Communication and Negotiation to Improve Agent-Based Models

Alejandro Rodr

ıguez-Arias

, Noelia S

anchez-Maro

and Bertha Guijarro-Berdi

nas

Universidade da Coru

na, CITIC, Campus de Elvi

na, s/n, 15008 A Coru

na, Spain

Keywords:

Agent-Based Modeling, Multi-Agent System, Agent Communication Language, HUMAT, Sustainability.

Abstract:

Agent-based models (ABM) play a fundamental role in studying and modeling complex real-world systems,

primarily relying on reactive agents. Despite their simplicity, the interactions between agents and their envi-

ronment enable the simulation of diverse systems, contributing to their widespread adoption, particularly in

the social sciences. Similarly, though distinct in purpose, multi-agent systems (MAS) are designed to tackle

complex, diverse, and distributed problems by leveraging communication, negotiation, and coordination capa-

bilities. Both types of approaches have been used successfully in numerous areas; the power of ABM lies in

thousands of interacting agents, while MAS usually employs a smaller number of agents with more capabil-

ities. Including MAS agents’ capabilities in ABM agents allows the generation of more realistic simulations

that aid in the study of the modeled systems. In this paper, we present a generic ABM model whose agents

possess more capabilities, such as communication and negotiation, allowing this enhanced ABM to address

more complex modeling problems. To exemplify the usefulness of this enhanced ABM, we propose to use it as

a sandbox-tool to test “case-if” scenarios in a model that studies the evolution of a society’s opinion on a given

subject, speciﬁcally in this example, the implantation of superblocks in the city of Vitoria-Gasteiz (Spain).

1 INTRODUCTION

Science aims to understand real-world systems, their

patterns, and responses to various conditions. Direct

analysis is often slow, costly, or impractical, mak-

ing virtual systems essential for experimentation via

modeling.

In AI, agent-based modeling (ABM) is a key

method to represent complex systems. ABM depicts

real-world phenomena through agents, their environ-

ment, and their interactions. Agents are deﬁned by

unique variables, actions, and environmental inter-

pretations, enabling the creation of diverse, heteroge-

neous societies that mirror real-world systems (Rails-

back and Grimm, 2019; Wilensky and Rand, 2015).

Multi-agent systems (MAS) (Dignum, 2017;

Wooldridge, 2009) address complex, distributed

problems by enabling collaborative strategies among

agent groups, facilitating realistic interactions for

solving real-world challenges. To support this, inter-

agent communication languages (ACLs) have been

developed (Soon et al., 2019).

While ABM focuses on social sciences and MAS

https://orcid.org/0000-0002-0140-7473

https://orcid.org/0000-0003-4025-1405

https://orcid.org/0000-0001-8901-5441

on engineering, the demand for advanced ABM in so-

cial sciences is growing (Steinbacher et al., 2021). Ta-

ble 1 summarizes their key differences. In this paper

we will focus on the difference in agent complexity,

and the negotiation, coordination, and cooperation ca-

pabilities associated with MAS.

Table 1: Key differences between ABM and MAS.

Aspect ABM MAS

Purpose

Study emergent

behaviors.

Solve practical,

problems.

Field

Research and

simulation.

AI, engineering

Focus

Emergent phe-

nomena.

Agent coopera-

tion/competition.

Complexity

of Agents

Simple rule-

based behaviors.

Advanced

decision-

making.

Environment

Static or simpli-

ﬁed.

Dynamic and

real-time.

An ABM example requiring more complex agents

is a forest ﬁre propagation model. Fires act as agents,

with the system simulating their evolution under vary-

ing atmospheric conditions. Adding agents like ﬁre-

ﬁghters or ﬁre trucks would enable analysis of their

1388

Rodríguez-Arias, A., Sánchez-Maroño, N. and Guijarro-Berdiñas, B.

Communication and Negotiation to Improve Agent-Based Models.

DOI: 10.5220/0013377500003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1388-1395

ISBN: 978-989-758-737-5; ISSN: 2184-433X

effectiveness in extinguishing ﬁres, necessitating ad-

vanced communication, cooperation, and coordina-

tion capabilities to optimize resource management.

Another example is modeling local fauna in an ABM,

where invasive species are agents. Incorporating

hunter agents who negotiate and coordinate to eradi-

cate invasive species introduces MAS functionalities.

In both cases, starting from a standard ABM with re-

active agents, adding MAS-like agents (ﬁreﬁghters or

hunters) enhances the model’s ability to address com-

plex real-world problems.

In this work, we build on a conceptual ABM de-

veloped in prior studies (Antosz et al., 2019; Antosz

et al., 2020; Bouman et al., 2021; Rodr

ıguez-Arias

et al., 2024). The model represents societies based on

sociodemographic characteristics and personal needs,

enabling the analysis of opinion evolution on speciﬁc

topics. It also highlights the inﬂuence of key group

entities or individual agents whose impact on opinion

dynamics is particularly signiﬁcant.

The ABM was implemented using NetLogo

(Wilensky and Rand, 2015), a popular platform for

agent-based models but lacking MAS functionalities

and goal-oriented architectures (e.g., Belief-Desire-

Intention). This work enhances the ABM by: (a)

equipping NetLogo agents with basic communica-

tion, negotiation skills, and interaction protocols; (b)

integrating these agents into the model to enable com-

plex behaviors; and (c) showcasing these features

through a use case. The example used is citizen ac-

ceptance of a superblock project

. However, the en-

hanced ABM is adaptable to various problems and so-

cieties (Bouman et al., 2021; Rodr

ıguez-Arias et al.,

2024).

2 STATE OF THE ART

Agent-Based Models (ABMs) are widely used for

modeling and simulation across various ﬁelds. In

health, Escudero et al. (Escudero et al., 2016) mod-

eled HIV transmission in New York City (1996–2012)

using an ABM. In sociology, Crooks (Crooks, 2010)

used ABMs with vector GIS to study residential seg-

regation in urban settings. In emergency manage-

ment, Dawson et al. (Dawson et al., 2011) developed

an ABM for policy analysis to improve ﬂood incident

management.

Agents in agent-based models typically use a reac-

tive architecture, relying on simple stimulus-response

rules (Kaelbling et al., 1987). These agents perceive

A superblock is a group of city blocks reorganized to

prioritize pedestrians over vehicles

their environment and act without memory or plan-

ning. This approach suits models where emergent

phenomena arise from numerous simple interactions,

enabling the observation of collective patterns with-

out complex individual behaviors.

Cognitive architectures mimic human cognitive

processes through three components: 1) the agent’s

perception of the world and its ability to sense

it, 2) memory for storing information, and 3) a

decision-making model. A popular example is the

Belief–Desire–Intention (BDI) architecture, where

agents operate based on beliefs (assumed truths), de-

sires (goals), and intentions (committed actions) (Rao

and Georgeff, 1997).

Some ABMs use hybrid architectures, blending

cognitive and reactive elements. This allows agents

to plan and make decisions deliberatively while react-

ing to immediate stimuli. Hybrid architectures com-

bine the strengths of both approaches, providing ﬂex-

ibility and adaptability for simulating complex sys-

tems (Guessoum, 1997). For instance, Bussmann et

al. propose a hybrid architecture for autonomous mo-

bile robots (Chin et al., 2014).

Our goal is to implement an enhanced hybrid ar-

chitecture that incorporates MAS communication lan-

guages, aiming to increase the complexity and capac-

ity of traditional ABMs.

3 THE STANDARD ABM

The original ABM (Antosz et al., 2019; Antosz et al.,

2020; Bouman et al., 2021; Rodr

ıguez-Arias et al.,

2024) consists of two distinct agent types: 1) Humats,

representing individuals in the society under study,

based on the HUMAT architecture (Antosz et al.,

2018); and 2) critical nodes, representing entities or

individuals with a signiﬁcant inﬂuence on the opinion

dynamics of the Humats. The following subsections

detail both agent types.

3.1 Humat Agents

A Humat must make a decision regarding the topic

under study, using sociodemographic characteristics

and psychosocial needs. At least three psychologi-

cal needs are considered: (1) experiential needs (e.g.,

personal well-being), (2) values (e.g., concern for en-

vironmental quality), and (3) belongingness (e.g., so-

cial group afﬁliation). Each need is deﬁned by two

factors: (1) the importance I the Humat places on it,

and (2) the satisfaction S it derives from it. For in-

stance, COVID-19 measures like mask use and social

distancing reduce virus spread but complicate social-

Communication and Negotiation to Improve Agent-Based Models

1389

ization and family visits. The importance individuals

assign to these needs inﬂuences their decision to ad-

here to preventive health measures.

To represent societal heterogeneity, each Humat

has sociodemographic properties that, along with psy-

chosocial needs, deﬁne it. The HUMAT architecture

is ﬂexible enough to accommodate different sociode-

mographic properties and needs, depending on the

problem being modeled (e.g., (Antosz et al., 2020;

Bouman et al., 2021)).

Each Humat is placed in a virtual environment,

which can be geographical, social, or a combination

of both. By default, the model creates a 2D virtual

space that can represent various settings (e.g., an of-

ﬁce, city, or forest). Additionally, each Humat is part

of one or more social networks, allowing it to inﬂu-

ence and be inﬂuenced by other Humats.

The life cycle of a Humat can be divided into two

main phases: (1) self-evaluation and (2) information

exchange.

Phase 1. Self-Evaluation. In each life cycle a Hu-

mat agent must assess its internal needs in order to

choose a decision or behavioural alternative that best

satisﬁes its individual needs. This is done by calcu-

lating the overall satisfaction (O) expected from each

behavioural alternative using the equation (1).

∑

n=1

b,n

∗ I

∈ [−1, 1], b = 1, 2 (1)

where S is the satisfaction value of an Humat for a

need n and a behavioural alternative b and I is the im-

portance that the Humat attaches to the need n. N

is the number of needs that motivate the decision-

making. Humats will choose the behavioural alter-

native with the highest expected overall satisfaction.

After choosing the behavioural alternative that

most satisﬁes their needs, Humats evaluate whether

this alternative generates any cognitive dissonance in

them. A dissonance occurs when a behavioural alter-

native generates positive or negative evaluations for

one need and the opposite sign for the others, where

the evaluation E of a particular need is calculated as

b,n

= S

b,n

∗ I

∈ [−1, 1] n = 1. . . N, b = 1, 2

(2)

resulting in a positive (E

) or negative (E

−

) evalua-

tion. The strength of an Humat’s dissonance is calcu-

lated following the equation (3).

+ c

∈ [0, 1] (3)

where

= min



∑

n=1

b,n



∑

n=1

−

b,n



= max



∑

n=1

b,n



∑

n=1

−

b,n



When the strength of the dissonance exceeds a

certain tolerance threshold, the Humat will try to act

to resolve it by communicating with other Humats in

its same social network, as described below.

Phase 2. Information Exchange. When an agent

faces a dilemma due to dissonance exceeding their

tolerance threshold, they seek resolution by commu-

nicating with their social network.

The persuasiveness (P) of one Humat over another

during communication depends on: (a) the weight α,

which balances an individual’s opinion against exter-

nal inﬂuence, (b) the trust (T ) of the inﬂuenced Humat

in the communicator, and (c) the similarity of needs

enced agent’s satisfaction is updated using equation

(4). The α weight ranges from 0 to 0.4, ensuring an

individual’s opinion always outweighs external inﬂu-

ence.

b,n

= α ∗ T ∗C

b,n

(4)

Respect to similarity C, if the evaluation of a need

n (see equation (2)) has a different sign in both agents,

then the similarity of that need is 0, otherwise, it fol-

lows equation (5):

b,n

= 1 − |I

b,n,e

− I

b,n,o

| (5)

where I is the importance of the need, b is the be-

havioural alternative, o is the agent that inﬂuences.

After the communication, the new satisfaction

value of the inﬂuenced agent e is calculated as shown

in equation (6).

b,n,e

(t + 1) = (1 − P

b,n

) ∗ S

b,n,e

(t) + P

b,n

∗ S

b,n,o

(t)

(6)

3.2 Critical Nodes

The second type of agent is the critical node, which

represents inﬂuential collectives or key agents shap-

ing opinion evolution. Examples include local media,

inﬂuencers, or academic ﬁgures like professors.

Critical nodes hold individual opinions on the

topic being modeled (e.g., promoting sustainable

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1390

Table 2: Critical node communication act parameters.

Parameter Value

Behaviour Supporter/Opponent

Reach

Integer in [0,100]

Start month Integer in [1,12]

Start year Integer in [1,12]

End month Integer in [1,12]

End year Integer in [1,12]

Frequency per month Integer in [1,2]

Primary critical node Any critical node

Secondary critical node Any critical node

transportation), which can evolve over time. They in-

ﬂuence Humats through communication actions, op-

erating based on a communication plan rather than

seeking to reduce cognitive dissonance.

Each critical node is deﬁned by parameters such as

the percentage of Humats it can reach, its geograph-

ical coordinates, and a communication plan. The so-

cial network of a critical node is determined by the

Humats it can inﬂuence.

A communication plan consists of a series of

dated communicative acts, each characterized by

scope (percentage of the social network affected), fre-

quency, and date parameters (as shown in table 2). A

communicative act can be for or against one of the

decisions that the Humats can take. Up to two critical

nodes can be involved in a communication: the pri-

mary node initiates the process, while the secondary

node carries out the communication through its net-

work. An example of a communication plan in the

superblock modeling case is provided in section 5.

3.3 Model Limitations

Humats are purely reactive agents, receiving commu-

nications from other Humats or critical nodes, updat-

ing their internal state, and responding accordingly.

In contrast, critical nodes exhibit a more proactive

behavior by following a ﬁxed communication plan

throughout the simulation.

While this setup allows for the analysis of simple

opinion dynamics, it limits the model’s ability to rep-

resent complex and adaptive behaviors. This is par-

ticularly challenging when studying scenarios requir-

ing collaboration, negotiation, or strategic planning.

For example, the current model cannot capture how

agents might adapt strategies, form coalitions, or en-

gage in negotiation. As a result, it struggles to repre-

sent real-world dynamics where agents must balance

personal interests with community goals, respond to

unforeseen events, or adjust based on feedback. To

address these limitations, agents need communication

and reaction capabilities typical of MAS.

4 ENDOWING AGENTS WITH

NEGOTIATION CAPABILITIES

The goal of this work is to enhance ABM agents, en-

abling them to create and execute more complex plans

and behaviors. This requires hybridizing the model

with multi-agent system features by equipping agents

with intelligence, a new architecture, and a more so-

phisticated communication language, as detailed be-

low.

4.1 BDI-Like Agents in Netlogo

The ﬁrst step is to convert some agents into more

complex agents by implementing a BDI-like architec-

ture. The BDI architecture consists of 3 main compo-

nents (Rao and Georgeff, 1997; De Silva et al., 2020):

• Beliefs, which represent information about the

state of the world held by the agent.

• Desires, that are the agent’s design objectives, i.e.

those goals to be achieved.

• Intentions, which are the tasks that are part of the

agents’ plan to achieve certain objectives.

Since NetLogo does not natively support BDI

agents, we propose a table-based implementa-

tion where each agent maintains beliefs as “con-

cept”–“value” pairs in a table, representing its per-

ceptions of the world. Desires are deﬁned by a set

of beliefs the agent intends to be true.

Finally, intentions are managed using a First-In,

First-Out queue structure, where each entry consists

of a pair (”operator to use” and a ”stop condition”).

The operator speciﬁes the function the agent should

execute, and the stop condition determines when the

intention should be removed from the queue.

4.2 Adding Communication Capacity

The second step in enhancing agent intelligence is

to provide more complex communication capabilities.

Currently, agents can only use the inquire/signal ac-

tions of the HUMAT architecture, enabling inﬂuence

communications but not real information exchange.

To address this, agents need a structure to manage and

send messages, along with an agent communication

language (ACL) to deﬁne them.

An (ACL) is a standard that enables agents to ex-

change information about plans, goals, and beliefs

(Genesereth and Ketchpel, 1994). ACLs deﬁne the

types and meanings of messages exchanged between

agents. Most ACLs are based on speech act theory

(Austin, 1975), where messages are communicative

Communication and Negotiation to Improve Agent-Based Models

1391

acts intended to prompt the receiving agent to take

action. In this work, we use the FIPA-ACL commu-

nication language from the Foundation for Intelligent

Physical Agents (FIPA).

FIPA deﬁnes a set of performatives that specify

the type of communicative act, indicating the intent

of messages (Wooldridge, 2009). These acts have

a well-deﬁned meaning independent of the message

content (O’Brien and Nicol, 1998), as outlined in the

FIPA communicative acts library speciﬁcation (FIPA,

2000). The range of performatives varies from sim-

ple information exchange to requests for tasks. Along

with the performative label, a FIPA ACL message

includes a set of parameters, which depend on the

agent’s current situation for effective communication.

NetLogo does not natively support ACLs. To

enable communication, we developed a library

for creating and modifying messages. In NetL-

ogo, a message is represented as a table, with

the ”performative” parameter being the only re-

quired attribute. The library includes functions

to add additional FIPA-deﬁned parameters, such as

sender, receiver, and content, alongside the performa-

tive. This library is available at https://github.com/

alejandrorodriguezarias/EnhancedABM

Each communicating agent must have functional-

ity to manage its conversation structure and handle

message exchange. A conversation records all mes-

sages sent or received under the same conversation

identiﬁer. An agent can engage in multiple conversa-

tions with different agents, or even multiple conversa-

tions with the same agent. When an agent receives a

message, it is added to the existing conversation if the

agent is the intended recipient, or a new conversation

is initiated if it’s the ﬁrst message in that interaction.

Standardized communication frameworks, known

as communication protocols, enable agents to ex-

change messages, negotiate, and make decisions efﬁ-

ciently. One widely used protocol is the Contract Net

Protocol (CNP) (FIPA, 2002), which manages task

allocation among autonomous agents in a distributed

system. In the CNP (illustrated in Figure 1), an initia-

tor agent delegates a task to one or more participants.

The initiator broadcasts a request for proposals with

evaluation criteria. Participants respond with propos-

als, and the initiator selects the best agent based on

the bids, awarding the contract. The selected agent

performs the task and reports back with the results.

This structured approach allows for efﬁcient task

distribution and coordination in multi-agent systems,

while also promoting a competitive environment

where agents can dynamically evaluate their partici-

pation based on their current state and capabilities.

Figure 1: Contract Net protocol ﬂowchart.

5 USING THE ENHANCED ABM:

THE SUPERBLOCKS PROJECT

The original model, developed for the European

SMARTEES project, has been used to analyze citi-

zens’ responses to various social innovations and sim-

ulate alternative political scenarios, such as new com-

munication strategies by critical nodes to assess their

impact on acceptance. In this paper, we apply the im-

proved model to the case of Superblocks, focusing on

Vitoria-Gasteiz, which was a pioneer in implementing

this urban innovation. The project faced signiﬁcant

public criticism, particularly regarding policies that

increased on-street parking costs. The model reﬂects

the project’s implementation (2006-2013) and was

validated by stakeholders and experts in the SMAR-

TEES project.

During the SMARTEES project, numerous pro-

moters and detractors of social innovations con-

tributed to ensuring the model’s reliability. They pro-

posed various scenarios to explore the social inno-

vation from different perspectives. Despite the suc-

cessful outcomes and the interesting model developed

(Antosz et al., 2019; Antosz et al., 2020; Bouman

et al., 2021; Rodr

ıguez-Arias et al., 2024), some re-

alistic scenarios could not be implemented due to

the limitations of the agents, which lacked commu-

nication and negotiation capabilities. In the follow-

ing subsections, we will describe the original ABM

adapted to the Vitoria-Gasteiz Superblocks case and

demonstrate how the agents’ new capabilities enable

the exploration of expanded alternative scenarios.

5.1 Virtual Environment

The model includes a virtual environment, consist-

ing of two components: (1) the geographical environ-

ment, represented by the city of Vitoria-Gasteiz using

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1392

its census sections and a 50x50 2D board (see ﬁgure

2), and (2) the social networks of friends and neigh-

bors to which citizens belong.

Figure 2: Vitoria-Gasteiz represented by its census sections.

Citizens are represented by a human silhouette, green if they

are in favour of the project and red if not. Critical nodes are

represented by an image of a building.

Neighbour networks were created using social cir-

cles (Hamill and Gilbert, 2009), based on agent prox-

imity. Friendship networks were generated as random

networks with constraints: (1) a minimum number of

friends per agent and (2) homophily on age and edu-

cation, with a small chance of random links. Humats

(see section 3.1) can communicate and inﬂuence oth-

ers within their network.

5.2 Agents of the Model

In the Victoria-Gasteiz case the Humats represents

citizens. Citizens use the HUMAT architecture for

decision-making and inﬂuence diffusion (see section

3.1). These agents are initialized using data from sur-

veys conducted in Vitoria-Gasteiz during the SMAR-

TEES project.

As we have explained in the section 3, using HU-

MAT architecture we have enough ﬂexibility to rep-

resent individuals with different sociodemographic

characteristics and needs. In the table 3 we can see the

properties that characterise a citizen in the superblock

model. Different problems may need different char-

acterisation, as seen in (Antosz et al., 2020; Bouman

et al., 2021).

Similarly, the critical nodes are agents (see section

3.2) that represent promoters, and institutions relevant

to the development of the social innovation. In the su-

perblocks case, the following critical nodes were in-

Table 3: Variables deﬁning a Humat agent in the su-

perblock’s project.

Variables Values

Age (years) integer in [18-120]

Gender male/female

Education level

primary/secondary/ tertiary

Economic Activity

employed/jobless /inactive

Location census tract code

Homeowner yes/no

Years in the neigh-

bourhood

<3 / 3- 10/ 10-30 / >30

cluded:

• The city council, as the main promoter of this so-

cial innovation project.

• Merchants’ associations. Throughout the imple-

mentation of the project they showed a clear re-

jection of the proposed measures to implement the

superblocks. They were the main opponents.

• Other associations, from neighbourhood associa-

tions to cycling associations. Their opinion varied

throughout the implementation of the project.

• Local press: they were the main disseminators of

information about the project.

In the table 4, we can see as an example part of

a communication plan employed in the superblock

modeling problem. In this example, the City Council

acts as a primary critical node, sometimes contracting

an advertising campaign to the press (secondary criti-

cal node) to give support and promote (behavior) the

superblocks project.

Critical nodes implement the BDI architecture

(explained at section 4) and the necessary structure

to communicate using the FIPA-ACL protocols (see

section 4.2). Speciﬁcally, in this case they have the

mechanisms to implement the contract net protocol.

5.3 Model Results

After the model was implemented, the lack of detailed

historical data on the evolution of citizen opinion led

to its validation through expert feedback in a series of

workshops conducted during the SMARTEES project

(Dumitru et al., 2021). Calibration and validation of

the model’s parameters can be found in (Rodr

ıguez-

Arias et al., 2024).

The timeline (blue curve) in Figure 3 illustrates

the historical evolution of citizen acceptability, as re-

produced by the model, with results averaged over

100 model executions. To develop the communica-

tion plans of the critical nodes, an analysis of the

newspaper library was conducted to ensure their re-

alism. Each cycle of the model represents 15 days in

Communication and Negotiation to Improve Agent-Based Models

1393

Table 4: Example of a critical node communication plan.

Primary critical

node

Behaviour

Start

month

Start

year

End

month

End

year

Frequency

per month

Reach

Secondary

critical node

City council Supporter 11 2006 12 2006 1 10% City council

City council

Supporter 1 2008 12 2008 1 1% Local press

City council Supporter 2 2009 2 2009 1 20% Local press

the simulation. As shown, public acceptance was ini-

tially high, consistent with the feedback from promot-

ers and experts. This was due to a general consensus

that led to the underrepresentation of opposing view-

points in the media discourse. Consequently, nega-

tive messages were infrequent (reﬂected in the limited

number of communicative acts), resulting in a rapid

acceptance of the superblocks. However, in Novem-

ber 2009, a new trafﬁc policy was approved, introduc-

ing trafﬁc and parking restrictions and raising public

parking costs. Resistance to this policy was initially

strong, particularly from the retail sector, but it dimin-

ished as the superblock was fully established and its

beneﬁts became more apparent to the public.

5.4 Alternative Policy Scenarios

The goal of modeling superblock implementation is

to provide a sandbox tool for stakeholders to test poli-

cies or ”what-if” scenarios aimed at improving citizen

acceptance. These scenarios modify the real, expert-

validated case.

For instance, with the enhanced BDI agents, con-

sider a scenario where the critical node ”city coun-

cil” uses this architecture to maintain citizen accep-

tance above 45%. It monitors acceptance levels and, if

they fall below 45%, executes a plan: launching a me-

dia advertising campaign for the superblocks project.

Since multiple media outlets could run the campaign,

the city council uses the Contract Net Protocol to ne-

gotiate and ﬁnalize the agreement.

For that, the municipality will launch a commu-

nication to all press agents (critical nodes) request-

ing a two-month advertising campaign in favour of

the project. Each press agent evaluates this request

and communicates a proposal with two parameters:

(1) the scope of the campaign and (2) the cost of the

campaign. The city council will select the proposal it

is most interested in and this press agent will initiate

the campaign.

This scenario evaluates whether a more reactive

developer can mitigate discontent during the project’s

most controversial phase (see Figure 3). The CNP

serves as an example of how the model beneﬁts from

enhanced agents with ACL-based communication,

but other FIPA protocols could be applied. For in-

stance, critical nodes could negotiate between cyclists

0,3

0,4

0,5

0,6

0,7

0,8

0,9

01/01/2006 01/01/2007 01/01/2008 01/01/2009 01/01/2010 01/01/2011 01/01/2012 01/01/2013

% of ci�zens accep�ng social innova�on

Timeline Case-if-scenario

Figure 3: Comparison of the evolution of citizen acceptance

between the real scenario (timeline) of Victoria-Gasteiz and

the new “case-if” scenario.

and merchants opposing the superblocks project. Ad-

ditionally, Humats could adopt BDI features to orga-

nize protest groups or political demonstrations.

6 CONCLUSIONS

Agent-Based Models (ABMs) are powerful tools for

representing complex systems and are widely used in

the social sciences. These models rely on reactive

agents and derive their strength from the interactions

between agents and their environment. However, cer-

tain phenomena require agents with enhanced capa-

bilities to accurately capture the dynamics involved.

For example, while a traditional ABM might predict

the spread of a ﬁre, incorporating agents such as ﬁre-

ﬁghters who can communicate and coordinate their

actions signiﬁcantly enriches the model’s realism and

predictive power.

We developed a NetLogo library enabling agents

with a hybrid BDI (Beliefs, Desires, Intentions) ar-

chitecture. This includes message generation, stor-

age, modiﬁcation, and communication via FIPA-

ACL, along with negotiation protocols like FIPA-

ACL Contract Net for collaboration. These agents

were integrated into an ABM framework, expanding

its modeling capabilities. In the presented example,

the enhanced framework supports stakeholders in im-

plementing social innovations by providing deeper in-

sights and facilitating informed decision-making.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1394

ACKNOWLEDGEMENTS

This work was supported by the Horizon 2020

SMARTEES project (grant no. 7639) of the European

Commission. We also acknowledge funding from the

Xunta de Galicia and ERDF funds of the European

Union through grants for research groups (ED431C

2018/34, ED431C 2022/44), CITIC as a Research

Center of the University System of Galicia (ED431G

2023/01), and the Ministry for Digital Transformation

and Civil Service under ‘Next-GenerationEU’/PRTR

(grant TSI-100925-2023-1).

REFERENCES

Antosz, P., Jager, W., Polhill, G., Jiaqui, G., Salt,

D., Alonso-Betanzos, A., S

anchez-Maro

no, N., and

Guijarro-Berdi

nas, B. (2018). Report on the concep-

tual model of the smartees simulation and data types

to be included. Deliverable 7.1, SMARTEES project.

Antosz, P., Jager, W., Polhill, G., Salt, D., Alonso-Betanzos,

A., S

anchez-Maro

no, N., Guijarro-Berdi

nas, B., and

Rodr

ıguez, A. (2019). Simulation model implement-

ing different relevant layers of social innovation, hu-

man choice behaviour and habitual structures. Deliv-

erable 7.2, SMARTEES project.

Antosz, P., Jager, W., Polhill, G., Salt, D., Scalco, A.,

Alonso-Betanzos, A., S

anchez-Maro

no, N., Guijarro-

Berdi

nas, B., and Rodr

ıguez, A. (2020). Smartees

simulation implementations. Deliverable 7.3, SMAR-

TEES project.

Austin, J. L. (1975). How to do things with words, vol-

ume 88. Oxford university press.

Bouman, L., Antosz, P., Jager, W., Polhill, G., Salt, D.,

Scalco, A., Alonso-Betanzos, A., S

anchez-Maro

no,

N., Guijarro-Berdi

nas, B., and Rodr

ıguez-Arias, A.

(2021). Report on scenario development and exper-

iments for selected cases. Deliverable 7.4, SMAR-

TEES project.

Chin, K. O., Gan, K. S., Alfred, R., Anthony, P., and

Lukose, D. (2014). Agent architecture: An overviews.

Transactions on science and technology, 1(1):18–35.

Crooks, A. T. (2010). Constructing and implementing an

agent-based model of residential segregation through

vector gis. International Journal of Geographical In-

formation Science, 24(5):661–675.

Dawson, R. J., Peppe, R., and Wang, M. (2011). An agent-

based model for risk-based ﬂood incident manage-

ment. Natural hazards, 59:167–189.

De Silva, L., Meneguzzi, F. R., and Logan, B. (2020). Bdi

agent architectures: A survey. In Proceedings of the

29th International Joint Conference on Artiﬁcial In-

telligence (IJCAI), 2020, Jap

ao.

Dignum, V. (2017). Social agents: Bridging simula-

tion and engineering. Communications of the ACM,

60(11):32–34.

Dumitru, A., Lema Blanco, I., Albulescu, P., Antosz, P.,

Bouman, L., Colley, K., Craig, T., Jager, W., Mac-

singa, I., Meskovic, E.and Mischkowski, N., Pelle-

grini Masini, G., Polhill, G., Quinti, G., Salt, D.,

Somervail, P., and Wilson, R. (2021). Policy recom-

mendations for each cluster of case-studies of case-

studies. Insights from policy scenario workshops. De-

liverable 5.2, SMARTEES project.

Escudero, D. J., Lurie, M. N., Mayer, K. H., Weinreb, C.,

King, M., Galea, S., Friedman, S. R., and Marshall,

B. D. (2016). Acute hiv infection transmission among

people who inject drugs in a mature epidemic setting.

Aids, 30(16):2537–2544.

FIPA, O. (2002). Fipa contract net interaction protocol

speciﬁcation. Document no. SC00029H.

FIPA, T. (2000). Fipa communicative act library speciﬁca-

tion. Change, 2000(01/18).

Genesereth, M. and Ketchpel, S. (1994). Communications

of the acm. Software Agents, 37(7):48–54.

Guessoum, Z. (1997). A hybrid agent model: a reactive and

cognitive behavior. In Proceedings of the Third In-

ternational Symposium on Autonomous Decentralized

Systems. ISADS 97, pages 25–32. IEEE.

Hamill, L. and Gilbert, G. (2009). Social circles: A sim-

ple structure for agent-based social network models.

Journal of Artiﬁcial Societies and Social Simulation,

12(2).

Kaelbling, L. P. et al. (1987). An architecture for intelligent

reactive systems. Reasoning about actions and plans,

pages 395–410.

O’Brien, P. D. and Nicol, R. C. (1998). Fipa—towards a

standard for software agents. BT Technology Journal,

16(3):51–59.

Railsback, S. F. and Grimm, V. (2019). Agent-based and

individual-based modeling: a practical introduction.

Princeton university press.

Rao, A. S. and Georgeff, M. P. (1997). Modeling rational

agents within a bdi-architecture. Readings in agents,

pages 317–328.

Rodr

ıguez-Arias, A., S

anchez-Maro

no, N., Guijarro-

Berdi

nas, B., Alonso-Betanzos, A., Lema-Blanco, I.,

and Dumitru, A. (2024). An agent-based model to

simulate the public acceptability of social innovations.

Expert Systems, page e13731.

Soon, G. K., On, C. K., Anthony, P., and Hamdan, A. R.

(2019). A review on agent communication language.

Computational Science and Technology: 5th ICCST

2018, Kota Kinabalu, Malaysia, 29-30 August 2018,

pages 481–491.

Steinbacher, M., Raddant, M., Karimi, F., Camacho Cuena,

E., Alfarano, S., Iori, G., and Lux, T. (2021). Ad-

vances in the agent-based modeling of economic and

social behavior. SN Business & Economics, 1(7):99.

Wilensky, U. and Rand, W. (2015). An introduction to

agent-based modeling: modeling natural, social, and

engineered complex systems with NetLogo. MIT

press.

Wooldridge, M. (2009). An introduction to multiagent sys-

tems. John Wiley & Sons.

Communication and Negotiation to Improve Agent-Based Models

1395