals

autonomously making local decisions. These process

fractals may involve actors from multiple depart-

ments. Two dependent organizational problems

evolve in the context of process fractals: (i) the intra-

organizational structure of each process fractal and

(ii) the overall inter-fractal procedure that isn’t under

control of a single fractal. Process fractals, thus,

have to coordinate their actions for an appropriate

global output. Table 1 presents a meta-model for

modelling supply chains consisting of autonomous

process fractals. The logistics term transshipment

describes the actions connecting different logistics

process steps and, thus, allows the output of one

fractal to be used as an input for the next one.

While the participation of an agent in an organi-

zation is due to its autonomy dependent on its will-

ingness to participate, the presented meta-model

takes a look at higher level organizational proce-

dures. The process fractals within an organization

usually also have a certain degree of freedom in their

actions: In general, they cannot decide on the partic-

ipation at a certain process itself, but the fractals are

able to decide on their internal processes. The organ-

izational management has only limited influence on

the internal processes and thus has to motivate the

fractals to behave in a manner beneficial for the

organization. The meta-model visualizes this aspect,

which is especially important for agents in organiza-

tions, as they are usually able to handle considerably

more parallel tasks in different organizational pro-

cesses than their human co-workers.

Table 1: Meta-model of organizational procedure.

Label Symbol Description

Process

Fractal

Self-contained and self-organized

series of activities that involves

actors and is available via interfaces

Actor

Autonomous organizational entity

in a process fractal that has the

competency to make individual

decisions within a given scope

Interface

Coupling point of a process fractal

that allows for incoming or out-

going products/services from or to

another process fractal

Interac-

tion Path

Bidirectional communication link

between two actors of a fractal

Trans-

shipment

Transition of a product/service from

one process fractal to another one

4 LINKAGE OF STRUCTURE

AND PROCEDURE

The preceding sections presented two types of frac-

tals that occur in enterprises: (i) Organizational frac-

tals that may be embodied in departments or teams,

hence, hierarchical structures and (ii) process frac-

tals that have to solve specific logistic tasks as inter-

dependent steps of an overall enterprise output. Or-

ganizational structures and organizational proce-

dures are interconnected in various ways. In general,

there are numerous processes in enterprises that

involve multiple departments. Thus, one cannot

assume that the process fractals follow the same

pattern as the organizational structure. Consequent-

ly, organizational structure and procedure have to be

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

126

linked for a holistic view on the enterprise. Figure 4

shows the linkage of structure and procedure via an

identifier of the acting position. The example in-

volves a care robot that supports processes in a hos-

pital. Due to their ethical high responsibility for their

patients, physicians and their corresponding organi-

zational fractal are usually characterized by highly

autonomous decisions. Thus, hospital management

face the problem of efficient overall processes in-

volving multiple process fractals. As agents might

be included in more the one fractal, but lack legal

responsibilities and need mechanically readable

problem descriptions, the linkage between organiza-

tional structure and procedure needs to consider

these circumstances. However, the problems stated

in the context of hospitals may be also transferred to

other domains, e.g. autonomous production islands

in a manufacturing enterprise.

When autonomous agents are involved in organi-

zations, other factors have to be considered: Auton-

omous agents usually act in behalf some natural or

juristic person. Hence, somebody is responsible for

the agent’s action and has to stand in for contracts

that have been signed. In the example of autono-

mous cars, this leaves the question open who is re-

sponsible in case of an accident: The owner, the

manufacturer or the driver – who actually didn’t

drive. The same question arises in organizations:

Autonomous agents will usually work in some envi-

ronment, receiving orders from hierarchically super-

ordinate instances, e.g. a care robot in a hospital (see

Figure 4). However, the agent needs access to some

(hardware) resource and needs to be maintained.

These tasks are usually undertaken by the developer

or operator of the agent, e.g. the hospital internal IT

department or an external service provider. These

assignments have to be represented in the organiza-

tional structure, therefore autonomous agents in

organizations will be assigned to at least two organi-

zational departments: (i) IT department, developing

and maintaining the agent or manage service con-

tracts with the corresponding provider, and (ii) the

department, the agent is actually working for.

In the comparison with autonomous cars, the IT

department takes the role of an “owner”, when the

agents is developed by an external provider, as their

main task is to manage service contracts. The role

external service provider in this case is similar to a

manufacturer of autonomous cars. However, if an

autonomous agent is completely developed and

maintained within the organization, the IT depart-

ment would play a dual role as “owner” and “manu-

facturer”. In both cases, it is important to clarify who

is responsible for the agent’s activity, as the agent

itself isn’t able to take the juristic responsibilities.

Is the agent completely developed and main-

tained within the organization, internal organization-

al regulations may help to solve the problem of un-

clear responsibilities. However, if the agent has been

developed outside of the organization and autono-

mously decided to join an organization, the respon-

sibilities in a legal sense are unclear and strongly

depend on the involved countries. When country of

residence of the developer, the provider or that of

the application differ, it might be even unclear which

law is applicable.

Figure 4: Linkage-Example of organizational structure and

procedure.

5 CONCLUSIONS

In contrast to MAS that maximize flexibility, stable

structures for the interaction of autonomous agents

are necessary to create reliability in economical

applications We contribute to this problem by

providing models for three different perspectives on

MA-organizations: (i) the constitutive processes that

allow for creation, adaption and dissolution of or-

ganizations, (ii) the organizational structure that

represents a hierarchical order of capabilities and

resource requirements, and (iii) the organizational

procedures that represent fractal operational pro-

cesses. We discuss the problems arising with con-

necting organizational structure and procedure as

competences and responsibilities may remain un-

clear. The complexity increases significantly when

autonomous agents are assigned to multiple organi-

zational fractals and contribute to various fractals.

Autonomous Agents in Multiagent Organizations

127

Visualizing the emergent difficulties is a signifi-

cant step towards efficient integration of autono-

mous agents in organizations. However, further

research needs to develop new methods to address

these upcoming challenges. Working with the mod-

els presented enables their evaluation in an organiza-

tional context. Managing the integration of autono-

mous agents in economically orientated organiza-

tions is key to their application in general. Merely if

companies see the benefits of applying autonomous

agents in real world scenarios, their full potential

may be exploited. This particularly includes the

transition from solely flexible – and from an external

point of view unpredictable – behavior to stable

organizational processes.

ACKNOWLEDGEMENTS

This work has been supported by (1) the project

SmartSite (BMWi, FKZ 01MA13002) and (2) the

project InnOPlan (BMWi, FKZ 01MD15002), both

funded by the German Federal Ministry for Econom-

ic Affairs and Energy.

REFERENCES

Bond, A., Gasser, L., 1988. An Analysis of Problems and

Research in DAI. In: Bond, A., Gasser, L. (Eds.)

Readings in Distributed Artificial Intelligence.

Morgan Kaufman Publishers, San Mateo, pp. 3-35.

Brenner, M.J., Tushman, M.L., 2003. Exploitation, Explo-

ration, and Process Management: The Productivity

Dilemma Revisited. The Academy of Management

Review, vol. 28, no. 2, pp. 238-256.

Castelfranchi, C., 1995. Guarantees for Autonomy in

Cognitive Agent Architecture. In Proceedings of the

workshop on agent theories, architectures, and lan-

guages on Intelligent agents, pp. 56-70, Springer.

Cooren, F., Fairhurst, G.T., 2009. Dislocation and Stabili-

zation: How to Scale Up from Interactions to Organi-

zations. In: Putnam, L.L., Nicotera, A.M. (Eds.):

Building Theories of Organization – The Constitutive

Role of Communication. Routledge, New York.

Dennis, L.A., Fisher, M., Aitken, J.M., Veres, S.M., Gao,

Y., Shaukat, A., Burroughes, G., 2014. Reconfigurable

Autonomy. KI - Künstliche Intelligenz, vol. 28, issue

3, pp 199–207.

Fischer, K., Schillo, M., Siekmann, J., 2003. Holonic

Multiagent Systems: A Foundation for the Organisa-

tion of Multiagent Systems. In: Marik, V., McFarlane,

D., Valckenaers, P. (Eds.) Holonic and Multi-Agent

Systems for Manufacturing, LNAI 2744, pp. 71-80.

Hill, W., Fehlbaum, R., Ulrich, P., 1994. Organisa-

tionslehre. Paul Haupt. Bern, 5

edition.

Hübner, J.F., Boissier, O., Kitio, R., Ricci, A., 2010. In-

strumenting multi-agent organisations with organisa-

tional artifacts and agents – „Giving the organisa-

tional power back to the agents“. In Autonomous

Agents and Multi-Agent Systems 20, pp. 369-400.

Jennings, N.R., 2000. On agent-based software engineer-

ing. Artificial Intelligence 117, pp. 277-296.

Gerla, M., Lee, E.-K., Pau, G., Lee, U., 2014. Internet of

Vehicles: From Intelligent Grid to Autonomous Cars

and Vehicular Clouds. IEEE World Forum on Internet

of Things, pp.241 – 246.

Grochla, E., 1966. Automation und Organisation – Die

technische Entwicklung und ihre

betriebswirtschaftlich-organisatorischen

Konsequenzen. Gabler, Wiesbaden.

Kirn, S., Gasser, L., 1998. Organizational Approaches to

Coordination in Multi-Agent Systems. it+ti Intelligente

Agenten 4, pp. 23-29.

March, J.G., Simon, H., 1958. Organizations. John Wiley

and Sons, New York.

Müller-Hengstenberg, C.-D., Kirn, S., 2016. Rechtliche

Risiken autonomer und vernetzter Systeme – Eine

Herausforderung. De Gruyter, Oldenbourg.

Pfeffer, J., Salancik, G.R.,1978. The External Control of

Organizations - A Resource Dependence Perspective.

Harper & Row, New York.

Premm, M., Kirn, S., 2015. A Multiagent Systems Per-

spective on Industry 4.0 Supply Networks. In: Müller,

J.P., Ketter, W., Kaminka, G., Wagner, G., Bulling, N.

(Eds.) Multiagent System Technologies, LNCS 9433,

pp. 101-118.

Russel, S., Norvig, P., 2009. Artificial Intelligence – A

Modern Approach. Pearson, 3

edition.

Sandkuhl, K., Kirikova, M., 2011. Analysing Enterprise

Models from a Fractal Organisation Perspective – Po-

tentials and Limitations. In: Johannesson, P., Krogstie,

J., Opdahl, A.L. (Eds.) The Practice of Enterprise

Modeling, LNBIP 92, pp. 193-207.

Schillo, M., Fischer, K, 2004. A Taxonomy of Autonomy in

Multiagent Organisation. In: M. Nickles, M.

Rovatsos, and G. Weiss (Eds.): AUTONOMY 2003,

LNAI 2969, pp. 68–82.

Tapscott, D., Caston, A., 1993. Paradigm Shift: The New

Promise of Information Technology. cGraw-Hill, New

York.

Vernon, D., 2014. Artificial Cognitive System – A Primer.

The MIT Press, London.

Warnecke, H.J., 1993. Revolution der Unternehmenskultur

– Das Fraktale Unternehmen. Springer, Heidelberg.

Widmer, T., Premm, M., Kirn, S., 2016. A Formalization

of Multiagent Organizations in Business Information

Systems. In: Abramowicz, W., Alt, R., Franczyk, B.

(Eds.) Business Information Systems, LNBIP 255, pp.

265-276.

Witte, H., 2001. Logistik. Oldenbourg, München.

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

128