A Multi-level Model for Multi-agent based Simulation

Thomas Huraux

1,3

, Nicolas Sabouret

and Yvon Haradji

EDF R&D, Clamart, France

LIMSI-CNRS, University of Paris-Sud, Orsay, France

LIP6 - Pierre and Marie Curie University, Paris, France

Keywords:

Multi-level Modeling, Multi-agent Systems, Simulation.

Abstract:

In this paper, we consider the problem of modeling complex systems at several levels of abstraction. We design

SIMLAB, a multi-level model for multi-agent based simulation. Our approach is based on the coexistence of

different levels during simulation to enhance the model with complementary experts’ opinion. We present

how a same concept can be deﬁned independently of its granularity using the notion of modeling axis. We

consider recursive agents with interactions and inﬂuences which captures the inter-level dynamics. We also

propose observations to detect and to reify macroscopic entities.

1 INTRODUCTION

The simulation of complex phenomena is based on

the use of models that allow experts to understand

systems and to anticipate changes. These models are

highly constrained by the modeler’s view of the sys-

tem and by the computational limits, in terms of mem-

ory or computation time required to run simulations.

However, we know that considering a complex sys-

tem as a whole to reproduce the immense diversity of

its behavior is unfeasible (Batty and Torrens, 2001).

This is the reason why simulations are based on

partial representations. Rather than considering the

whole picture, people use simpliﬁcations and aban-

don some aspects that they consider less relevant for

a given study. The role of experts in the modeling

process is then to make choices about what should be

considered in the model.

Multi-agent systems (MAS) provide a well-suited

paradigm for simulating such complex phenomena.

One of the main strength of multi-agent simulation

is that it uses distinct entities to represent every con-

cept that has been identiﬁed as important by the mod-

eler. Moreover, the behavior of every agent, i.e. ev-

ery signiﬁcant entity to the modeler, can be recorded

and analyzed. As a consequence, and in contrast to

other approaches, the produced model does not take

the form of a “black box”: as Edmonds explains (Ed-

monds, 2001), MAS offer an ideal support to interact

with experts based on the agent’s behavior analysis.

Another advantage of multi-agent based simula-

tion is that it can use multi-level organisation. In-

deed, for a given complex system, expert knowledge

is available at different granularity levels, depending

on the domain, based on the idea that expert knowl-

edge is largely unusable if we just consider the mi-

croscopic level (Conte et al., 2001). For instance in

social science, studies go from psychology (individ-

uals) to sociology (groups) through ergonomics (ac-

tivity). In multi-level agent-based simulation, entities

that correspond to different abstractions’ level will be

represented by agents at different levels in the MAS

organization.

However, in most existing work, the analysis of

the emergent phenomena in multi-agent systems only

considers expert knowledge from one level at a time.

In this paper, we claim that some behaviors cannot be

understood without a multi-level context, i.e. a posi-

tioning in what forms the agent (entities at the lower

levels) and what constrains it (entities from upper lev-

els). Hence, our hypothesis is that representing and

co-simulating different levels in the same model may

produce agents with more complete behaviors.

The work presented here relies on two core ideas.

First, common characteristics of the same concept

shall be deﬁned independently from its level of ab-

straction. Second, levels should coexist during the

simulation not as different visualizations of a single

phenomenon, but to connect different aspects, differ-

ent expert knowledge. As we explain in the section

2, most of the existing models were not designed to

co-simulate complementary points of view. They are

139

Huraux T., Sabouret N. and Haradji Y..

A Multi-level Model for Multi-agent based Simulation.

DOI: 10.5220/0004814501390146

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 139-146

ISBN: 978-989-758-016-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

designed to perform abstractions or aggregations that

control the microscopic level. In section 3, we present

our multi-level model. Then, we illustrate in section

4 its properties with a simple example and in section

5 we discuss the perspectives of our works.

2 POSITIONING

Several frameworks and methodologies have been de-

veloped for engineering multi-agent systems. For

instance, the MaSE (Wood and DeLoach, 2001)

methodology proposes an iterative progression to

build a multi-agent system. The methodology is di-

vided into six phases but none of them is the question

of choosing levels of abstraction, even less a possi-

ble coexistence of these levels. Yet we believe that it

should play a central role in the design process of sys-

tems. In Gaia (Zambonelli et al., 2001), AGRE (Fer-

ber et al., 2005) or INGENIAS (Pav

on and G

omez-

Sanz, 2003), the use of organizations can be consid-

ered as an addition in the system of other levels of

abstraction. In these works, organizations are groups

of agents with tasks or/and goals in common. Agents

are then subordinated to organizations in which they

belong and this organization strongly relies on the

agents’ goals. In our work, we propose a multi-level

model for simulation with intentional agents without

explicit goals manipulation.

Multi-level models in the literature are often

domain-dependent. For instance the RIVAGE model

in (Servat et al., 1998) which consider water as a

multi-level set of agents for modeling ﬂow, erosion

and inﬁltration on heterogeneous soils. Another ex-

ample : the model in (Tranouez, 2005) simulate liq-

uid ﬂow where authors use vortex and macroscopic-

vortex models. These speciﬁc models cannot be

reused in a different context. But what is important

with these models is that they are not designed to

lower computational cost but to facilitate studies us-

ing several levels. This idea is at the heart of our work.

Actually, it is not one of our objectives in this current

paper to build a generic model.

In contrast, other works use multi-level modeling

to speed up simulations. For instance, the level-of-

detail approach used by (Navarro et al., 2011) to real-

ize large scale urban simulation. The system selects

by itself the level of representation for each agent al-

lowing a signiﬁcant computational gain. In such sys-

tems, only one level is activated at the same time

and so several levels do not coexist. We notice the

same problem with the hierarchical multi-level model

SWARM (Minar et al., 1996) and the SimulBogota

model (Gil-Quijano et al., 2007) based on agent ag-

gregation. In these models, macroscopic entities take

the lead of microscopic ones which loose their auton-

omy. Here, even though macroscopic levels are ex-

plicitly represented, levels do not really coexist.

In (Nguyen et al., 2011), authors choose a multi-

modeling approach using heterogeneous models to

represent several levels. Agents are aggregated and

constitute path using equations of ﬂuid ﬂow to sim-

ulate evacuation in case of tsunami alert of the viet-

namese city Nha Trang. This approach is studied

in (Siebert et al., 2010) : the meta-model AA4MM

use artifact to couple models of several levels. In

these models, inter-level interactions allow coupling

between models with functions of information shar-

ing. Although multi-modeling enable to join existing

models from different domain, it can be presumably

more interesting to get domain experts together to ex-

change looks on a same model.

The interested reader can obtain additional refer-

ences in the literature review on multi-level agent-

based modeling by G. Morvan (Morvan, 2012).

Based on the literature, we identify two limits in the

existing models :

1. Levels do not really coexist during the simulation

with often controlled microscopic levels.

2. Interactions and perceptions between levels are

not clearly elucidated.

For that reason we propose in this article a multi-level

model for multi-agent simulations (SIMLAB stands

for SIMLAB Is Multi-Level Agent Based) in which

various level agents can coexist and their interactions

reﬂect the properties of common notions between en-

tities, as understood by the modeler.

3 THE SIMLAB MODEL

Our multi-level agent-based model is based on the fol-

lowing fundamental concepts. First, some character-

istic elements of the modeled entities are shared by

agents from different levels. To this purpose, we in-

troduce the notion of modeling axis, which captures

the representation of transverse agents’ components.

Second, we propose to distinguish between agent in-

teractions (i.e. exchange of informations or requests

to the purpose of the system’s execution) and intra-

axis inﬂuence of properties, which captures the dy-

namics of these transverse components. Last, in order

to have a dynamic reorganization in the MAS, we pro-

pose to use observations that detect and reify macro

entities that make sense for the experts.

The following subsections present the general

agent model and these four components of our multi-

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

140

level MAS model.

3.1 Agent Model

All entities in the system are recursive agents (agents

composed of other agent with properties and inter-

actions transferred from one level to another). Let

Ω be the set of all agents. For each agent ω ∈ Ω,

we have Ω

sup

the set of direct super-agents of ω and

Ω

sub

the set of direct sub-agents. We denote @ the

non-transitive inter-level relation: ω

@ ω

means

that ω

∈ Ω

sub

and ω

∈ Ω

sup

. We assume that @

is acyclic, but no other restriction is speciﬁed in our

model. In particular, an agent can have several super-

agents.

Every agent ω ∈ Ω is characterized by: a set of

properties P (see 3.1.1), a set of actions A (see 3.1.2)

and interactions I (see 3.1.3), and a set of observa-

tions V (see 3.4).

3.1.1 Agent Properties

Agent properties are the variables that can be manipu-

lated by the agent. They capture the characteristic no-

tions of the modeled entity. A property can be atomic

(real, boolean, ...) or more complex (list, set, or even

another agent). We denote ω.p the property p of agent

ω. We denote P the set of all porperties and P (ω) the

set of properties of ω.

3.1.2 Agent Actions

Each agent has a set of internal actions to modify its

properties. To describe the effect of an action, we in-

troduce the operator “:=” to change state properties :

given a value ω.p := v means that v is allocated to the

property p of the agent ω. We denote A(ω) the set

of actions of ω. Each a ∈ A(ω) is a set of property

allocations. Moreover, we associate a precondition

Pre(a) (i.e. a function of P (ω) to B) to each action a

and, at every step of execution, an agent performs all

its actions whose preconditions are satisﬁed.

3.1.3 Interactions between Agents

Our interaction mechanism is a simpliﬁed version of

the classical intentional communication model pro-

posed by FIPA (FIPA consortium, 2003).

Every agent ω ∈ Ω is associated with a set of re-

actions R (ω), i.e. actions that can only be triggered

by interactions. Like actions, reactions are a set of

properties allocations and are associated with a pre-

condition (Pre(r)).

Moreover, every agent ω has a set of inter-

actions I (ω), where each i ∈ I (ω) is a tuple

htarget, reactionsi with target ∈ Ω the recipient and

reactions ⊆ R (target) a set of reactions by the recip-

ient. As for reactions and internal actions, we asso-

ciate a precondition Pre(i) to each interaction.

At every step of execution, for all i such that Pre(i)

is satisﬁed, for all r ∈ reactions(i) such that Pre(r) is

satisﬁed, all the effects of r are performed by target.

Example. Let n and h be two agents ∈ Ω corre-

sponding to an individual and a heater, h.temp ∈ P (h)

the temperature property of h and increase ∈ I (n) the

possible interaction of n on h deﬁne as :

increase = hh,{h.temp := h.temp ∗ 1.1}i

If the individual performs the increase interaction

on the heater, it will increase the temperature by 10%.

3.1.4 (De)Activation of Agents

Each agent can be activated or de-activated during

the simulation, either for selecting a suitable simu-

lation’s level for a given study, or for reducing the

computation load. We note ω.active the property that

states whether an agent is active or not, and we have

two set of actions f

and f

o f f

which are triggered

when the agent activates (resp. de-activates). When

de-activated, an agent continues to inﬂuence its peer

properties but it will no longer interact with other

agents (this interaction task is moved to the responsi-

bility of its super-agents) or perform internal actions.

3.2 Modeling Axis

A modeling axis is intended to represent consistency

between super and sub-agents. It consists of a set

of entities that correspond to the same concept. In

other words, it regroups all the abstraction levels of

a same aspect of the phenomenon being studied. For

instance, individuals, families and social groups are

elements of the population axis.

In our model, we have the to following properties:

• agents in the same axis are connected with the re-

lation @,

• agents in the same axis share a set of common

properties.

Notation. We denote χ

⊆ Ω the modeling axis cor-

responding to concept c (e.g. χ

population

3.2.1 Agents Relation in an Axis

Every axis χ

forms a connex subset of Ω for the rela-

tion @: all agents ω

∈ χ

are connected to the others

∈ χ

via @.

AMulti-levelModelforMulti-agentbasedSimulation

141

We denote lvl(ω) the agent’s level in its axis,

with lvl(ω) = 0 for lower-level agents, and lvl(ω) =

max

m∈Ω

sub

(lvl(m)) for higher-level agents. The mod-

eler can activate or de-activate agents of a given level,

for the purpose of its study or for performance issues.

(i) ⊆ χ

represents the set of all agents with lvl = i.

3.2.2 Shared Properties in an Axis

It is important that different modeling levels formed

by the relationship “@” mean something to the mod-

eler. Indeed, there is an inﬁnite number of levels for a

same phenomenon, and they are not all relevant in the

context of a given study. What links these different

levels within a modeling axis, conceptually, is sharing

some properties. Consider the example of human ac-

tivity, it can be modeled as a sequence of tasks, or at a

higher level as a set of habits. With these two models

of the same phenomenon, there are common proper-

ties like conditions, preferential periods, frequency of

realization, etc.

Every axis χ

is characterised by a subset of prop-

erties P

⊂ P such that:

∀ω ∈ χ

,∀p ∈ P

, p ∈ P (ω) (1)

We call P

the set of shared properties of the axis χ

3.3 Intra-axis Inﬂuences

To describe inter-level relations in a modeling axis,

we introduce the notion of inﬂuence. It is intended

to represent how agent properties are inﬂuenced by

its super or sub-agents. In MAS, we ﬁnd the no-

tion of inﬂuence in (Morvan et al., 2011) where the

inﬂuence is the “ desire” of an agent to modify its

environment. The semantic is different here, we are

closer to the deﬁnition of social inﬂuence and the in-

ﬂuence of individuals on the group as it is found, for

example, in work on social computing (Wang et al.,

2007). Through this mechanism, agents of different

levels will be able to co-evolve during the simulation.

In the model, we deﬁne F as the set of inﬂu-

ences. An inﬂuence f ∈ F is characterized by a tu-

ple hω

,ω

src

, p,in f li. It allows an agent ω

to in-

ﬂuence the value of a property of one of its super or

sub-agents called ω

, based on some of its properties.

src

⊆ P (ω

) is the set of properties of ω

used to

compute the inﬂuence on ω

, p ∈ P (ω

) the changed

property and in f l the inﬂuence function which change

the property. We denote f (P

src

) the value that ω

must

integrate in p (with the relation :=).

Example. Let n be a sub-agent of f . They corre-

spond to an individual and its family. t ∈ [0,1] is an

individual’s property which represents its tendency to

increase the heater. prio ∈ {co,sp} describes if the

family gives priority to its comfort or its spending.

We deﬁne ∈ F the following inﬂuence:

h f , n, prio,t,



t ∗ 0.5 prio = sp

t prio = co

It means that the individual’s tendency to increase the

heater is halved when its family give priority to spend-

ing.

When an agent is de-activated, it continues to have

an inﬂuence, unlike the actions and interactions that

the agent can not achieve anymore. The idea is that

the agent’s behavior is always potentially inﬂuenced

by the presence in the system of its super and sub-

agents, regardless of their activation status.

Note that inﬂuences apply on properties, not agent

instances of properties. As a consequence, shared

properties have the same inﬂuences for all agents in

the axis.

Remark. Considering the inﬂuences as a directed

graph with P the set of nodes (corresponding to the

set of properties, potentially involved in inﬂuences),

we must constrain the model to prevent the presence

of cycle. Let (p

, p

,..., p

n−1

, p

) be a sequence of

nodes in which two consecutive nodes p

and p

i+1

are

connected by an inﬂuence. The system must verify

the following acyclic property :

@(p

, p

,..., p

n−1

, p

) | p

= p

(2)

3.3.1 Inﬂuences Integration

Each agent property will be inﬂuenced by other prop-

erties in the same modeling axis. The agent is respon-

sible for aggregating all these inﬂuences to determine

the value of its property, using a σ combination func-

tion (which can be a sum, a product or any combina-

tion chosen by the modeler) :

ω.p := σ(E(p)) (3)

with E(p) representing the set of all inﬂuences.

3.3.2 Recursive Properties

The concept of inﬂuence allows us to deﬁne recursive

properties. A recursive property is a property that is

deﬁned for all agents of an axis (n.b. recursive proper-

ties are also shared properties) and is inﬂuenced by all

the source property values in the lower-level agents.

The inﬂuence of a recursive property is always ori-

ented from the sub-agent to the super-agent.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

142

The introduction of shared properties and recur-

sive properties in the model, with the inﬂuence’s

mechanism, allows the modeler to establish consis-

tency in behavior between different levels (e.g. es-

tablishing a link between the preferences of an in-

dividual and those of a group he belongs). On one

side the shared properties constitute the structural el-

ements of a modeling axis. On the other, the inﬂuence

and recursive properties link levels together and form

a common dynamic.

We denote P

rec

⊆ P

the sub-set of recursive

properties of the χ

axis. For each shared property

p between an agent a and its sub-agents Ω

sub

, there is

a corresponding inﬂuence. Formally :

∀p ∈ P

rec

∀i > 0 ∀a ∈ χ

(i) ∀ω

∈ Ω

sub

hω

,a, P

src

, p,in f li ∈ F (4)

3.3.3 Inﬂuence of a Recursive Property

Changing the value of a recursive property translates

into a recursive function whose computation is from

the agent to the sub-agents. Indeed, an agent ω ∈ Ω

recursively computes the inﬂuence f (see 3.3) for

each of its properties, such as :

f (ω) =



ω.σ( f (sub

),..., f (sub

)) Ω

sub

0 Ω

sub

(5)

with sub

∈ Ω

sub

, σ the combination function and 0

the neutral element for the inﬂuenced property.

3.4 Self-observation and

Transformations

We consider that to be effective, our approach to

multi-level modeling requires a dynamic structure.

Agents should be able to reorganize themselves in re-

sponse to the entrance or exit of agents, to allow dy-

namic formation of super-agents, or with an eye to

reify emergent phenomena. To this end, each agent

can be associated with one or more observations rep-

resented as a tuple hS ,M,φ,ti with S ⊆ Ω the mea-

sured agents. Each measurement is then compared

(using the following function : M : S → R) with an ac-

tivation threshold φ ∈ R to potentially trigger a trans-

formation t ∈ T (see 3.4.1).

3.4.1 Transformation of the System

A transformation modiﬁes the organizational struc-

ture of the system. In the multi-level representation

domain, we ﬁnd this type of operation with holonic

systems based on Koestler’s work (Koestler, 1967),

for instance the CRIO model in (Gaud, 2007). Let T

be the set of transformations. A transformation t ∈ T

is a couple hcond,∆

i with cond a condition on the

MAS structure for the completion of t and ∆

a set

of modiﬁcations on the organizational structure of the

system.

When an agent triggers a transformation, it is al-

ways directed to a single target agent. Transforma-

tions are deﬁned independently of agents: for a given

transformation’s type, there is potentially |Ω|

trans-

formations. Agents have ﬁve types of transforma-

tions:

(1) Create. The target agent is added by to the set Ω.

(2) Join. The target agent is added to the set of super-

agents of the triggering agent.

(3) Merge. The triggering agent and the target agent

group together to form a new super-agent. This

transformation is composed of a create transfor-

mation following by two join.

Note that two agents cannot “merge” if they al-

ready have a common super-agent with the same

properties: common super-agents at a given level

must have different properties. However, two in-

dividuals could form both a family and a group.

(4) Leave. The target agent is removed from the set

of super-agents of the triggering agent.

(5) Delete. The target agent is removed from the set

of agents Ω.

3.4.2 Triggering a Transformation

When the measurement function M, applied to one of

the agents of S , exceeds the threshold φ, the observa-

tion triggers the associated transformation t ∈ T :

∃x ∈ 2

|M(x) ≥ φ → t(x)

4 ILLUSTRATIVE EXAMPLE

To illustrate our proposal, let us provide a simple ex-

ample of multi-level modeling : we consider the ther-

mal comfort of individuals and groups. Individuals in

a room has a thermal comfort which will encourages

them to adjust their heater according to the group.

Applying the proposed model we introduce two

modeling axis : the population and the environment.

The population is composed of individuals that can

form groups and environment is limited to heater lo-

cated in rooms. As shown in ﬁgure 1, each individual

(which inﬂuence each other within the group) will in-

teract with the heater to adjust the thermostat accord-

ing to its thermal comfort. Then, the heater inﬂuences

AMulti-levelModelforMulti-agentbasedSimulation

143

Figure 1: Overview diagram.

the room temperature. Finally the room interacts with

the individual to change the perceived temperature.

4.1 Population Axis

We denote χ

pop

this modeling axis. It contains all

populations on several levels. We denote com f ∈

{−2,−1, 0,1, 2} the recursive property representing

the level of thermal comfort. As shown in 3.3.2, it

means that this property is inﬂuenced by all the source

property values in the lower-level population agents.

4.1.1 Individual

At the microscopic level we have the individuals.

An individual is characterized by ct ∈ {−1,0,1} a

cold-tolerance level, pt ∈ N the perceived tempera-

ture, tend ∈ [0,1] the tendency to adjust the heating,

ψ ∈ [0, 1] a probability to change it and cr ∈ χ

env

the

current room. The cold-tolerance changes the level of

thermal comfort felt by individuals.

Action. An individual i ∈ χ

pop

updates continu-

ously its thermal comfort following :

i.com f := i.pt − (21 + i.ct)

For this example, we simply consider 21

◦

C as the

ideal temperature. This is an obviously caricatural

description of thermal comfort mechanism but our ex-

ample is intended to illustrate the possibilities of our

model.

An individual modiﬁes its tendency to adjust the

heating according to group’s thermal comfort.

Reaction. An individual modiﬁes its perceived tem-

perature accordingly to the room’s interaction (see

4.2.2).

Interactions. An individual can act upon heater h

with the following interactions:

• inc = hh,{h.th := h.th + 1}i

• dec = hh,{h.th := h.th − 1}i

The heater will inﬂuence the room’s temperature

which, in turn, impacts the individual’s thermal com-

fort. A constant discomfort threshold is used as a pre-

condition to trigger these interactions from the indi-

viduals.

Inﬂuence. Because thermal comfort is a recursive

property, individual i inﬂuences the thermal comfort

of the group g following :

hi,g,{i.com f }, g.com f ,i.com f i

Obervation. We consider a simple observation

which adds an individual i to a group when it enters

into a room : hGroups ⊂ χ

pop

,M,1, joini with

M : g ∈ Groups →



1 g.room = i.cr ∧ i /∈ Ω

sub

0 otherwise

4.1.2 Group

A group is an agent composed of several individuals

in the same room room.

Action. A group g ∈ χ

pop

updates continuously its

level of thermal comfort using the average function

on inﬂuences to its comfort property (see σ in 3.3).

Inﬂuence. A group g inﬂuences the probability of

its members to adjust the heating following :

∀i ∈ Ω

sub

hg,i,{g.com f }, i.ψ,i.tend ∗

|g.com f |

4.2 Environment Axis

Let χ

env

be the environment axis. It contains the

heaters and the rooms as seen as groups of heaters.

Rooms are deﬁned in a static manner, they do not

evolve during the simulation.

4.2.1 Heater

A heater is characterized by a thermostat th ∈ [0, 9].

Reaction. A heater modiﬁes its thermostat accord-

ingly to individuals’ interactions.

Inﬂuence. A heater h inﬂuences the temperature of

the room r with :

hh,r,h.th, r.t,r.t + (h.th ∗ α)i

with α the heater’s efﬁciency.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

144

4.2.2 Room

We deﬁne a room as a super-agent of heater. It is char-

acterized by a current temperature t.

Action. A room r ∈ χ

env

updates continuously its

temperature using an average function as σ on all the

inﬂuences coming from heaters :

r.t = f (sub

),..., f (sub

) ∀sub

∈ Ω

sub

We consider that rooms’ temperature is only de-

pendent on heaters’ thermostat.

Interaction. A room r change the perceived tem-

perature by individuals :

updtT =

hr,{∀i ∈ Ω

group

sub

| group.cr = r},{i.pt := r.t}i

4.3 Discussion

This example shows how much the notion of model-

ing axis is implicitly present when we designing mod-

els. As it often happens with MABS, we have sev-

eral concepts in this example. First, individuals and

groups are entities which represent actors in the sys-

tem. Second, heaters and rooms that represent the en-

vironment with our model. It is very staightforward to

deﬁne the population and the environment axis corre-

sponding to these concepts. As a result, we tend to

obtain a multi-level representation which do not dis-

tort the modeled system. Moreover, shared properties

facilitate the modeling task by encouraging the reﬂec-

tion around what is common to all levels in a model-

ing axis. Once this work is completed, there is no

need to deﬁne these properties for each agent. Recur-

sive properties extend this simpliﬁcation of modeling

to the dynamic’s design. This is illustrated in the ex-

ample : the thermal comfort is computed on the in-

dividual level and it is transmitted to the group. One

can see how practical it could be for complex system

modeling.

Interactions are at the heart of agent based model-

ing. However, it is not sufﬁcient when entities of sev-

eral levels are introduced into the model. This is why

we introduce inﬂuences in addition to interactions. In

fact, the difference is purely conceptual : On the one

hand, interactions contribute to the system’s dynamics

(in this example, leading to a change of thermostat or

thermal comfort). On the other hand, inﬂuences con-

tribute to the coherence between levels. For instance,

the comfort level of the group is always connected to

the individuals’ comfort. In fact, it would be possible

to represent inﬂuences using interaction (e.g. an indi-

vidual interact with the group to change its comfort)

but the main idea here is to separate them in order to

simplify the modeling process. Thus, modeling strat-

egy can then be summarized in four steps :

1. What are the concepts involved to deﬁne the cor-

responding modeling axis and then deﬁne levels

for each of them ?

2. In which ways entities interact ?

3. How the behavior of agents inﬂuences their super-

agent, and vice versa ?

4. What organizational elements should be dynamic

and what could be the observations and the asso-

ciated transformations ?

The balance between inﬂuences (i.e. non-autonomous

variable modiﬁcation) and interactions (i.e. au-

tonomous exchange of information) need to be fur-

ther studied, to decide how the system designer can

have agent that are aware of other levels and more or

less sensitive to inﬂuences. This is one of our current

research objectives.

The role of observations as part of the organiza-

tional process, was not fully illustrated in our exam-

ple. In order to better understand this feature, we

must consider a more complete and complex exam-

ple. Consider now that individuals can perform tasks

(i.e. elements of everyday life activity : sleeping, eat-

ing, watching TV, ...), modifying their current room

and their thermal comfort (some task such as clean-

ing leads to increasing the amount of body heat pro-

duction and therefore changes the perceived tempera-

ture). Hence, third concept is introduced in our model

leading to create a new modeling axis : the activity

axis. Such an axis allows us to represent human activ-

ity on several levels of abstraction (e.g. tasks, habits,

lifestyles, ...). With activities, we only need to provide

observations to individuals such that they automati-

cally form groups when they perform the same activ-

ity pattern. Groups can be interesting for the mod-

eler in many studies. For instance in a study on en-

ergy consumption, groups are highly relevant as they

can help domain experts to identify some typologies

of consumers and their related energetic behavior. It

must be emphasized that it is important that agents are

able to perceive themselves and modify their organi-

zations. Indeed, modeling is facilitated because some

levels can be created automatically. Moreover, an ex-

plicit representation of reiﬁed entities gives them vis-

ibility in addition to a real role in the simulation due

to possible inﬂuences.

AMulti-levelModelforMulti-agentbasedSimulation

145

5 CONCLUSIONS

In this paper we propose SIMLAB, a novel model

for multi-level agent based simulation. This model

has the particularity to explicitly deﬁne the notion of

modeling axis allowing representation on others as-

pects than only actors. In addition, inﬂuences are

designed to represent inter-level relationships and in-

ﬂuences of recursive properties are spread from one

level to another. Moreover, observations make the

model capable of modifying the system’s organiza-

tion, as detecting and reifying macro-entities. As we

have explained, it would be interesting to represent

multi-level entities within a same model allowing ex-

perts of different domain to work together at the level

that suits them best.

In the SMACH platform (Amouroux et al., 2013),

we are currently working on the application of this

multi-level model to simulate human behavior. The

SMACH simulation platform is intended to study

household activities and their relation with electrical

consumption depending on speciﬁc pricing policies or

appliance use. We are extending the SMACH model

to study population, activity and environment at sev-

eral levels of representation. Our goal is to evaluate

possible incentives to diminish peak hours electric-

ity demand. We may be able to evaluate our model

in a three axis representation of human activity. For

instance, such model may well be able to reproduce

interesting social phenomena without having to sim-

ulate thousands of agents introducing explicit social

entities.

REFERENCES

Amouroux, E., Huraux, T., Semp

e, F., Sabouret, N., and

Haradji, Y. (2013). Simulating human activities to

investigate household energy consumption. Proc. of

ICAART, 13.

Batty, M. and Torrens, P. M. (2001). Modeling complexity:

the limits to prediction. Cybergeo: European Journal

of Geography.

Conte, R., Edmonds, B., Moss, S., and Sawyer, R. K.

(2001). Sociology and social theory in agent based

social simulation: A symposium. Computational &

Mathematical Organization Theory, 7(3):183–205.

Edmonds, B. (2001). The use of models-making mabs more

informative. In Multi-agent-based simulation, pages

15–32. Springer.

Ferber, J., Michel, F., and B

aez, J. (2005). Agre: Integrating

environments with organizations. In Environments for

multi-agent systems, pages 48–56. Springer.

FIPA consortium (2003). FIPA Communicative Act Library

Speciﬁcation and FIPA ACL Message Structure Spec-

iﬁcation. Technical report.

Gaud, N. (2007). Syst

emes Multi-Agents Holoniques : De

l’Analyse

a l’Implantation. PhD thesis, Universit

e de

Technologie de Belfort-Montb

eliard.

Gil-Quijano, J., Piron, M., and Drogoul, A. (2007). Mecha-

nisms of automated formation and evolution of social-

groups : A multi-agent system to model the intra-

urban mobilities of Bogota city. Idea Group Inc.

Koestler, A. (1967). The Ghost in the Machine. Hutchinson

& Co.

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

(1996). The swarm simulation system : a toolkit for

building multi-agent simulations. GEMAS Studies in

Social Analysis, Working Paper 96-06-042.

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

bibliography. Technical report, LGI2A, Univ. Artois,

France.

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

Irm4mls: the inﬂuence reaction model for multi-

level simulation. In Multi-Agent-Based Simulation XI,

pages 16–27. Springer.

Navarro, L., Flacher, F., and Corruble, V. (2011). Dynamic

level of detail for large scale agent-based urban simu-

lations. Proc. of 10th Int. Conf. on Autonomous Agents

and Multiagent Systems (AAMAS 2011), pages 701–

708.

Nguyen, T. N. A., Zucker, J.-D., Du, N. H., Drogoul, A.,

and Vo, D.-A. (2011). An hybrid equation-based and

agent-based modeling of crowd evacuation on road

network. International Conference on Complex Sys-

tems.

Pav

on, J. and G

omez-Sanz, J. (2003). Agent oriented soft-

ware engineering with ingenias. In Multi-Agent Sys-

tems and Applications III, pages 394–403. Springer.

Servat, D., Perrier, E., Treuil, J.-P., and Drogoul, A. (1998).

When agents emerge from agents : Introducing multi-

scale viewpoints in multi-agent simulations. LNCS,

1534:183–198.

Siebert, J., Ciarletta, L., and Chevrier, V. (2010). Agents

and artefacts for multiple models co-evolution: build-

ing complex system simulation as a set of interacting

models. In Proceedings of the 9th International Con-

ference on Autonomous Agents and Multiagent Sys-

tems: volume 1-Volume 1, pages 509–516. Interna-

tional Foundation for Autonomous Agents and Mul-

tiagent Systems.

Tranouez, P. (2005). Contribution

a la mod

elisation et

a la

prise en compte informatique de niveaux de descrip-

tion multiples. PhD thesis, Universit

e du Havre.

Wang, F.-Y., Carley, K. M., Zeng, D., and Mao, W. (2007).

Social computing: From social informatics to social

intelligence. Intelligent Systems, IEEE, 22(2):79–83.

Wood, M. F. and DeLoach, S. A. (2001). An overview

of the multiagent systems engineering methodology.

In Agent-Oriented Software Engineering, pages 207–

221. Springer.

Zambonelli, F., Jennings, N. R., and Wooldridge, M.

(2001). Organisational abstractions for the analysis

and design of multi-agent systems. In Agent-Oriented

Software Engineering, pages 235–251. Springer.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

146