EVALUATING THE COST OF SUPPORTING INTERACTION AND

SIMULATION THROUGH THE ENVIRONMENT

Flavien Balbo, Fabien Badeig

Universit

e Paris-Dauphine, LAMSADE, Place du Mar

echal de Lattre de Tassigny, 75016 Paris, France

Universit

e Paris Est, GRETIA, INRETS, 2 Rue de la Butte Verte, 93160 Noisy le Grand, France

Julien Saunier

Universit

e Paris Est, LEPSiS, UMR INRETS LCPC, 58 Bld Lef

ebvre, 75015 Paris, France

Keywords:

Environment, Evaluation, Communication, Activation.

Abstract:

The environment has emerged as a powerful ﬁrst-order abstraction in Multi-Agent Systems (MAS), as well as

a critical building block. One beneﬁt is to reduce the complexity of the agents by delegating to the environment

a part of the tasks of the system. This delegating process provides a ﬂexible way to exchange information and

to coordinate the agents thanks to the environment. The counterpart is a centralization of a part of the MAS

processes inside the environment.

In this paper, we present the modeling of an environment for multi-agent communication and simulation. Our

proposition enables the addition of advanced features to the MAS like multi-party communications (communi-

cation) and contextual activation (simulation). We evaluate the cost of this environment process and compare

it to the execution of the same tasks in the agents for communication and simulation.

1 INTRODUCTION

In (Weyns et al., 2007), the authors give the follow-

ing deﬁnition of a multi-agent system (MAS) environ-

ment: “The environment is a ﬁrst-class abstraction

that provides the surrounding conditions for agents

to exist and that mediates both the interaction among

agents and the access to resources”. The environment

is a critical building block in the multi-agent systems

that encapsulates its own responsibilities. The agents

interact in their environment and the effects of their

actions are observed and evaluated through this envi-

ronment. Thus, the environment provides observabil-

ity and accessibility services.

The relation between the environment and the

agents is traditionally based on a perception - decision

- action cycle (Weiss, 1999) which is repeated by the

agents during their life time. The three phases of the

cycle are executed in each agent. The perception be-

ing the agent ability to observe its environment, its re-

sult is the computation of what can be called contexts.

A context is deﬁned as a set of information that can be

used to characterize a situation. This set of informa-

tion includes percepts and messages that are obtained

by the agent and its own state. Thanks to the context

computation, the agent decides to execute the suitable

action. The relation between the environment and the

agents based on this perception - decision - action cy-

cle does not exploit the potential of the environment

that remains a static entity. This relation can be en-

hanced by delegating the perception process to the en-

vironment (Weyns et al., 2004; Badeig et al., 2007;

Saunier and Balbo, 2009; Kesaniemi et al., 2009).

The objective is to support advanced features that are

difﬁcult to obtain without the environment.

We have proposed an environment modelling and

the adequate platform that are based on the delega-

tion of the perception process to the environment. Our

proposition supports two advanced features: 1) Multi-

Party Communications (MPC) (Branigan, 2006) takes

into account dyadic interaction (one to one), group

interaction (one to many) and overhearing (many to

one/many) within the same interaction process; 2)

Contextual activation (Laberge, 1975) applied to the

simulation process consists in activating the agents di-

rectly according to their context.

126

Balbo F., Badeig F. and Saunier J..

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE ENVIRONMENT.

DOI: 10.5220/0003182001260135

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 126-135

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

In our proposition, despite the delegation process

to the environment, the agents keep their autonomy

because the decision process and the action process

remain in the agents. The perception phase of the

cycle is modiﬁed to take into account the relation

with the environment. That means that the environ-

ment computes the contexts for the agents, and that

the agents can modify this computation process of

the environment to suit their needs: the agents de-

cides which contexts are important for them and act

by modifying in consequence the computation pro-

cess that is executed by the environment.

In our proposition, in the MPC case, the environ-

ment mediates the communication: the context is the

information about the agent receiver, the message and

any other environment information which is relevant

to this communication. if the context is validated,

the environment addresses the messages to the related

agents. In the simulation case, the environment is

the scheduler that manages the agent activation: the

context is the subset of information accessible to the

agents. The environment activates an agent in a spe-

ciﬁc context and according to this context, the agent

performs the suitable action.

To adapt the perception process to each agent, the

environment needs ‘tools‘ to compute their contexts.

In our proposition, these tools are ﬁlters that reify

the relation between an agent (described in the envi-

ronment) and its contexts. A ﬁlter contains the con-

straints on the context related to an action of an agent.

This action is the reception of a message for com-

munication ﬁlters and a speciﬁc agent action to per-

form after validation for activation ﬁlters. The agents

modify their relation to the environment thanks to the

addition and removal of their ﬁlters in the environ-

ment. This process is dynamic and the environment

activates only its current ﬁlters. These choices (ad-

dition/removal) belong to the decision process that is

managed by each agent.

The externalization of the computation of the per-

ception process improves qualitatively the design of a

MAS. The dynamic ﬁlter management by the agents

gives more ﬂexibility to the MAS design. More de-

tails on qualitative improvement are given in (Saunier

and Balbo, 2009) for communication and in (Badeig

et al., 2007) for simulation. The counterpart is a cen-

tralization of a part of the MAS processes inside the

environment. In this paper, we evaluate the cost of

this centralization for computing agent contexts, in

comparison to the execution of equivalent solutions

for communication and simulation.

Section 2 motivates the delegating process of tasks

to the environment for communication and simulation

and introduces an illustrative example with different

examples issued of a crisis management application.

Section 3 presents the environment model. Section

4 provides a theoretical evaluation. In section 5, we

provide an empirical evaluation, and we conclude in

section 6.

2 MOTIVATIONS

2.1 Communication

Recent research on multi-party communications

(Branigan, 2006; Saunier and Balbo, 2009) shows

how multi-agent communications can take advantage

of the complexity of the human communication pro-

cess. The main issue in supporting MPC is to take

into account dyadic interaction (one to one), group

interaction (one to many) and overhearing (many to

one/many) within the same interaction process. The

sender does not know all the agents that might be in-

terested in its messages. For example, an agent can

listen to messages without the agreement/knowledge

of the sender through overhearing. For a recipient, the

usefulness of a message may depend on the context of

the sender, the context of the message, and the context

of the recipient itself.

These challenges are related to the way the re-

cipients are chosen. MPC requires knowledge of the

needs of both the sender and the recipients. From the

sender viewpoint (direct interaction), it is a connec-

tion problem: which agents are related to my mes-

sage? The problem is to map the senders needs (infor-

mation, capabilities, resources, ...) to the address of

related agents. From the recipient viewpoint (indirect

interaction), it is a data extraction problem: which

messages are related to me? The problem is to map

the recipients needs to the content of the messages.

For each message, these problems have to be simulta-

neously solved by the communication infrastructure.

The environment is able to solve these problems by

mediating the communication in order to ﬁnd all the

receivers of each message. Nevertheless, in the cog-

nitive agents community, few works explicitly present

the environment as an interaction support. For direct

interaction, the environment is often associated to an

infrastructure that supports point to point communi-

cation. For indirect interaction, cognitive agents use

speciﬁc services that are based on the management

of a shared collection of data (e.g. (Picco and Bus-

chini, 2002)) that may be understood as a part of the

environment. There is therefore a separation between

the solutions to realize direct and indirect interaction

although the environment provides a suitable frame-

work to unify them (Platon et al., 2005).

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE

ENVIRONMENT

127

If the computation of the context does not take into

account ambient conditions, it can classically be done

inside the agents. This solution implies that the agents

receive all messages and ﬁlter them. An evaluation of

the environment support consists in the cost compar-

ison of the context ﬁltering either in each agent after

the reception or in the environment during the trans-

mission process.

2.2 Agent-based Simulation

In a simulation, the scheduling policy deﬁnes the acti-

vation order of the agents. Once activated, the agents

behave according to their context. How the agents are

activated and what information is available to com-

pute the agent context depend on the agent-based sim-

ulation (ABS) framework.

In most ABS frameworks, a scheduler activates

the agents one after the other and each activated agent

computes its context to choose one action to execute

at each simulation step. The classical ABS frame-

works are designed to support this activation process.

For example, in the platforms CORMAS (Bousquet

et al., 1998) and MASON (Luke et al., 2005), the

scheduler activates a standard method for each agent.

This method is specialized by the designer to adapt

the agent behavior. In the Logo-based multi-agent

platforms such as TurtleKit (Gutknecht and Ferber,

2001) or the STARLOGO system

, an agent has an au-

tomaton that determines the next action to perform.

The choice of these platforms to be agent-

oriented, with a light environment support, implies

that the computation of the context is repetitive be-

cause it is computed in each agent at each time cy-

cle during the simulation execution. This computa-

tion is done in each agent even if the agent context

does not change between two time cycles and/or if

several agents share the same context during a time

cycle. Here, the evaluation consists in the cost com-

parison of the context ﬁltering either in each agent

after the activation or in the environment during the

scheduling process.

Example. In this paper, we consider a crisis man-

agement application where several emergency ser-

vices must be coordinated in order to reduce the crisis

effects. A crisis situation is a dynamic phenomenon

deﬁned by the initial situation, which depends on

place and time, and by the impact on population and

infrastructure. We focus our example on a speciﬁc

point which is the victim evacuation. This task con-

sists in coordinating two agents playing the role medi-

cal porter. The goal of the medical porters is to shift a

http://education.mit.edu/starlogo/

victim to an emergency vehicle. This action requires

one medical porter with the skill medical monitoring

and one medical porter with the skill victim handling.

The ﬁrst skill allows the medical porter to monitor the

victim and to inform the hospital of the evolution of

the victim health. The second skill is necessary to

handle correctly the victim.

Each simulation component (agents, messages

and objects) is situated on a grid, i.e. a two-

dimensional space. In this example, we consider that

the victims belong to the set of the objects because

they are not autonomous. The agents (medical porter)

act in this environment and cooperate in order to evac-

uate victims. A medical porter can move randomly

or towards a given direction, and it possesses only

one skill. It can either monitor or handle a victim,

but the two skills are required to evacuate the victim.

The agents can communicate in order to ﬁnd a partner

with the complementary skill. Each agent has a ﬁeld

of perception that limits its perception of the environ-

ment. A medical porter is able to perform one of the

following actions in a time cycle of simulation: 1) the

action move randomly, 2) the action move towards a

location, 3) the action wait, and 4) the action evacuate

a victim.

We use this application along the paper to illus-

trate our study of the cost of the environment support-

ing interaction and simulation.

3 ENVIRONMENT MODELING

To support multi-party communications and contex-

tual activation, we propose to use the environment as a

privileged intermediary to control interaction (EASI,

Environment as Active Support for Interaction) and

activation (EASS, Environment as Active Support for

Simulation). The EASS model (Badeig et al., 2007)

embeds the EASI model (Saunier and Balbo, 2009).

The environment manages meta-informations on the

MAS (agents, messages, context) and uses them to

compute the agent context(s). In this section, we

give the background elements to understand the as-

sessment of the cost of the environment. More details

about the models can be found in [?]

The environment model EASI is thus deﬁned by

hΩ,D, P,F i with:

• Ω = {A ∪ M S G ∪ O} = {ω

,...,ω

} the set of

entities (with A the set of agents, M SG the set of

messages and O the set of objects, i.e. all entities

that are not agents or messages),

• D = {d

,...,d

} the set of domain descriptions of

the properties,

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

128

• P = {p

,..., p

} the set of properties,

• F = { f

,..., f

} the set of ﬁlters.

Entity: An entity ω

∈ Ω is deﬁned by he

where e

is a reference to an element of the MAS

and e

is the description of that element. An ele-

ment of the MAS can be an agent, a message or an

object. A reference is its physical address on the plat-

form or other objects (such as URL, mailbox, etc.).

The description e

is a set of couples hp

i where

∈ P and v

is the value of the property for the en-

tity. Any agent of the MAS has its own processing

and knowledge settings. It is connected to the envi-

ronment thanks to its description that the environment

stores and updates. This description e

is used for the

routing of the informations to the agent with the ref-

erence e

(EASI) or its activation (EASS). The agents

can modify their description dynamically.

Property: A property p

∈ P : Ω →

∪ {unknown,null} is a function whose description

domain d

∈ D can be quantitative, qualitative or a

ﬁnite set of data. The unknown value is used when

the value of the property cannot be set, and null is

used to state that the property is undeﬁned in the

given description. In order to simplify the notation,

only the value of the description domain d

is given

to specify a property.

Filter: A ﬁlter identiﬁes the entities according to

their description (e

) and realizes the interaction be-

tween the concrete objects (e

). A ﬁlter f

∈ F is

a tuple f

= h f

,[ f

],[ f

],n

i with n

the ﬁlter name.

The assertion f

: A → {true, f alse} identiﬁes the re-

lated agents, the assertion f

: M SG → {true, f alse}

identiﬁes the related messages, and f

: P (Ω) →

{true, f alse} is an optional set of assertions identify-

ing other entities of the context. A ﬁlter is considered

valid for any tuple hagent,[message],[context]i such as

(agent)[∧ f

(message)][∧ f

(context)] is true.

The choice of the action specializes the ﬁlter. Ev-

ery agent a whose description validates f

is activated

if there exists a set of entities in the context such that

is true. After its activation, the agent validates the

action related to this context and executes it (or not).

It is a communication ﬁlter if the action is to add the

message m that satisﬁes f

in the message box of the

agent a. The agents can add and remove dynamically

the ﬁlters. Thus, they adapt their relation with the en-

vironment according to their needs.

In a crisis simulation, the agents have observable

properties: their identiﬁer, which is unique, their loca-

tion, their availability (true/false values) which spec-

iﬁes if an agent rescues a victim, their skill (posi-

tion/monitoring), their ﬁeld of vision, and their inter-

nal time. The messages have the following properties:

their identiﬁer that is unique, their sender identiﬁer,

the message type (’request’/’accept’), their location

which is the location of the sender when the message

is put in the environment, and the skill required by

the sender in the case of communication. The victims

have three properties: their identiﬁer, their location

and their status (stretcher/vehicle). The victims to be

evacuated must be diagnosed; this is expressed by the

property diagnosed (true/false values). The following

table sums up the properties of the MAS components

descriptions:

property Medical porter Message Victim

id x x x

location x x x

availability x

vision x

time x x

skill x x

sender

type x

status x

diagnosed x

In the simulation example, let f

be a communica-

tion ﬁlter related to the following context: an agent is

interested by a message if it is available ( f

), if the re-

quest message is close to the agent a and a has the re-

quested skill ( f

) and if the victim has been diagnosed

( f

). In this example, f

contains one entity which is

a victim identiﬁed by the property id

of the message.

The request messages have information about the vic-

tim location and the skill of the medical porter.

The processing of the context is traditionally done

by the agents, but in EASI and its extension EASS it

is done by the environment. In this way, the computa-

tion can be done with information that is shared by the

agents (the entities descriptions) and with a degree of

mutualization. However, it implies the centralization

of the computation and the management of the infor-

mation update. Now, we propose to evaluate the cost

of this centralization.

4 THEORETICAL ASSESSMENT

The theoretical assessment is the comparison between

the processing by the environment of a communi-

cation ﬁlter for n

agents and the processing by n

agents of all the messages. In order to take into ac-

count the context dynamics, this theoretical assess-

ment also studies the update process of the MAS.

The theoretical assessment is based on two criteria.

The ﬁrst, noted Cost

, is the number of tests per-

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE

ENVIRONMENT

129

formed during the ﬁltering process and the second,

noted Cost

, is the number of resulting messages.

The objective is to identify in which cases it is better

to mediate the communication through the environ-

ment than to manage it in the agents.

Following the deﬁnition given in section 3, a ﬁlter

f has tests related to one agent description ( f

), one

message description ( f

) and a subset of entities ( f

To simplify the explanation, we assume there is only

one entity to match f

and this entity cannot be an

agent. This last assumption implies that there is no

additional cost for the update process if the message

processing is executed by the agents. The following

generic ﬁlter illustrates our assessment, it is shown in

a classical “if” structure:

Tests N

i f ∧

i=1..n

(?m) (1)

and ∧

i=1..n

(?a) (2)

and ∧

i=1..n

a,m

(?a,?m) (3)

and ∧

i=1..n

(?c) (4)

and ∧

i=1..n

m,c

(?m,?c) (5)

and ∧

i=1..n

a,c

(?a,?c) (6)

The elements ?a,?m and ?c are variables that re-

spectively match an agent, a message and an entity.

Let T

be a test on a x entity and T

x,y

a test between

two entities. For the ﬁlter example f

: an agent (?a) is

interested by a request message (?m) (1) if it is avail-

able (2), if this request message is close to the agent

and the agent ?a has the skill speciﬁed in the message

(3) and if the victim to evacuate (?c) (5) is diagnosed

(4). T

(?a) is the result of the ﬁrst test on an agent

related to its availability and ∧

i=1..n

(?a) are all the

tests on the agent receiver. T

a,m

(?a,?m) is the result

of the ﬁrst test on an agent and a message related to

the requested skill and ∧

i=1..n

(?a,?m) are all the

tests on the receiver and the message.

Evaluation of the Number of Tests. A ﬁlter is

composed of two types of tests. The ﬁrst type is a

comparison between the value of a property and a

constant: the tests (1)(2)(4). The second type of test

implies a matching process: the tests (3)(5)(6). For

each of these types the cost in number of tests is re-

lated to the number of entities that have to be tested

and on the result of the previous tests on the same set

of entities. For example, the number of agents that

is tested by T

(?a) is related to the number of agents

that have validated the test T

(?a). Let t

be the per-

centage of entities that validates the test i.

Let |C

| be the number of tests and C

be the num-

ber of evaluated tests of (i), the value of C

is:











1 +

∑

j=1

∏

k=1

if |C

| > 1

1 if |C

| = 1

0 if |C

| = 0

The cost of a ﬁlter is the sum of the costs of its

tests (sequentially validated). The following table

gives the cost of the generic ﬁlter if one message has

to be processed by the environment (column E) or by

the n

agents (column A) .

N E A tested entities

(1) C

∗C

= 1

(2) n

∗C

all the agents

in E and each

agent individu-

ally (A)

(3) p

∗ n

∗C

∗ n

∗C

= percent-

age of agents

that have the re-

quested state

(4) n

∗C

∗ p

∗

)∗ n

∗C

=number of

entities that are

tested in (4).

= percent-

age of agents

that are related

to the message.

(5) p

∗ n

∗C

∗ p

∗

) ∗ (p

∗

) ∗C

=percentage

of entities re-

lated to the con-

text (4)

(6) (p

∗ p

∗

) ∗(p

∗

∗n

)∗C

∗ p

∗

) ∗ (p

∗

∗n

)∗C

=percentage

of entities vali-

dating (5)

There are two differences between these two pro-

cessing: 1) the factorization of the tests if the process-

ing is executed by E; 2) the implicit elimination of the

agents if the processing is executed by the agents. If

the processing is executed by E the tests are factor-

ized, for example (1) is executed only once and is re-

peated for n

agents. Each test (i) that is false implies

an elimination of the agent that executes the the pro-

cessing. For example in (5) only the subset of agents

where the test (4) is validated continue the processing.

Following the hypothesis that the ﬁltering and the

shared information are the same, the matching done

by the agents is more costly than by the environment

for the tests (1)(4)(5) and are the same for (2)(3)(6).

We conclude that the mediation by environment is al-

ways better than the broadcast when at least one agent

is interested in the message. Nevertheless the im-

plicit agent elimination implies that if the agents do

not share all information some part of the ﬁlter evalu-

ation can be less costly for the agents than for the en-

vironment. This parameter will be empirically tested

in section 5.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

130

Moreover, the environment and the agents need

the same information to process the ﬁlters and the re-

sult has to be balanced by the cost of the access to this

information.

Evaluation of the Number of Messages. If the ex-

change of information is done via messages, there are

two types of messages to take into account: 1) the

messages related to the interactions in the MAS such

as the request messages in our example f

; 2) the

messages related to the update process.

Firstly we evaluate the number of messages that

are the result of the interaction in the MAS. Let n

be the number of messages sent by the agents, let p

be the average percentage of agents that have to re-

ceive the messages. If the matching process is exe-

cuted by the environment then each message is sent

to the environment, and the environment transmits it

to the p

∗n

receivers. The total number of messages

is therefore n

(1 + p

∗ n

). If the matching process

is executed by the agents, then each message is sent

to all the agents which locally execute the matching

process. The total number of messages is therefore

∗ n

. Except if p

is close to 1 which means that

the message is related to all the agents, the environ-

ment mediation is less costly. The more the selection

is important, the more the mediation by the environ-

ment is beneﬁcial.

Secondly we evaluate the number of messages that

are the result of the update process of the MAS. To

simplify, we consider only the update process of the

descriptions of the agents and not of the context, al-

though the principle is the same. In our example, it

implies that the state of the victims is not updated (or

that its cost is the same if it is computed by the envi-

ronment or by the agents). When an agent is updated

all its properties are updated in the environment. Re-

member that the ﬁlter example does not contain tests

on the state of the agents to compute the context ( f

Hence, the computation by the agents does not im-

ply additional costs because we suppose the agents

do not need to access updated information about the

other agents, which is an underestimation of the cost

of the ”agent” solution.

When the communication is mediated by the en-

vironment, the agents update their descriptions by

putting in the environment new descriptions that re-

place the old ones. This action is associated to the

sending of a message. One agent can make between

0 and n updates during a time period. Let f req

the average frequency of the description update dur-

ing a time period. There is therefore f req

∗n

update

messages if the interaction is mediated by the environ-

ment during the reference time period.

To have the total cost according to the number of

Figure 1: Comparison of the cost with p

= 0.6 and n

100.

messages criteria we have to sum these costs. Let

Cost

= n

(1 + p

)+ f req

be the total environ-

ment cost and Cost

= n

be the total agents cost.

To compare these two costs, we study the sign

of the subtraction of the two costs Cost

= Cost

−

Cost

in order to ﬁnd when the environment medi-

ation is more costly than the local computation in

function of the number of messages and of the fre-

quency of the update process. The resulting formula

is Cost

= n

(1 + n

− 1)) + f req

. Figure 1

gives the plan corresponding to this function in the

three dimensional space (message, agent update, cost)

if the number of agents (n

) and the average propor-

tion of receiver (p

) are respectively ﬁxed to 100 and

0.6. These assessment parameters are unfavourable

to the mediation by the environment: 1) the cost of

the update process is not taken into account when the

context is computed by the agents; 2) the agents are

interested by more than half the messages. Neverthe-

less we can see that there are few cases in which the

use of the environment is more costly than the local

matching according to the number of messages crite-

ria.

The function Cost

, f req

) = 0 enables to ﬁnd

the number of messages dedicated to the interaction

in the MAS (m) that have to be mediated by the envi-

ronment to compensate the cost of the update process

( f req

). The relation between these two parameters

is n

= −(n

/(1 + n

(1 − p

))) ∗ f req

, the value of

is therefore inversely proportional to the propor-

tion of the agents interested in the messages. Figure 2

illustrates this proposition. The environment area rep-

resents the cases where it is better to use the environ-

ment than to mediate the messages in function of the

number of messages. For instance, with f req

= 2,

if there is more than 5 messages related to the inter-

action in the MAS, then it is better to use the envi-

ronment to mediate the communication even if that

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE

ENVIRONMENT

131

Figure 2: Comparison of the number of messages in func-

tion of the agent update frequency with p

= 0.6 and n

100.

means that there are f req

= 200 messages related

to the update process.

This study has been done in the worst case for

the environment because the agents are not taken into

account to evaluate the interaction context ( f

). If it

were the case, there would be no modiﬁcation of the

update cost of the environment. However, when the

matching is executed by agents, the number of mes-

sages in the update process would be increased by

f req

− 1) (the cost of broadcasting the descrip-

tion updates), thus improving the comparative envi-

ronment performance.

In this section the assessment has been done with

the hypothesis that the agents have the same informa-

tion to evaluate a ﬁlter. In the next section, we depict

how these results relate to real experimentation where

this hypothesis is not valid.

5 EMPIRICAL ASSESSMENT

In section 4, we have shown that using the environ-

ment to compute the contexts is less costly than com-

puting them locally in the agents when the agents up-

date frequency is not too high. In other words, the en-

tities dynamics is the parameter that determines which

solution is the best for a multi-agent system. This

section consists in evaluating the cost of the context

computing on a real example that is a simulation of

the victim evacuation in the crisis situation. The sim-

ulation run-time results aggregate the previous calcu-

lations.

A prototype of our ABS framework has been im-

plemented as a plugin for the multi-agent platform

Madkit (Gutknecht and Ferber, 2001). This plugin is

composed of an environment component with an API

that enables the agents to add/retract/modify their de-

scriptions and ﬁlters. We have chosen to implement

the matching process within a Rules-Based System

(RBS). The instantiation of the model into a RBS is

straightforward: the descriptions are the facts of the

rule engine, and the ﬁlters are its rules. Rule ﬁring is

based on the efﬁcient RETE algorithm (Forgy, 1982).

It is a network-based algorithm designed to speed the

matching of patterns with data. RETE uses a static

discrimination network, generated by the language

compiler, that represents data dependencies between

rule conditions.

We compare our model which encompasses con-

textual activation by the environment with a model

which encompasses a classical activation process to

evaluate the cost of supporting the simulation process

through the environment. We have tested two sim-

ulation scenarios, without and with communication,

using ten ﬁlters: seven ﬁlters for contextual activa-

tion ( f

to f

), two ﬁlters for communication ( f

recept

and f

) and one ﬁlter for classical activation

classicalactivation

. Except for the ﬁlter f

classicalactivation

each activation ﬁlter allows to activate an agent med-

ical porter in a speciﬁc context to perform an action.

The ﬁlter f

classicalactivation

is used in the scenarios S

and S

in order to simulate a classical activation pro-

cess. This ﬁlter activates each agent once by simula-

tion cycle. It does not take into account the context.

The scenarios are deﬁned below. S

and S

are

scenarios without inter-agent communication, and S

and S

are scenarios with inter-agent communication:

• –S

– { f

classical activation

} + Local Agent Context

Analysis (LACA),

• –S

– { f

, f

• –S

– { f

classical activation

} + LACA + Broadcast,

• –S

– { f

, f

} + { f

recept

, f

}

illustrates a classical scenario with an activation

phase and a local agent context analysis; S

is a sce-

nario with a contextual activation inside the environ-

ment. We describe the ﬁlters of S

. The ﬁlter f

al-

lows to activate the action move randomly that is trig-

gered when a medical porter has no victim next to it

(context victim seeking). The ﬁlters f

and f

corre-

spond to the association of the action move towards

a location with the contexts closest victim discovery

( f

) and handled victim discovery ( f

). The context

closest victim discovery happens when an agent per-

ceives the closest victim to evacuate in its perception

ﬁeld. The context closest victim discovery depicts the

situation where an agent medical porter perceives a

victim that is handled by another medical porter. The

ﬁlter f

activates the action wait and triggers when a

medical porter is alone close to a victim to evacuate

(context victim proximity). The ﬁlter f

corresponds

to the association between the action evacuate victim

and the context rescue victim. This context appears

when two medical porters with the complementary

skill are close to a victim ready to be evacuated. This

context is related to the following information: the

availability of the porters, the location of the porter

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

132

and the victim. The comparison between S

and S

enables to study the cost of the activation process.

and S

are the extensions of respectively the

scenarios S

and S

with the communication ﬁlters

recept

and f

, and the associated activation ﬁl-

ters f

and f

. The ﬁlters f

recept

and f

provide

the support for the communication process: f

recept

en-

ables the reception of the request messages and f

enables the reception of their answer(s). When the

agents communicate, the simulation has to manage

two new agent actions that are activated by the ﬁlters

and f

. The ﬁlter f

is related to the action con-

tact: an agent puts a request message in the environ-

ment and waits for an answer. The ﬁlter is activated

when the agent is close to a victim to evacuate but

does not have another agent nearby to evacuate this

victim. The ﬁlter f

is related to the action coordina-

tion: an agent answers to the contact agent and moves

towards the victim location. The context of this ﬁl-

ter is the reception of a request message containing

the location of the victim. The comparison between

and S

enables to study the cost of the activation

when the medical porters use communication proto-

col to contact another medical porter with the skill

required.

Activation. We have run three series of simulations

characterized by two parameters: number of agents

medical porters and ﬁeld of vision. For each series

of simulations, we evaluate our model using the aver-

age run-time over 50 simulations with similar param-

eters. In the ﬁrst group, we have experimented with 5

medical porters with the skill “health monitoring”, 5

medical porters with the skill “victim handling” and 5

victims, in the second group with 15 of each, and in

the third group with 20 of each.

Figure 3 shows the result for the third experiment.

As we can see, the scenario S

is faster than the clas-

sical scenario S

, except for the lowest ﬁeld of vision.

For a ﬁeld of vision from 8 to 30, S

run-time curves

(dotted curves) are below S

run-time curves (solid

curves). The increase of the ﬁeld of vision improves

the knowledge of the agents on the environment and

should improve their efﬁciency: the probability for an

agent to perceive victims and other agents increases.

This is true with the use of the environment in the sce-

nario S

, a larger ﬁeld improves the efﬁciency of the

agents: as the ﬁeld of vision increases from 5 to 30,

run-time decreases from around 4000s to 800s. But

with the classical activation process in the scenario

, the improvement of victim perception barely off-

sets the cost of local context analysis. We can observe

that in most cases, the cost of the context computing

in each agent is more expensive in terms of simula-

tion run-time than the cost of computing the context

Figure 3: Simulation run-time with the strategies S

and S

for the third group.

inside the environment. The results are similar for the

ﬁrst and second group of experimentations.

Communication. In the following experimenta-

tion, we compare the use of broadcast and the use

of the environment for the communication. In these

tests, only the communication ﬁlters f

recept

and f

of the scenario S

are evaluated and compared to the

broadcast of the scenario S

. The message exchange

to ﬁnd partners has been instantiated:

• The agents send “request” messages. The agents

interested in these messages are those which are

close to the sender and have a different skill, if the

victim can be moved. In scenario S

, the message

is broadcasted to all the agents, while in S

, they

are managed by f

recept

• Those agents answer with either an accept or a

reject message through an addressed message. In

scenario S

, the message is sent to the agent via

point-to-point communication, while in S

, they

are managed by f

Each data is an average on 50 simulations. Each

simulation is composed of 2000 steps. We have run

two series of simulations characterized by two param-

eters: number of agents and number of updates during

the simulation.

Firstly, we have tried to use the RETE algorithm to

implement the communication ﬁlter, and compared it

to the broadcast and addressed messaging capabilities

of MadKit. The following table sums up the simula-

tion run-time in function of the messaging implemen-

tation: EASI with a RETE tree, MadKit Messaging

support, ad hoc Environment and Broadcast.

Agent RETE Madkit Environment Broadcast

10 4624 1096 125 173

20 15215 4050 394 596

50 83785 26306 2377 4755

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE

ENVIRONMENT

133

These results show that the cost of a RETE tree off-

sets the gains in the message treatment. This discrep-

ancy with the previous results for activation can be

explained by the cost of the addition and removal of

the messages. In the activation part, the entities are

the same all along the simulation, and only rule ﬁring

and description modiﬁcations take place.

This has lead us to use an ad hoc implementa-

tion of the environment for communication, which

treats the calculations in the same way as described

in section 4. However, the comparative results show

that the MadKit implementation of the broadcast and

addressed messaging is not efﬁcient because of its

genericity. Therefore, we also re-implemented the

broadcast and point-to-point communications using

the same data structures as the environment, in order

to have a fair comparison between the environment

and the broadcast solutions.

In the following, we compare the two last solu-

tions. Firstly, we study the impact of a variation in the

number of agents. The results shown in Fig. 5 feature

10 to 100 agents. The curve shows a clear advan-

tage of the environment over the broadcast, and that

the more agents there are, the more EASI is compara-

tively interesting (from 27.8% for 10 agents to 53.9%

for 100 agents).

5000

10000

15000

20000

25000

30000

35000

0 10 20 30 40 50 60 70 80 90 100

Ag e nt s

Tim e (ms)

Environment

Broadcast

Figure 4: Comparative results of Broadcast and EASI in

function of the number of agents.

In the second phase (Fig. 5),we study the impact

of a variation in the number of updates, for 50 agents.

During each update, all the agents move and in the

case of the Environment implementation, they mod-

ify their description accordingly. The run-time gain

of EASI over broadcast gets from 50.7% for 20 up-

dates to 48.8% for 1000 updates. Therefore, the cost

of the update mechanism is not signiﬁcant in compar-

ison with the difference in efﬁciency.

These experimentations show a clear advantage to

the environment over the broadcast, and this result is

backed up by the theoretical results. It shows the ef-

1000

2000

3000

4000

5000

6000

0 100 200 300 400 500 600 700 800 900 1000

Upd ates

Tim e (ms)

Broadcast

Environment

B minus E

Figure 5: Comparative results of Broadcast and EASI in

function of the number of description updates.

ﬁciency of the context computing inside the environ-

ment in comparison with the classical approach where

the context is computed in each agent.

6 DISCUSSION

AND CONCLUSIONS

In the EASI/EASS models, the environment offers a

technical support for communication and activation.

Its processing is the result of the ﬁlters triggering ac-

cording to the MAS descriptions. In that way, the

agents add or remove dynamically ﬁlters and update

their description. Thanks to the ﬁlter triggering, an

agent receives a message or is activated in a speciﬁc

context. The processing of the messages and the ac-

tion execution remain to the agents responsibility.

This paper is focused on the cost of the centraliza-

tion, which is closely related to the cost of the update

process. Organizations, such as the agent-group-role

model (Ferber et al., 2002) enable to decrease the cost

but do not take into account ambient criteria like the

location of the agents (e.g. in the simulation). Let us

note that our modeling enables to reproduce a selec-

tion of percepts according to the organisation. Insti-

tutions (see e.g. (Esteva et al., 2004)) generally do

not share the same objective. The focus is on control

and not on the ﬁltering / matchmaking. However, the

technical solutions may be close to ours.

The use of a shared knowledge has already been

done within agents (for example Sycaras work, e.g.

(Sycara and Wong, 2000)) but in that case the up-

date of the properties is not considered. In this pa-

per, we focus on the cost of the update process that

could limit the interest to centralize speciﬁc informa-

tion. Furthermore, middle-agents are generally used

only as a ﬁrst step to ﬁnd contacts, and not to manage

all the communications. In our view, the environment

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

134

is a facility, which can be used to facilitate the inter-

action, apply norms, or verify some rules related to

the application design. These roles do not belong to

the same design level as the agents.

Theoretically, we have shown that if the commu-

nication takes the context into account, there is no

strictly dominant solution. It depends on the dynam-

icity of the multi-agent system, the number of agents

and the average percentage of agents interested in

each message. According to the number of tests cri-

teria, we have shown that the environment is always

better than the local context computation. According

to the number of messages criteria, the result has to

take into account the number of messages related to

the MAS activity and the number of messages related

to the update process. We have shown that the en-

vironment solution is generally better to mediate the

communication of the MAS activity and that few mes-

sages to mediate are needed to compensate the cost of

the update process.

To propose an empirical assessment of the cost of

the environment, we have studied the run-time crite-

rion in the crisis simulation example. We compare

the cost of the local context analysis for each agent

to a central and global control ensured by the envi-

ronment, and the cost of communication. The main

conclusion is that the environment cost is signiﬁcantly

lower than the local agent calculation of the context

perception, except when there are very few agents.

In the future, we intend to investigate different

ways to improve the environment performance. An

ongoing effort concerns the theoretical evaluation of

a RETE-based instantiation of the model. We also

study how to take advantage of the ﬁlter and entity

structures to speed up the matching process.

REFERENCES

Badeig, F., Balbo, F., and Pinson, S. (2007). Contextual

activation for agent-based simulation. In ECMS’07,

pages 128–133.

Bousquet, F., Bakam, I., Proton, H., and Page, C. L. (1998).

Cormas: Common-pool resources and multi-agent

systems. In Pobil, A. P. D., Mira, J., and Ali, M., edi-

tors, IEA/AIE (Vol. 2), volume 1416 of Lecture Notes

in Computer Science, pages 826–837. Springer.

Branigan, H. (2006). Perspectives on multi-party dialogue.

Research on Language & Computation, 4 (2-3):153–

177.

Esteva, M., Rodriguez-Aguilar, J., Rosell, B., and Arcos,

J. (2004). Ameli: An agent-based middleware for

electronic institutions. In Jennings, R., Sierra, C.,

Sonenberg, L., and Tambe, M., editors, Proceedings

of the third International Joint Conference on Au-

tonomous Agents and Multi-Agent Systems (AAMAS

2004), pages 236–243. ACM Press.

Ferber, J., Gutknecht, O., Jonker, C., Muller, J., and Treur,

J. (2002). Organization models and behavioural re-

quirements speciﬁcation for multi-agent systems.

Forgy, C. L. (1982). Rete : A fast algorithm for the many

pattern/many object pattern match problem. Artiﬁcial

Intelligence, 19:17–37.

Gutknecht, O. and Ferber, J. (2001). The madkit agent plat-

form architecture. In Revised Papers from the Inter-

national Workshop on Infrastructure for Multi-Agent

Systems, pages 48–55. Springer-Verlag.

Kesaniemi, J., Katasonov, A., and Terziyan, V. (2009). An

observation framework for multi-agent systems. Au-

tonomic and Autonomous Systems, International Con-

ference on, pages 336–341.

Laberge, D. (1975). learning and cognitive processes (Vol.

4), chapter 5 - Perceptual Learning and Attention,

pages 237–273. W.K. Estes.

Luke, S., Ciofﬁ-Revilla, C., Panait, L., Sullivan, K., and

Balan, G. (2005). Mason: A multiagent simulation

environment. Simulation, 81(7):517–527.

Picco, G. P. and Buschini, M. L. (2002). Exploiting tran-

siently shared tuple spaces for location transparent

code mobility. In COORDINATION ’02: Proceedings

of the 5th International Conference on Coordination

Models and Languages, pages 258–273, London, UK.

Springer Verlag.

Platon, E., Sabouret, N., and Honiden, S. (2005). Over-

hearing and direct interactions: Point of view of an

active environment. In Proceedings of Environment

for Multi-Agent Systems, Workshop held at the Fourth

Joint Conference in Autonomous Agents and Multi-

Agent Systems, volume 3830 of Lecture Notes in Arti-

ﬁcial Intelligence, pages 121–138. Springer Verlag.

Saunier, J. and Balbo, F. (2009). Regulated multi-party

communications and context awareness through the

environment. International Journal on Multi-Agent

and Grid Systems, 5(1):75–91.

Sycara, K. and Wong, H. (2000). A taxonomy of middle-

agents for the internet. In ICMAS ’00: Proceedings

of the Fourth International Conference on MultiAgent

Systems (ICMAS-2000), pages 465–466, Washington,

DC, USA. IEEE Computer Society.

Weiss, G., editor (1999). Multiagent Systems: A Modern

Approach to Distributed Artiﬁcial Intelligence. The

MIT Press, Cambridge, MA, USA.

Weyns, D., Omicini, A., and Odell, J. (2007). Environment

as a ﬁrst-class abstraction in multiagent systems. Au-

tonomous Agents and Multi-Agent Systems, 14(1).

Weyns, D., Steegmans, E., and Holvoet, T. (2004). Towards

active perception in situated multi-agent systems. Spe-

cial Issue of Journal on Applied Artiﬁcial Intelligence,

18 (9-10):867–883.

EVALUATING THE COST OF SUPPORTING INTERACTION AND SIMULATION THROUGH THE

ENVIRONMENT

135