Organization as a Multi-level Design Pattern for Agent-based Simulation

of Complex Systems

Vianney Sicard

1 a

, Mathieu Andraud

2 b

and S

ebastien Picault

1 c

INRAE, Oniris, BIOEPAR, 44300, Nantes, France

ANSES, Ploufragan-Plouzan

e-Niort Laboratory, Health and Welfare Research Unit, Ploufragan, France

Keywords:

Multi-level Agent-based Simulation, Design Patterns, Organizational System, Complex Systems.

Abstract:

This paper describes a generic design pattern to introduce organizational mechanisms into multi-level agent-

based simulation architectures, to help the modelling of highly structured complex systems. This pattern makes

it possible to specify how to couple any three levels of agents in a multi-level simulation architecture, through

their relationships to environments, taking into account organizational constraints. As a proof of concept, we

applied this pattern to the ﬁne-grained modelling of batch management in pig farms, and illustrate how the

pattern can be instantiated and composed at several agent levels to accurately handle a complex organization

in time and space. We thus demonstrate the beneﬁts of combining organizational concepts and multi-level

patterns to represent and simulate complex dynamic systems.

1 INTRODUCTION

Multi-level agent-based simulation (MLABS)

emerged during the last decade as a fruitful ap-

proach to model complex systems in a broad range

of ﬁelds (Morvan, 2012). This approach extends

Multi-Agent Systems (MAS) by providing an explicit

representation of the macroscopic level and of

each intermediary level, as agents endowed with

behaviours of their own. Several meta-models or

architectures have been proposed for the agentiﬁca-

tion of agent groups at multiple scales. However,

none of these models developed speciﬁc methods to

take explicitly into account organizational character-

istics that can be found in several complex systems

(e.g. in anthropized systems). In natural systems

indeed, organization is often studied as an emerging

phenomenon, which results from interactions of the

underlying agents and is not meant to be introduced

as such in the model. On the contrary, in human-

designed systems, organization is often an explicit

frame which impacts the behaviour and interactions

of individuals, and thus has to be modelled explicitly.

Conversely, organizational architectures are not

designed to cope with multi-level structures. In

https://orcid.org/0000-0002-4909-5544

https://orcid.org/0000-0003-2891-2901

https://orcid.org/0000-0001-9029-0555

multi-level approaches, the issue is to represent agent

groups corresponding to nested structural elements

(e.g. cells in a tissue, in an organ...) whereas in or-

ganizations, the issue is to represent agent groups

based on functional features or constraints (which

results in the concept of role). Besides, organi-

zational approaches often separate physical and so-

cial dimensions, whereas modelling complex ecosys-

tems require a strong coupling between those dimen-

sions (Bousquet and Le Page, 2004). Thus, multi-

level approaches and organizational architectures are

two complementary ways to address complex systems

modelling.

The purpose of this work is to propose a generic

design pattern to introduce organizational mecha-

nisms into a MLABS architecture. Essentially, an

agent group can be considered both a structural ag-

gregation of ﬁner-grained agents as in multi-level ap-

proaches, and a predeﬁned structured part of an or-

ganization. Agent states can thus either lead to a spe-

ciﬁc grouping (bottom-up multi-level aggregation), or

result from a speciﬁc grouping (top-down propaga-

tion of organizational constraints). As in MLABS

groups are agents encapsulating an environment, in-

troducing organizational features leads to specifying

the relationship between organizational and environ-

mental dynamics and constraints, without prior dis-

tinction between physical and social environments.

This paper is structured as follows: Section 2

232

Sicard, V., Andraud, M. and Picault, S.

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems.

DOI: 10.5220/0010223202320241

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 1, pages 232-241

ISBN: 978-989-758-484-8

analyses related works about MAS organization and

multi-level agent-based systems. Section 3 describes

the rationale for an organizational multi-level pat-

tern and the corresponding architecture. Section 4

presents an application of this pattern to modelling

a highly structured farming system (batch manage-

ment of a pig farm). Finally, we discuss methodolog-

ical and epidemiological modelling perspectives be-

fore concluding.

2 RELATED WORK

2.1 Multi-level Agent-based Systems

Multi-level agent-based simulation systems differ

from holonic systems (Fischer, 1999; Zhang and Nor-

rie, 1999) or recursive architectures such as SWARM

(Minar et al., 1996) by their ability to cope with

non-hierarchical structures, in order to represent non-

arborescent level couplings observed in complex sys-

tems. Apart from the huge number of ad-hoc MLABS

applications, a few meta-models propose an agentiﬁ-

cation of agent groups at multiple scales, between

atomic individuals and the whole system (Kubera

et al., 2011; Morvan et al., 2011; Drogoul et al.,

2013; Camus et al., 2015; Hjorth et al., 2020). Re-

cent developments also recommend design principles

based on patterns (Mathieu et al., 2018) as in other

ﬁelds of agent-based simulation (Juziuk, 2012; Kl

ugl

and Karlsson, 2009) to enhance the genericity and

reusability of conceptual solutions to recurrent issues.

In what follows, to exemplify the design pattern

we propose, we use the PADAWAN meta-model (Pi-

cault and Mathieu, 2011), which relies on a sim-

ple formalism and little speciﬁc assumptions, so that

the pattern can be transposed to other MLABS meta-

models. In this interaction-oriented model for multi-

scale simulation, the multi-level aspect is represented

by the ability for agents to encapsulate an environ-

ment. An agent is also located in one or more environ-

ments (without any prior distinction between physi-

cal or social environments), and this agent may itself

encapsulate another environment, and so on (Fig. 1).

The different levels between the whole system and the

ﬁnest-grained individuals are then represented by the

agents that, through the environment they encapsu-

late, host other agents.

Also, an interaction matrix is associated to each

environment to deﬁne possible interactions between

a source and a target agent family in this environ-

ment, as in the interaction-oriented simulation ap-

proach (Kubera et al., 2011). The key underlying as-

Agent

Environment

situated in

encapsulates

content

0..*

1..*

host

0..1

Figure 1: Typical multi-level architecture. Class dia-

gram showing the relationships between agents and envi-

ronments to manage multiple levels which can represent

non-hierarchical structures.

sumption here is that behaviours that agents can ex-

hibit can be speciﬁc to each level.

The main limitation of MLABS architectures in

general, as regards organizational features, is that they

focus on structural speciﬁcations for groups (as agents

hosting other agents through an environment) rather

than functional ones, thus ignoring notions such as

role. They provide powerful coupling mechanisms

between agents, environments and interactions, but

not much on e.g. constraints that could control which

agents and behaviours are allowed in each group.

Hence, MLABS architectures are well suited to model

highly structured systems where macroscopic agents

are built on top of microscopic agents, as an aggre-

gation, e.g. to simulate organisms or trafﬁc simu-

lations. The opposite design process, starting with

macro agents and trying to propagate constraints to-

wards the microscopic level, is still a challenge.

2.2 Organization in MAS

The common deﬁnition of organization is based on

three principles (Ferber et al., 2004): (1) The organi-

zational level describes the “what” and not the “how”;

(2) No agent description and therefore no mental is-

sues are provided at the organizational level; (3) An

organization provides a way for partitioning a sys-

tem, each partition (or group) constitutes a context

of interaction for agents. Organization can be under-

stood from two perspectives (Dignum et al., 2008):

as a process (i.e. a set of individuals with constraints

— structure, rules, models), or as an entity (with its

own requirements and objectives). An organization

provides a framework for structuring and managing

interactions between agents, and adjusting the level

of autonomy of the agents (H

ubner et al., 2009).

The AGR (“Agent-Group-Role”) meta-model

(Ferber and Gutknecht, 1998) deﬁnes organizations

as an additional abstraction level in the system, com-

posed of groups of agents having common goals or

tasks. It can be seen as a kind of dynamic frame-

work where agents are components. The environment

(assumed social) represents the context of communi-

cation between agents. Within a group, agents play

roles. A role describes the constraints (obligations,

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems

233

requirements, skills) that an agent must satisfy, the

beneﬁts (abilities, authorization, proﬁts) that an agent

receives when performing a role, and the associated

responsibilities. An agent can play several roles and

therefore be situated in several groups of the same

organization at the same time. But a group can be-

long only to one organization. AGRE (Ferber et al.,

2005) extends AGR by introducing a physical envi-

ronment in addition to social environments. The no-

tion of “Space” provides a generalization for physical

and social groups, both remaining strongly separated

(there can be only one physical environment but many

social ones). This approach provides a high level of

abstraction, and establishes the basis for the minimum

structure of an organization. However, the strict asso-

ciation between groups and organization and the dis-

tinction between the physical and social environments

are a limitation for representing systems with com-

plex structuring, because a same group cannot par-

ticipate in several organizations. Moreover, AGRE

does not propose a structured environment, restrict-

ing spaces to a context for a pattern of activities, and

is used for partitioning the system. This limitation

implies that the impact of the structure of the environ-

ment, and the dynamic of these structured environ-

ments are not taken into account.

In MOISE (Hannoun et al., 2000), the organiza-

tion is considered to be a system of rules compelling

the agent behaviour. These constraints correspond

to the role, i.e. the speciﬁcations of authorized be-

haviours of an agent in the organization through the

set of activities that the agent can perform. A group

is deﬁned by a set of roles and a sub-set of objectives

of these roles that can be achieved in the group. The

roles are quite similar of the notion in AGRE, but con-

straints are not directly linked to the consistency of

the organizational system. Agents have a set of con-

straints, called missions, they must take into account

for executing speciﬁc activities. Groups are a com-

position of agent with their roles and missions, and

are not explicitly linked to the notion of environment.

However, a group is a context of interaction between

agents, and can be understood as an environment (at

least social). Organization concept is formally di-

vided into organizational structure (OS) and organi-

zational entity (OE). OS is a graph deﬁned by a set of

roles which are nodes and links (acquaintance, com-

munication, authority) which are edges, and OE is

the implementation of the corresponding OS. MOISE

provides a deﬁnition of an organizational system and

a division of this system, but the notion of environ-

ment, structured and dynamic, is not taken into ac-

count.

Other approaches address more explicitly organi-

zation from the software engineering point of view,

such as ORA4MAS (H

ubner et al., 2009) and MA-

CODO (Weyns et al., 2010). MACODO, which

extends ORA4MAS, is a middleware for context-

driven dynamic agent organization and proposes an

abstraction of organization separating coordination

and structuring aspect from the local behaviour of

agents. Environments represent the software context

of communications, perceptions and actions between

agents. This approach provides several answers to

the issue of the dynamic relationship between agents,

organization level and environments, but no explicit

concepts to describe the dynamics of organizations.

Current organizational approaches do not take into

account multi-level aspects, such as the dynamics

of environments, the dynamic structuring of groups,

and the couplings between these different elements.

Modelling a highly structured system requires an ex-

plicit representation of interactions between atomic

elements and their environments, but also between

agents from each level. In these different approaches,

it is not possible to represent in a simple way the dy-

namics of the environments and the transversal cou-

pling between groups and organizations, particularly

in a multi-level context.

Therefore, we propose in this article a design pat-

tern to enable the ﬂexible introduction of organiza-

tional features into multi-level agent-based simula-

tions. The goal of this approach is to make it pos-

sible to represent and consider the speciﬁc coupling

between agents considered atomic, their environment

with its associated dynamics and the organizational

level, allowing that a group (agent that hosts several

agents from a sub-level) can have multiple purposes,

and can thus be located in several environments.

3 THE MULTI-LEVEL

ORGANIZATION PATTERN

The purpose of a design pattern is to provide a

generic, reusable and modular solution to solve a spe-

ciﬁc problem (Gamma et al., 1994), namely here in-

troducing organizational concepts into MLABS, al-

lowing organizations to have an explicit representa-

tion. This solution addresses a speciﬁc issue con-

cerning a subpart of a MLABS architecture and is de-

signed to be adaptable to other MLABS meta-models.

To allow an organization to have an explicit repre-

sentation in a MLABS, it has to meet three criteria: 1)

express and formalize the structural relationship be-

tween agents ; 2) constrain the behaviours and inter-

actions between agents with a notion of roles (at least

implicitly) ; 3) take part in the control of the environ-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

234

ment both structurally and functionally.

The pattern must also be compatible with classical

MLABS design, in particular as regards the grouping

of agents in the multi-level acceptance (Mathieu et al.,

2018). Thus, belonging to a group does not necessar-

ily imply being member of a speciﬁc organization, but

membership in an organization involves belonging to

a group.

As MLABS meta-models do not assume anything

speciﬁc on what the different levels represent, orga-

nizational features may thus concern potentially any

subset of agents in the multi-level system, depend-

ing on the application domain. Hence, transposing

organizational concepts into MLABS through a pat-

tern approach is highly relevant, since it only requires

to identify which agents are concerned by each orga-

nization, and specify the appropriate relationships be-

tween levels, i.e. how agents and environments relate,

with which constraints and dynamics.

3.1 Structure of the Multi-level

Organizational Pattern

The Organization Pattern describes the structural and

dynamical relations between three levels within the

MLABS: Organization, Groups and Atom (Fig. 2).

Org.

Group

Atom

Struct. env.

Atomic env.

Passport

member

composed

member through

situated in

encapsulated by

Figure 2: Structure of the multi-level organization pattern.

Agent Organization encapsulates a structured environment

which is composed of atomic environments, themselves en-

capsulated by Group agents. Agents involved as Atoms be-

long to an organization by their location in an atomic envi-

ronment, according to the constraints applied by the Orga-

nization. The Passport is a mediator between an atom and

its organization.

Organization is the agentiﬁcation of a structured en-

vironment, i.e. an environment that can be partitioned

into speciﬁc sub-environments that are called “atomic

environments”. The Atom is any agent which is mem-

ber of the organization. More precisely, organiza-

tion membership comes from the localization of the

Atom in an atomic environment of this organization,

according to the associated constraints. The atomic

environment contains information and has its own dy-

namics through the encapsulation by a Group agent.

Groups are located in the structured environment en-

capsulated by the Organization agent, which manages

the dynamics of the environments and its own con-

sistency through its constraints. As in (Ferber et al.,

2005), the organization is a frame within which the

agents behave. Information regarding Atom member-

ship in an Organization are stored in a Passport: e.g.,

location of the Atom, status regarding the constraints

of the organization, etc.

3.2 Atoms and Their States

An atom is any kind of agent that needs to be hosted

by an organization and is subject to the constraints of

that organization. The term “atom” means that, in the

pattern, we do not consider its underlying structure (it

can, or not, encapsulate an environment where other

agents can be located, etc.) because it is not relevant

as regards organizational features.

Atoms can interact with other agents according to

their location in environments. This location in an

environment implies the belonging to a group cor-

responding to the agent that encapsulates the corre-

sponding environment. The agent can act on the envi-

ronment, i.e. take or deposit information (§3.5).

As any agent, the atom is endowed with states,

which change according to the atom’s behaviour,

called “actual states” in the pattern. These actual

states may come to violate constraints of the organi-

zation which the atom belongs. The organization can

decide either that this violation is prohibitive (and ex-

clude the atom), or that the atom can nevertheless be

seen as temporarily complying with the organization’s

constraints (and let the atom stay in the organization).

In the latter case, the organization overrides this per-

ceptive discrepancy and assigns a “nominal state” to

the atom. In order to preserve the atom’s autonomy,

this nominal state cannot be imposed directly on the

atom, thus it is rather stored in a special data structure

called a “Passport” (§3.3).

For example, a boxer (atom) has a weight which is

an actual state (e.g. 187 pounds). According to the di-

vision where he plays (professional or amateur), seen

as an organization, he will receive a different nomi-

nal state (respectively cruiserweight or heavyweight).

This nominal state is obviously an interpretation of

the weight by either a professional or an amateur or-

ganization, not an intrinsic state of the boxer.

3.3 Organization, Constraints, Passport

The organization agent is the concretization of the

organization abstract level, endowed with its own

behavior and states. An organization encapsulates

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems

235

a structured environment which is partitioned into

sub-environments encapsulated by agents that are

groups, and where atoms are actually situated (Fig. 2).

The organization is in charge of its own integrity

and consistency control through constraints. Espe-

cially, the organization has to check that its members

(atoms) comply with the constraints, and if not, de-

cide whether to exclude the intruders or reconsider

how it perceives them by updating their passport.

The constraints allow the organization to admit

atoms and locate them, or to reject atoms if neces-

sary. They are a set of rules (r) regarding either actual

or nonimal states of atoms, that an atom must fulﬁl

to enter or remain in an organization, and correspond-

ing actions (a) that have to be performed when atoms

fulﬁl or violate the rules.

The passport is a mediator pattern between an or-

ganization and an atom. It stores information on the

location and nominal states, to represent the point of

view of the organization on the atom, at a given time.

The passport is owned by an atom but is only han-

dled by the organizational system (organization agent,

Fig. 2). The passport is composed of two items: the

history of successive locations of the atom over time,

and a visa which is the current status of the atom re-

garding an organization and its constraints, including

how actual states are interpreted into nominal states.

3.4 Roles

Roles are considered abstract behaviours that agents

can exhibit within a group (Ferber and Gutknecht,

1998). As in PADAWAN (Kubera et al., 2011), we

assume that behaviours that agents can perform in

a level are speciﬁed through an interaction matrix,

which is a function that assigns possible interactions

to a source family and a target family. Source/target

families, depending on the application context, can be

either a speciﬁc agent (e.g. “George Foreman”), or an

arbitrary name such as an agent class (“Boxer”) or an

actual or nominal state (“Heavyweight”). As a conse-

quence, roles in the organizational pattern are deﬁned

by the matching between behaviours speciﬁed in in-

teraction matrices and the combination of information

stored in atoms and passports.

3.5 Environment, Structured

Environment, Group

Environments are considered in accordance to (Math-

ieu et al., 2015), i.e. a space endowed with two func-

tions: placing agents and providing information, and

which can be physical or social indifferently. Further-

more, environments have their own dynamics linked

to their topology (neighbourhood, information trans-

mission, etc.).

A structured environment is composed of smaller

parts (atomic environments), with a speciﬁc topolog-

ical arrangement and a dynamic of its own. Its topo-

logical representation is a graph composed of vertices

that represent atomic environments carrying informa-

tion, and weighted edges that represent information

ﬂows, which constitutes the dynamics of the environ-

ment (Fig. 3).

Structured Environment (SE)

Atom

e1→atom

atom→e1

e1→e2

e2→e1

e1→e3

e3→e1

e1→SE

SE→e1

decay(I

)

generation(I

)

Figure 3: Representation of the dynamics of an informa-

tion I in an atomic environment e

, resulting from in/out

ﬂows due 1) to intrinsic dynamics (sources/sinks), 2) to ex-

changes with other atomic environments (e

, e

) or 3) with

the structured environment SE, and 4) to actions performed

by atoms located in e

. E.g., if I is a pheromone level, it

changes through deposition by atoms, diffusion to and from

neighbours respective to distance, and evaporation.

An environment is considered atomic according to a

given point of view of the pattern, i.e. depending on

the part of the system relevant as regards organiza-

tional concerns. Hence, the environment considered

atomic at a level used in the pattern (group or atom

level) can itself be actually structured to implement

the pattern recursively. Similarly, an agent consid-

ered atomic in an instance of the pattern for a given

triplet of levels, can be considered an organization in

another pattern instance, applied to another triplet of

agent levels (Fig. 4 & 5).

4 APPLICATION TO THE

MODELLING OF COMPLEX

FARMING SYSTEMS

Mechanistic models make it possible to understand

and predict the spread of pathogens at several scales

(from individuals to territories) under various scenar-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

236

subsystem 1

Atom

Group1

Org.1

Struct. env.1

Atomic env.1

subsystem 2

Atom

Group2

Org.2

Struct. env.2

Atomic env.2

Figure 4: Example of a group-focused pattern composition.

In this case, a Group level of the sub-system involved in

the ﬁrst pattern is itself structured as an Organization im-

plementing the pattern with other levels of the system. The

Atom level is the same in both patterns.

subsystem 1

Atom1

Group1

Org.1

Struct.env.1

Atomic env.1

subsystem 2

Atom2

Group2

Org.2

Struct. env.2

Atomic env.2

Figure 5: Example of an atom-focused pattern composition.

In this case, the Atom level of the ﬁrst pattern is itself han-

dled as an organization in the second pattern.

ios (control measures, climate...) (Keeling and Ro-

hani, 2008; Ezanno et al., 2020). Accounting for the

complexity of agricultural systems, with their strong

environmental and population structuring, dynamics

and constraints, and the coupling between these con-

cepts, is a real challenge to make models more realis-

tic and identify the mechanisms involved and possible

control levers.

MLABS appears a suitable solution for addressing

such complex pathosystems, as evidenced by EMUL-

SION

(Picault et al., 2019), an open source frame-

https://sourcesup.renater.fr/www/emulsion-public/

work dedicated to mechanic stochastic modelling in

epidemiology, already based on MLABS (Mathieu

et al., 2018). Thus, it was quite natural to implement

the multi-level organizational pattern on top of this

platform to address our own concerns in the ﬁeld of

complex farming systems.

4.1 Application to Batch Management

In what follows, we demonstrate how the multi-level

organizational pattern can be applied to model the

batch management in a French pig farm, and its added

value. Pigs are bred in batches, to guarantee a ho-

mogeneous evolution of the physiological states (i.e.

different steps in the animal like such as gestating

for sows or fattening for piglets). This involves that

batches must be consistent, i.e. that all animals are

in the same physiological state at the same time. The

animals, according to their type (sow or piglet), their

physiological state (depending on their age or repro-

ductive stage) and therefore their belonging to a batch,

are located in speciﬁc spaces (room, sector) corre-

sponding to environments. These environments have

a physical structuring, which directly impacts the en-

vironment dynamics and the relationship between the

environments and animals hosted (sharing informa-

tion like faeces, shedding of pathogens, airﬂow, etc.).

This type of management corresponds to a challeng-

ing realization for our issues:

• the physical environment is highly structured

(housing, litters, pens)

• the physical environment has its own dynamics

(pathogen spread, release, accumulation and de-

crease)

• the batches (social environments) must maintain

their consistency (homogeneity criteria)

• agents can have different statuses (actual: the real

value, and nominal: as considered for batch man-

agement) for a same state regarding the context

(e.g. depending on their belonging to a batch,

their physiological stage, their housing, etc.)

• environments and agents are closely coupled

(sharing information, etc.)

We consider a “typical” farm structure (Salines et al.,

2020) composed of ﬁve sectors corresponding to dif-

ferent physiological stages: 1) a mating sector where

sows are inseminated, 2) a gestating sector, 3) a ma-

ternity where sows and piglets are together for suck-

ling, 4) a post-weaning sector where piglets are sep-

arated from their mothers and 5) a fattening sec-

tor where pigs are fattened before being sent to the

slaughterhouse (Chambre d’Agriculture de Bretagne,

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems

237

Table 1: Typical pig herd batch management for 7 batches with a 21-day interval (Chambre d’Agriculture de Bretagne, 2010).

The duration in a sector for each batch is calculated to optimize room occupation, yet accounting for a time of cleaning and

disinfection.

Mating Gestating Maternity Post-weaning Fattening

sector sector sector sector sector

Physiological stage Insemination Gestating Suckling Post-weaning Fattening

Number of batches

2 4 2 3 6

to be housed

Entrance to the

sectors every ... (days)

35 or 42 77 or 84 35 or 42 56 or 63 119 or 126

Sector occupancy (days) 35 77 28

4 × 61 6 × 114

or 3 × 54 or 1 × 121

Duration of cleaning and 5 × 4

disinfection (days) or 2 × 1

7 and 1 14 and 7 2 5

2010). The number of batches determines the man-

agement (housing and timing). As an example, we

consider here a management in 7 batches with an in-

terval of 21 days. The evolution of a batch over time

depends on the duration spent in a sector, correspond-

ing to physiological stages (Table 1).

Before addressing pathogen spread and control,

our ﬁrst objective was to simulate this management

and to observe both the overall behavior of the system,

and the evolution of animals constituting batches, for

social (batches) and physical (housing) aspects. To do

so, we adapted the EMULSION framework to provide

the multi-level organizational pattern, and instantiated

the pattern with the appropriate levels of agents.

An animal corresponds to an atom, and we con-

sider two main organizations: one for batches (social)

and one for housing (physical). The “batch” organi-

zation is decomposed into several litters with one sow

per litter. Each litter is a (social) atomic environment

encapsulated by a group, and linked to the organiza-

tion by a structured environment which reﬂects the re-

lationships between atomic environments. The “hous-

ing” organization is decomposed into several sectors,

which are atomic spaces and also, recursively, orga-

nizations representing the decomposition of a sector

into rooms (Fig 6).

The evolution of the physiological state is time

dependent, speciﬁc duration for each step is directly

controlled by the atom and is, of course, dependent

of its type (sow or piglet). During the transition be-

tween gestating and suckling, sows produce offspring

(piglets).

The housing of atoms depends on their states

(physiological stage, type, batches) and on constraints

which deﬁne the coupling between the “batch” and

“housing” organizations:

• the location in a sector depends on the physiolog-

ical step of the animal

• all animals of a same batch are located in a same

room

• all animals of a same litter are located at the same

location

For example, to be located in the maternity sector, an-

imals can be either a sow or a piglet, and they have

to be in the suckling physiological step. Some an-

imals from different batches can potentially satisfy

these constraints and be in the same sector at the same

time, possibly with a different duration. This is why

the different batches are distributed in the different

rooms (one batch per room). The batch management

is optimized according to the occupation of the rooms,

accounting for a time for cleaning and disinfection

(Chambre d’Agriculture de Bretagne, 2010).

As a proof of concept of the organizational pat-

tern, we represent the occupancy of sectors and rooms

within 452 days for a 7-batch management. Due to

the 21-day time gap, we start with the ﬁrst batch in

suckling step, and initialize each batch considering

this state: the second batch is in gestating since 21

days, the third batch is in gestating since 42 days, etc.

The herd is composed of 15 sows per batch (i.e. 105

sows in total) with a fertility rate of 10, which can

increase the total population to over 1,000 animals.

Simulation outcomes (Fig. 7) show exactly the

main housing described in the literature (Chambre

d’Agriculture de Bretagne, 2010) and observed on

the ﬁeld. The multi-level organization pattern proved

convenient to represent, model and simulate the speci-

ﬁcities of the batch management farm system. The

organizational pattern makes it possible both to ob-

serve the spatial behavior (allocation) and to follow

the “social” organization aspect at different levels,

from batches to animals. Consequently, this ﬁne-

grained modelling capability allows precise aspects of

communication and interaction between agents to be

taken into account, whether they are atoms, organiza-

tions or encapsulated environments.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

238

subsystem Housing

Animal

Sector

Housing

Struct. housing

Atomic sector

subsystem Sector

Animal

Room

Sector

Struct. sector

Atomic room

subsystem Batch

Animal

Litter

Batch

Struct. batch

Atomic litter

Figure 6: Composition of three instances of the multi-level organization pattern for the modelling of the batch management

of a French pig farm. Two main organizations are deﬁned: Housing corresponding to a physical organization, and Batch

corresponding to a social organization. The organization of sectors within housing provides a ﬁner grain to control this level.

lllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllll

room 2

room 1

0 50 100 150 200 250 300 350 400 450 500

time

Mating room

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

room 4

room 3

room 2

room 1

0 50 100 150 200 250 300 350 400 450 500

time

Gestating room

lllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllll

lllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllll

room 2

room 1

0 50 100 150 200 250 300 350 400 450 500

time

Maternity

llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

room 3

room 2

room 1

0 50 100 150 200 250 300 350 400 450 500

time

Post−weaning room

llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

room 6

room 5

room 4

room 3

room 2

room 1

0 50 100 150 200 250 300 350 400 450 500

time

batch

batch1

batch2

batch3

batch4

batch5

batch6

batch7

Fattening room

Figure 7: Results of housing coming from the simulation of the batch management farm system. Animals are located in the

different sectors and rooms according to their states with regard to the organizations (Chambre d’Agriculture de Bretagne,

2010). It is thus possible to accurately track the system at different levels (batch, litter, animal, etc.)

4.2 Perspectives

We have demonstrated that is possible to accurately

represent and consider the complex organizations

such as highly structured farming systems in an ex-

isting MLABS architecture, based on a design pat-

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems

239

tern approach. Our next objective will be to study the

spread of pathogens within the system, taking into ac-

count interactions between individuals, environments

and organization levels. This new opportunity to ad-

dress assumptions that could not be easily taken into

account until now, should help to better understand

pathogen transmission in these complex systems, es-

pecially to evaluate various control scenarios in a re-

alistic and efﬁcient way.

Reducing health risk at the animal, farm and ter-

ritorial levels is essential to the viability of live-

stock farms and production sector. The Porcine Re-

productive and Respiratory Syndrome (PRRS) is a

viral contagious disease and extremely widespread

in areas with a dense pig population (Rose et al.,

2015). To understand and predict pathogen transmis-

sion and to compare the effectiveness over time of

realistic control strategies, mechanistic epidemiologi-

cal modelling is essential to complement the expertise

of health managers, quantifying epidemiological and

economic impacts. The possibility of considering the

complex structuring of herds and the farming system

provides new opportunities for controlling the spread

of the pathogen.

5 CONCLUSIONS

We have analysed the difﬁculties encountered in ex-

plicitly taking into account the organizational charac-

teristics that can be found in several kinds of complex

systems.

To overcome existing limitations, we propose

an organizational system for multi-level agent-based

simulation that take into account the representation

and implementation of the dynamic relationships be-

tween agents, organization levels and environments.

This proposal comes as a design pattern so that it can

be reused and adapted in other multi-level architec-

tures, and can be composed to cope with highly com-

plex structures (sub-organizations or concurrent or-

ganization). Besides, it fully relies on the fact that

all entities of a MLABS are represented by agents,

thus supporting the structural homogeneity of the sys-

tem. This pattern was easily implemented in an exist-

ing simulation platform based on existing multi-level

patterns (EMULSION framework), demonstrating the

ﬂexibility and consistency of our proposal with exist-

ing MLABS approaches.

The proof of concept implemented in EMUL-

SION also contributes to the multi-scale epidemio-

logical modelling of complex ecosystems. The repre-

sentation of the interaction between agents, environ-

ments and levels of organization is an important step

to represent highly structured populations and envi-

ronments to enhance classical modelling paradigms.

That will provide a deeper understanding of the dy-

namics of anthropized environments and help to as-

sess realistic control scenarios.

Finally, introducing organizational features in

MLABS contributes to reduce the gap between a

structural and a functional approach to complex sys-

tems. Using a design pattern approach to do so makes

it possible to extend the concept of organization as a

speciﬁc relationship between several levels reiﬁed by

agents, rather than a dedicated component added to

a MAS, hence leading to a more homogeneous and

ﬂexible architecture.

ACKNOWLEDGEMENTS

This work is supported by a grant from the Animal

Health division of INRAE (French national research

institute for agriculture, food and environment) and

the French region Pays de la Loire.

REFERENCES

Bousquet, F. and Le Page, C. (2004). Multi-agent simula-

tions and ecosystem management: a review. Ecologi-

cal Modelling, 176(3-4):313–332.

Camus, B., Bourjot, C., and Chevrier, V. (2015). Consid-

ering a Multi-Level Model as a Society of Interacting

Models: Application to a Collective Motion Example.

Journal of Artiﬁcial Societies and Social Simulation,

18(3):7.

Chambre d’Agriculture de Bretagne (2010). Les conduites

en bandes en production porcine - coh

erence de la

cha

ıne de b

atiments, Organisation du travail, Truies

en groupe.

Dignum, V., Meyer, J.-J. C., Weigand, H. G., Dignum, F.,

and Meyer, J.-J. C. (2008). An Organization-oriented

Model for Agent Societies.

Drogoul, A., Amouroux, E., Caillou, P., Gaudou, B., Grig-

nard, A., Marilleau, N., Taillandier, P., Vavasseur, M.,

Vo, D.-A., and Zucker, J.-D. (2013). GAMA: multi-

level and complex environment for agent-based mod-

els and simulations. In Gini, M. and others, edi-

tors, Proceeding of the International Conference on

Autonomous Agents and Multi-Agent Systems (AA-

MAS’2013), pages 1361–1362.

Ezanno, P., Andraud, M., Beaun

ee, G., Hoch, T., Krebs,

S., Rault, A., Touzeau, S., Vergu, E., and Widgren, S.

(2020). How mechanistic modelling supports decision

making for the control of enzootic infectious diseases.

Epidemics, 32:100398.

Ferber, J. and Gutknecht, O. (1998). A meta-model for the

analysis and design of organizations in multi-agent

systems. In Proceedings International Conference

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

240

on Multi Agent Systems (ICMAS’98), pages 128–135.

IEEE Comput. Soc.

Ferber, J., Gutknecht, O., and Michel, F. (2004). From

Agents to Organizations: An Organizational View of

Multi-agent Systems. In Goos, G., Hartmanis, J., van

Leeuwen, J., Giorgini, P., M

uller, J. P., and Odell,

J., editors, Agent-Oriented Software Engineering IV,

volume 2935, pages 214–230. Springer Berlin Heidel-

berg, Berlin, Heidelberg.

Ferber, J., Michel, F., and Baez, J. (2005). AGRE: Integrat-

ing Environments with Organizations. In Weyns, D.,

Van Dyke Parunak, H., and Michel, F., editors, Envi-

ronments for Multi-Agent Systems, Lecture Notes in

Computer Science, pages 48–56, Berlin, Heidelberg.

Springer.

Fischer, K. (1999). Robotics and Autonomous Systems.

Robotics and Autonomous Systems, page 11.

Gamma, E., Helm, R., Johnson, R., and Vlissides, J.

(1994). Design Patterns, Elements of Reusable

Object-Oriented Software. Addison Wesley.

Hannoun, M., Boissier, O., Sichman, J. S., and Sayettat, C.

(2000). MOISE: An Organizational Model for Multi-

agent Systems. In Monard, M. C. and Sichman, J. S.,

editors, Advances in Artiﬁcial Intelligence, Lecture

Notes in Computer Science, pages 156–165, Berlin,

Heidelberg. Springer.

Hjorth, A., Head, B., Brady, C., and Wilensky, U. (2020).

LevelSpace: A NetLogo Extension for Multi-Level

Agent-Based Modeling. Journal of Artiﬁcial Societies

and Social Simulation, 23(1):4.

ubner, J. F., Vercouter, L., and Boissier, O. (2009). In-

strumenting Multi-agent Organisations with Artifacts

to Support Reputation Processes. In H

ubner, J. F.,

Matson, E., Boissier, O., and Dignum, V., editors,

Coordination, Organizations, Institutions and Norms

in Agent Systems IV, Lecture Notes in Computer Sci-

ence, pages 96–110, Berlin, Heidelberg. Springer.

Juziuk, J. (2012). Design Patterns for Multi-Agent Systems.

Keeling, M. J. and Rohani, P. (2008). Modeling Infectious

Diseases in Humans and Animals. Princeton Univer-

sity Press.

ugl, F. and Karlsson, L. (2009). Towards Pattern-

Oriented Design of Agent-Based Simulation Models.

In Braubach, L., van der Hoek, W., Petta, P., and

Pokahr, A., editors, Multiagent System Technologies,

volume 5774, pages 41–53. Springer Berlin Heidel-

berg, Berlin, Heidelberg. Series Title: Lecture Notes

in Computer Science.

Kubera, Y., Mathieu, P., and Picault, S. (2011). IODA: An

interaction-oriented approach for Multi-Agent Based

Simulations. Journal of Autonomous Agents and

Multi-Agent Systems, 23(3):303–343.

Mathieu, P., Morvan, G., and Picault, S. (2018). Multi-

level agent-based simulations: Four design patterns.

Simulation Modelling Practice and Theory, 83:51–64.

Mathieu, P., Picault, S., and Secq, Y. (2015). Design

patterns for environments in multi-agent simulations.

In Chen, Q., Torroni, P., Villata, S., Hsu, J., and

Omicini, A., editors, Proceedings of the 18th Confer-

ence on Principles and Practice of Multi-Agent Sys-

tems (PRIMA 2015), volume 9387, pages 678–686.

Springer.

Minar, N., Burkhart, R., Langton, C. G., and Askenazi, M.

(1996). The Swarm Simulation System: A Toolkit for

Building Multi- Agent Simulations. page 12.

Morvan, G. (2012). Multi-level agent-based modeling -

A literature survey. arXiv:1205.0561 [cs]. arXiv:

1205.0561.

Morvan, G., Veremme, A., and Dupont, D. (2011).

IRM4MLS: The Inﬂuence Reaction Model for Multi-

Level Simulation. In Multi-Agent-Based Simulation

XI, volume 6532 of LNCS, pages 16–27. Springer.

Picault, S., Huang, Y.-L., Sicard, V., Arnoux, S., Beaun

ee,

G., and Ezanno, P. (2019). EMULSION: Transparent

and ﬂexible multiscale stochastic models in human,

animal and plant epidemiology. PLOS Computational

Biology, 15(9):e1007342.

Picault, S. and Mathieu, P. (2011). An Interaction-Oriented

Model for Multi-Scale Simulation. In Walsh, T., edi-

tor, Proceedings of the 22nd International Joint Con-

ference on Artiﬁcial Intelligence (IJCAI’2011), pages

332–337. AAAI.

Rose, N., Renson, P., Andraud, M., Paboeuf, F., Le Potier,

M., and Bourry, O. (2015). Porcine reproductive

and respiratory syndrome virus (PRRSv) modiﬁed-

live vaccine reduces virus transmission in experimen-

tal conditions. Vaccine, 33(21):2493–2499.

Salines, M., Andraud, M., Rose, N., and Widgren, S.

(2020). A between-herd data-driven stochastic model

to explore the spatio-temporal spread of hepatitis E

virus in the French pig production network. PLOS

ONE, 15(7):e0230257.

Weyns, D., Haesevoets, R., Helleboogh, A., Holvoet, T.,

and Wouter, J. (2010). The MACODO Middleware for

Context-Driven Dynamic Agent Organizations. ACM

Transactions on Autonomous and Adaptative Systems

(TAAS), 5(4):16.

Zhang, X. and Norrie, D. H. (1999). Holonic Control at

the Production and Controller Levels. In In Proceed-

ings of the 2nd International Workshop on Intelligent

Manufacturing Systems, pages 215–224.

Organization as a Multi-level Design Pattern for Agent-based Simulation of Complex Systems

241