Seeing Through the Smoke: An Agent Architecture for Representing

Health Protection Motivation Under Social Pressure

Veronika Kurchyna

1,2,∗

, Stephanie C. Rodermund

1,2

, Ye Eun Bae

Patrick Mertes

, Philipp Fl

ugger

, Jan Ole Berndt

and Ingo J. Timm

1,2

Cognitive Social Simulation, German Research Center for Artiﬁcial Intelligence, Behringstr. 31, 54296 Trier, Germany

Business Informatics, Trier University, Universitaetsring 15, 54296 Trier, Germany

Keywords:

Agent Decision-Making, Social Simulation, Protection Motivation, Social Pressure.

Abstract:

Representing and emulating human decision-making processes in artiﬁcial intelligence systems is a challeng-

ing task. This is because both internal (such as attitude, perceived health or motivation) and external factors

(such as the opinions of others) and their mutual interactions affect decision-making. Modelling agents ca-

pable of human-like behavior, including undesirable actions, is an interesting use case for designing different

AI-systems when it comes to human-AI-interactions and similar scenarios. However, agent-based decision-

models in this domain tend to reﬂect the complex interplay of these factors only to a limited extent. To

overcome this, we enrich these approaches with an agent architecture inspired by theories from psychology

and sociology. Using human health behavior, speciﬁcally smoking, as a case study, we propose an agent-based

approach to combine social pressure within Protection Motivation Theory (PMT) to allow for a theory-based

representation of potentially harmful behavior including both internal and external factors. Based on smoking

in social settings, we present experiments to demonstrate the model’s capability to simulate human health

behavior and the mutual inﬂuences between the selected concepts. In this use case, the resulting model has

shown that social pressure is a driving inﬂuence in the observable system dynamics.

1 INTRODUCTION

In recent years, AI applications have had an increas-

ing impact on various aspects of people’s professional

and everday lives. Especially in health-related con-

texts there is a growing interest in designing inter-

active systems, e.g., in the ﬁeld of assisted living or

health care monitoring (Jovanovic et al., 2022; Qian

et al., 2021). Unlike AI systems, however, human be-

havior is often irrational and guided by motivations or

other internal processes, making it sometimes unpre-

dictable and challenging for AI systems to respond

appropriately to such behaviors. Agent-based social

simulation (ABSS) has proven to be well-suited for

examining the behaviors of individuals in response to

social inﬂuences, individual needs, and dispositions

(Davidsson, 2002; Squazzoni et al., 2014). There-

fore, it can be used for investigating how cognitive

and social factors interact and inﬂuence human be-

havior. ABSS is especially popular for its controlled

environment and wide range of possibilities, which

enables experiments that are difﬁcult or impossible

to implement in reality, and for testing mechanical

∗

Corresponding author

characteristics of psychological and social science

theories (Smith and Conrey, 2007). By simulating

autonomous individuals, a complex system of vari-

ous interactions emerges (Macy and Willer, 2002).

For example, ABSS can be used to simulate a virus

spreading among people (Tapp et al., 2022), family

planning (Berndt et al., 2018), and crowd evacuation

in the presence of ﬁre (Wagner and Agrawal, 2014).

Similarly, in health contexts, people’s decisions are

not only inﬂuenced by their own internal factors, such

as attitude and health status, but also by external fac-

tors. Social pressure, for instance, refers to the expec-

tations of the immediate social environment (Ajzen,

1991; Tesser, 1980) and it can induce individuals to

conform to such expectations or resist them, depend-

ing on their dispositions.

By combining psychological theories explaining

cognitive processes and concepts of social pressure,

models can represent the mechanisms of real systems

in an abstract manner. For instance, Protection Mo-

tivation Theory (PMT) is a widely used framework

in psychology. It illuminates how individuals assess

threats and make decisions about protective behaviors

in situations involving health risks (Floyd et al., 2000;

Hedayati et al., 2023), but also in other domains, for

Kurchyna, V., Rodermund, S., Bae, Y., Mertes, P., Flügger, P., Berndt, J. and Timm, I.

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure.

DOI: 10.5220/0012347400003636

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

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 315-325

ISBN: 978-989-758-680-4; ISSN: 2184-433X

315

example information security (Mou et al., 2022) or

pro-environmental behaviors (Kothe et al., 2019).

However, the nature of PMT is non-cyclical, not

considering dynamic environments or the outcomes

of one’s actions. To address this issue, we use social

pressure as a source of system change which creates

a feedback-loop for agents. Smoking was chosen as

use case, since both personal traits (internal factors;

PMT) and social circumstances (external factors; so-

cial pressure) have an impact on the likelihood of the

behavior (Xu et al., 2015). More precisely, individu-

als may be inclined to smoke more when perceiving

a positive attitude towards smoking in their social en-

vironment and, conversely, are more likely to desist

facing a negative attitude of others towards smoking;

cf. e.g., (Ganley and Rosario, 2013).

This paper is structured as follows: Section 2 in-

troduces concepts of social pressure, protective be-

haviors, and their implementation in ABSS to offer

practical requirements for the conceptual agent archi-

tecture combining PMT and social pressure for health

behaviors in general, which is introduced in Section

4. The implemented model is presented and analyzed

in Section 5 using calibration and sensitivity analysis

to demonstrate the concept for the use case of smok-

ing. Then, the implications of the experiments are

discussed as the results of the study. Finally, the ﬁnd-

ings are presented, focusing on potential avenues for

future research to expand upon this approach in Sec-

tion 6.

2 BACKGROUND: SOCIAL

PRESSURE AND

SELF-PROTECTION

We discuss concepts of social pressure as well as

PMT with a special focus on smoking as use case. To

situate this work in the context of current research,

computational models applying these psychological

mechanisms are presented.

Protective Behavior and Smoking. This paper ad-

dresses the use case of smoking, as a high propor-

tion of deaths, especially in the northern hemisphere,

is still attributable to the consequences of smoking

(e.g. in 2019, this ratio was about 17% in Germany

(Radtke, 2023)). One way of examining the possible

factors that lead to an individual’s decision to smoke

is by using PMT to categorize different inﬂuencing

factors.

Generally, those factors can be grouped into in-

trinsic and extrinsic factors. Intrinsic factors do not

Intrinsic Rewards

Extrinsic Rewards

Severity

Vulnerability

Threat Appraisal

Response Efficacy

Self-Efficacy

Response Cost

Coping Appraisal

Figure 1: Threat (a) and Coping (b) Appraisal according to

(Rogers, 1983).

just include factors such as their own attitude towards

smoking, but also more fundamental concepts such

as self-efﬁcacy, which denotes the belief that one can

successfully perform an action, e.g., quitting smoking

(Rogers, 1983). Meanwhile, extrinsic factors include

situational circumstances or the social environment’s

attitude towards a behavior (Xu et al., 2015).

According to PMT, individuals perform two cog-

nitive processes: threat appraisal and coping ap-

praisal (see Figure 1). These processes are triggered

by information from internal or external sources and

determine whether agents react adaptively (beneﬁ-

cial behavior) or maladaptively (harmful behavior)

(Rogers, 1983).

In the use case of smoking, threat appraisal de-

termines in how far adverse effects from harmful

activities are acknowledged. Subjective perceptions

of severity (e.g., how serious are the health conse-

quences of smoking in general?) and vulnerability

(how likely am I to suffer negative outcomes?) inﬂu-

ence a person’s willingness to react adaptively. High

intrinsic rewards in the form of pleasure or satisfac-

tion or extrinsic rewards, such as the feeling of be-

longing to a smoking group increase the probability

of maladaptive behavior, in which a person will at-

tempt to deny or downplay the risks of their choices.

Analogously, coping appraisal evaluates the rec-

ommended protective behavior (not smoking) and the

agent’s estimated ability to cope with and prevent

the threat from occurring. This is determined by

factors such as response efﬁcacy (i.e., does it really

bring health beneﬁts to me if I quit smoking?) and

self-efﬁcacy (i.e., will I be able to quit smoking?).

High efﬁcacy increases the likelihood of adaptive re-

sponses, which is a healthy behavior (Rogers, 1983).

Response costs, such as negative reactions from peers

due to not smoking, may hinder adaptive responses.

Protection motivation can be expressed as ei-

ther single or multiple actions, both one-time and

repeated, or even inaction. That is, coping with a

threat may require actively doing something, such as

quitting smoking, or refraining, such as not starting

to smoke (Rogers, 1983).

Social Pressure and Smoking. As mentioned previ-

ously, the attitude of peers impacts the appraisal of

different strategies in the form of pressure. Social

pressure can be deﬁned as perceived normative pres-

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

316

sure one receives from members of its immediate so-

cial environment (Tesser, 1980). When there is a con-

ﬂict between the social network’s opinion and the per-

sonal attitude towards a behavior, a person can expe-

rience cognitive dissonance, where different interests

conﬂict with each other (Festinger, 1957). As a way

to reduce this cognitive dissonance, a person typically

has two options. On the one hand, they may change

the comparison groups by seeking a new social en-

vironment with a preferably similar attitude. On the

other hand, they yield to social pressure by adjusting

their own disposition which can eventually lead to a

change in behavior(Festinger, 1957; Tesser, 1980).

3 MODELS OF SOCIAL

PRESSURE AND PROTECTIVE

BEHAVIOR

Models of social pressure are mainly expressed

through normative multiagent systems, where the be-

havior of agents is rewarded for norm-compliance and

sanctioned for deviation. An important distinction be-

tween types of models lies in the choice of mecha-

nism to express pressure. On the one hand, explicit

norms for behavioral regulation (see, e.g. (Castel-

franchi et al., 1998)) are often distributed by commu-

nication between agents. On the other hand, perceived

social pressure is more subtle, and is often tied to a

threshold to determine when agent decisions are im-

pacted by pressure (see, e.g. emotional autonomy).

This leads to agents acting according to perceived so-

cial pressure, even in case of conﬂict with the agent’s

own attitude (Dechesne et al., 2013).

Besides explicit and perceived norms, there are

also different ways rewards and punishments are in-

cluded into the model. Typically, there is a distinction

between direct sanctioning, such as monetary pun-

ishment, or indirect sanctioning in the form of emo-

tional penalties. In such cases, decreased trust in non-

conforming agents leads to reduced odds of coopera-

tion, such as the conclusion of contracts (Nardin et al.,

2016), in the future, sometimes even leading up to

the exclusion of agents from the social network (Per-

reau de Pinninck et al., 2010).

Norms, rewards and punishments can be commu-

nicated (Savarimuthu et al., 2009) or only suspected

and perceived through some kind of ﬁlter (Hashimoto

and Egashira, 2001) just as the subjectively perceived

pressure does not have to correspond to the actually

exerted pressure or attitudes of the social network.

Additionally, some individuals may have a stronger

impact on social pressure than others, e.g., based on

the strength of relationships or similarities to the per-

son being inﬂuenced (Savarimuthu et al., 2008). It

is important to note that changing one’s attitude does

not necessarily equal changing behavior, and will vary

between different models (Sheeran and Webb, 2016).

As a result of these different mechanisms, agents

can develop cognitive dissonance, as described in

Section 2, where personal preferences no longer align

with the environment. Typical dissolution strategies

include changing attitude (Dechesne et al., 2013) to-

wards adaption of other’s opinions or exclusion from

the social network (Perreau de Pinninck et al., 2010).

In the conceptual framework of PMT, the social

network plays an important role by presenting an en-

vironmental source that can affect decisions, e.g., in

the form of verbal persuasion or observational learn-

ing, such as in Hear et al. (Haer et al., 2016) in

the context of ﬂood risk management and the ef-

fect of different communication strategies and coun-

termeasures based on communication (Badham and

Gilbert, 2015). Mostly, PMT is used in connection

with climate-related decisions, e.g., with regard to

the implementation of preventive measures (see, e.g.,

(Wens et al., 2020; McEligot et al., 2019)) and the

general vulnerability to the consequences of climate

change (Kr

omker et al., 2008), or in health protective

behavior in the context of infectious diseases ((Ab-

dulkareem et al., 2018; Martin-Lapoirie et al., 2023)).

The impact of a social network on an agent’s deci-

sion is often not considered at all (see, e.g., (Kr

omker

et al., 2008; McEligot et al., 2019)) or is only regarded

as an additional source of information (Abdulkareem

et al., 2018). While works like (Haer et al., 2016;

Badham and Gilbert, 2015) often include the social

network as a factor in decision-making, both models

lack a clear deﬁnition and the effects of social pres-

sure with regard to the development of cognitive dis-

sonance as well as the agent’s responses to it.

Similarly to (Kurchyna et al., 2022), who exam-

ined another pair of cognitive and social theories for

the use case of heart disease, this work investigates

the interplay of cognitive dissonance, PMT, and the

realism of resulting system dynamics.

4 A CONCEPTUAL

AGENT-BASED MODEL OF

PROTECTION MOTIVATION

AND SOCIAL PRESSURE

Detached from the use case of smoking, this sec-

tion introduces the conceptual model of health pro-

tection motivation under consideration of social pres-

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure

317

sure. Agents experience social pressure from their en-

vironment and determine whether to perform a health-

related behavior (e.g., to smoke or not to smoke)

based on the concepts of PMT (see Figure 2).

The environment consists of a set of agents A. In-

spired by the approach presented by L

opez y L

opez et

al. (L

opez y Lop

ez et al., 2002), the initialization of

these agents is based on pre-deﬁned archetypes char-

acterized by a core desire. To reduce model com-

plexity for demonstration purposes, this model fo-

cuses on desires that we assume to be relevant for

most use cases: hedonism (desire

∈ [0, 1]), security

(desire

∈ [0, 1]), and conformity (desire

∈ [0, 1]) as

deﬁned by Schwartz (Schwartz, 1992). For example,

those who care more to have fun have a strong de-

sire for hedonism, those who are more careful when

it comes to health and safety have a strong desire for

security while those who favor social cohesion have

a strong desire for conformity. The closer a value

to 0, the weaker the desire an agent has while the

closer value to 1, the stronger the desire is. Agents

who desire pleasure for themselves the most (Heidari

et al., 2020) may neglect their need for safety (Mad-

dux and Rogers, 1983). Additionally, a desire to con-

form to expectations may interfere with the satisfac-

tion of either desire (Tesser, 1980). The archetype de-

termines the dominant desire of the agent, while the

other two desires are present to a lesser degree and

may inﬂuence the choice of an agent in edge cases.

There is no explicit categorization of agents within

the model, but is rather implied through their desires.

However, for the ease of communications, we refer

to the agents by their primary desire, the strongest

one. Hence, the agent’s environment can be deﬁned

as A = {Hedonists,Con f ormists, Sel f −Protectors}.

Each agent has a network N consisting of two dis-

joint groups: friends and locals. Friends (F ⊂ A) is

a broad category which includes persons the agent

has friendly relations with (like family members or

friends). This set’s size is ﬁxed to num f riends ∈ N.

Locals (L ⊂ A) contains people the agent encoun-

ters in its local environment. This may include col-

leagues, people on public transport or various ser-

vice providers. Belonging to this set is determined

by neighborhood in the 2D-space of the simulation.

While the social network of friends remains persistent

for the agent, random movement at each step leads

to a frequent variation of the local environment. For

more complex use cases, routines and the formation

of long-term relationships between agents would re-

place the random movements.

Over the course of the simulation, agents experi-

ence situations in which social pressure is present and

their actions are chosen based on PMT. This choice

of action is inﬂuenced not only by the network of the

agents, but also by their own attributes. Agents have

an initial attitude towards a health behavior, based on

their hedonism and security desires while conformity

desire determines how likely they will follow major-

ity opinion. This attitude att is mapped as a numer-

ical value to an interval [0, 1]. An agent’s opinion

towards a behaviour can range from strong aversion

(0) to strong favor (1), introducing nuance that al-

lows agents to choose a harmful behaviour, such as

smoking, even if they are not in complete favour of

it. This attitude is a proxy for the actual behavior and

it allows tracking of changes over time. Due to the

phenomenon of the intention-behavior gap (Sheeran

and Webb, 2016), the translation of attitudes into ac-

tions requires further examination beyond the scope

of a proof of concept. Another important attribute is

the general health status ∈ [0, 1] of the agent, inﬂu-

encing the way agents assess their own vulnerability

towards a threat. Here, the value 1 represents perfect

health, while values towards 0 represent increasingly

poor health such as chronic conditions and typical

health risk factors. Due to the delay between action

and effect, which makes it difﬁcult to observe the ac-

tual causal relationship, the health status is kept static

throughout the simulation.

In reality, factors such as economic status corre-

late with smoking and its frequency (Casetta et al.,

2017). However, under the assumption that needs

are a mediating factor between characteristics and

observable actions, such factors are implied through

both the agent’s archetype and their initial attitude

towards smoking. Depending on the purpose of the

study, especially explainability and the tailoring of

targeted interventions, implied characteristics may

not sufﬁce, and explicit implementation would be rec-

ommended in such a case.

The following paragraphs discuss the main com-

ponents of the proposed concept, i.e., the social pres-

sure component and the PMT components.

4.1 Social Pressure Component

Social pressure in this model is modeled by evaluating

the perceived attitudes against the agent’s own posi-

tion. In this case, there is no ﬁltering through subjec-

tive interpretation or other approaches from opinion

dynamics, and instead, the actual attitude of the agent

is being perceived as abstraction from how exactly

this social pressure is formed and exerted, focusing

on its effects.

Hence, agents consider the attitudes of re-

mote peers (att

∈ [0, 1]) as well as those of lo-

cals in their current physical environment (att

∈

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

318

Environment

Agent

Social

Pressure

Local

Attitude

Peer

Attitude

PMT

Extrinsic Reward

Intrinsic Reward

Response Efficacy

Self-Efficacy

Severity

Vulnerability

Response Cost

Threat

Appraisal

Coping

Appraisal

Maladaptive

Behavior

Adaptive

Behavior

Action

Change

Environment

Adapt

Attitude

Do Nothing

!𝛼

!𝛽

Health

Status

Figure 2: Agent-based model of PMT under social pressure.

[0, 1]). While att

takes the bidirectional inﬂuence

connection strength ∈ [0, 1] of agents on each other

into account, there is no such distinction for locals.

As a result, friends are weighted differently and can

have a larger or smaller inﬂuence compared to a local,

unlike some alternative approaches in which random

encounters always have less impact than friends (Li

et al., 2022), permitting situations of accute peer pres-

sure exerted by local contacts. The attribute reﬂecting

connection strength has an impact on the perceived

attitude of surrounding agents, and in certain cases

this can be perceived to be stronger than it actually

is and vice versa. For instance, att

is determined as

follows (Equation 1):

att

|F|

∑

i=1

att

∗ connection strength

|F|

(1)

Social pressure results from the sum of both aver-

age attitudes of locals and friends weighted according

to the number of connections of each type. The result-

ing social pressure ∈ [0, 1] is thus determined by the

distance between the agent’s own attitude att and the

values for att

and att

which are included proportion-

ally to the size of the whole network N (see Equation

2).

pressure = |att −

att

∗ |F| + att

∗ |L|

|N|

| (2)

4.2 Protection Motivation Components

As described in Section 4, agents can choose be-

tween two behavior modes: maladaptive or adap-

tive behavior. The decision for either is made based

on factors such as perceived social pressure, health

status and personal attitudes. To decide for one of

two behavior modes, the agent ﬁrst computes the

threat appraisal (appraisal

threat

) and coping appraisal

(appraisal

coping

). A behavior ∈ [m, a] is maladaptive

(m), if appraisal

threat

> appraisal

coping

and adaptive

(a) if appraisal

coping

> appraisal

threat

(see Figure 2,

Function α).

Appraisal

threat

is based on rewards that can be ob-

tained from performing a non-beneﬁcial behavior (in-

trinsic i and extrinsic e) in the presence of individu-

ally perceived severity (severity

∈ [0, 1]) and vulner-

bility (vul ∈ [0, 1]). Based on the theoretical equation

in Figure 1, threat appraisal of agents is calculated as

follows:

appraisal

threat

∗ ((reward

+ reward

)−

(severity

+ vul))

(3)

In this use case, reward

is the own attitude rep-

resenting the pleasure from smoking, while reward

is the social pressure towards smoking — conforming

to this expectation is perceived as reward. Severity

is moderated by individual perceptions (Badham and

Gilbert, 2014) and deﬁned as

severity

= severity

+ ((1 − desire

) ∗ (desire

− severity

))

(4)

The objective severity (severity

) is moderated by

desire

and desire

to express the individual levels of

anxiety (Badham and Gilbert, 2014). Vul is the as-

sessment of personal applicability of a threat, and thus

individual health (Rogers, 1983) inﬂuences its com-

putation as follows:

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure

319

vul = 1 − (health + (desire

− health) ∗ health) (5)

Appraisal

coping

examines the self-efﬁcacy

(efﬁcacy

) and the response efﬁcacy (efﬁcacy

) of an

action while evaluating the response cost (cost

res

) of

the action. If appraisal

threat

is high, quitting non-

beneﬁcial behavior is perceived as more effective,

which leads to an adaptive response. It is deﬁned as

follows:

appraisal

coping

efﬁcacy

+ efﬁcacy

− cost

res

(6)

where efﬁcacy

is deﬁned as

efﬁcacy

= att ∗

vul + severity

(7)

with perceived attitude (att p ) deﬁned as

att p =

att

∗ |F| + att

∗ |L|

|N|

(8)

and cost

res

deﬁned as

cost

res

att + (desire

∗ att p)

(9)

signifying both losing out on pleasure as well

as defying social norms. If a person has a highly

favourable attitude towards a non-beneﬁcial behavior,

desisting is difﬁcult. Likewise, a strong attitude (in

either direction) in the environment att p is hard to

resist, with both cases increasing response cost cost

res

(Huang and Wen, 2014).

Efﬁcacy

depends on both own attitude and the

opinions of the environment, as deﬁned in Equation

10.

efﬁcacy











1 − att if att > 0.5 ∧ att p ≤ 0.5

1 − att p if att ≤ 0.5 ∧ att p > 0.5

1 −

att+att p

else

(10)

Here, we set 0.5 as the transition point where an

agent’s attitude toward a certain behavior switches

from not favorable to favorable or vice versa.

In each step of the simulation, all agents make a

decision to act and impact on their immediate sur-

rounding agents and friends. Each agent calculates

the threat and coping appraisals and acts according to

their decision. There are three potential actions as de-

ﬁned in Figure 3, Function β, based on the cognitive

dissonance theory by Festinger (Festinger, 1957) ear-

lier introduced in Section 2. :

1. Create new consonant cognitions: An agent cuts

ties with friends that exert too much pressure on it

and ﬁnds new friends as well as moves on to an-

other location to switch their surrounding agents

(change environment). The user-deﬁned variable

contact termination ∈ [0, 1] determines at which

degree of difference the entire local network and

friends with large attitude differences are replaced

to preserve the set num f riends.

2. Change cognition: An agent bows to the social

pressure and adjusts its own desires and attitudes

towards the group mean (adapt attitude). By

adapting att, desire

and desire

, the agent re-

duces the distance between its internal states and

its social network. The amplitude of changes

is moderated by a user-determined change rate

(c r ∈ [0, 1]). As a result, cognitive dissonance is

reduced.

3. Reduce the importance of dissonant cognitions:

An agent does nothing if its own attitude is strong

enough to resist external inﬂuence or if pressure

is low due to similar attitudes (do nothing).

In Figure 3, both the actions change environment

as well as adapt attitude are chosen by the agent if

its attitude att is either speciﬁcally high or low (be-

low or above a determined threshold thd a and 1-

thd a) and the social network differs from that atti-

tude in a way, that it exceeds the value of c r. The

agent decides for a major change in its network only

if min stay ∈ [0, 10] steps have passed since previ-

ous changes to prevent a constant exchange of the

agent’s social network. If this is true, there are

two cases, in which the agent might adapt its net-

work structure: In the ﬁrst case (a), the agent has

a strong attitude towards the respective behavior and

where appraisal

threat

is stronger than appraisal

coping

The behavior thus is deﬁned as maladaptive, i.e., the

agent performs unhealthy behaviours. Here, agents

with a negative attitude towards the harmful behav-

ior are excluded from the agents sphere of inﬂuence

to reduce cognitive dissonance. The second case

(b) has a contrasting effect. The agent has a neg-

ative attitude towards the behavior. However, be-

cause the majority of its network has a positive atti-

tude and the agent’s appraisal

coping

is greater than its

appraisal

threat

, the agent shows an adaptive behavior

self: unhealthy behavior self: unhealthy behavior

others: unhealthy

behavior

others: unhealthy

behavior

others: healthy behavior

self: healthy behaviorself: healthy behavior

others: healthy behavior

Others Attitude

Self Attitude

Maladaptive:

change environment

Maladaptive:

adapt attitude

Adaptive:

adapt attitude

Adaptive:

change environment

(a)

(b)

Figure 3: Decision for maladaptive or adaptive action de-

pends on the level of attitude (healthy or unhelathy behav-

ior) and the social pressure of the network (corresponds to

Figure 2, Function β).

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

320

and excludes friends and locals that might inﬂuence it

towards the harmful behavior. Adapt attitude makes

a similar distinction. If the agent’s att is rather favor-

able towards smoking and the social network exerts

much pressure while appraisal

coping

is stronger than

the appraisal

threat

, leading to the adaptive behavior,

the agent adapts its attitude and desires to reduce cog-

nitive dissonance to its network. In the second case,

the agent’s appraisal

threat

is higher, which leads to

the maladaptive behavior. The agent adapts its att to-

wards unhealthy behavior to be more positive-minded

towards it. In contrast to the option of adapting the

network of friends and locals, adapt attitude can be

selected in each point in time. If none of these cases

applies, the agent chooses do nothing.

5 SIMULATING SOCIAL

PRESSURE TOWARDS

SMOKING WITHIN

PROTECTION MOTIVATION

THEORY

The implementation of the concept is demonstrated

with smoking as use case. Over the course of the sim-

ulation, with steps representing an abstract concept

of ‘situations’, agents experience social pressure. For

instance, people feel compelled by their peers to ei-

ther smoke or refrain from doing so, depending on

the group’s attitude. The act of smoking itself is not

modelled explicitly but is expressed via the attitude

as proxy. The model was implemented using NetL-

ogo 6.3.0. For the explanation of the model validation

process, its in- and outputs are described along with

the calibration process and exploration using sensitiv-

ity analysis.

5.1 Experiment Design and Calibration

The primary goal of this experiment is not a credible

simulation of smoking (and its cessation), but rather

to provide a use case in which this combination of

social pressure and PMT is explored. The analyzes

and experiments are performed with this goal in mind.

The implemented model includes a set of param-

eters which can be conﬁgured by users, listed and de-

scribed in Table 1. These parameters mostly deal with

the composition of agent archetypes and the rate at

which its state changes. To examine the model be-

havior, three types of output variables were observed:

• Smoking behavior: The number of habit-

ual smokers (att ≥ 0.75), casual smokers

(0.25 < att < 0.75) and convinced non-smokers

( 0.25 ≥ att) in the population

• Desires: The average desire for conformity, secu-

rity and hedonism in the population, which con-

veys the overall population trend

• Actions: The number of times the ac-

tions do nothing, change environment and

adapt attitude chosen by the agents

A primary objective of this model’s calibration is

to examine whether there exists a parameter combina-

tion that results in a stable behavior without a severe

stagnation. While habitual behaviors, such as smok-

ing, have long delays between action and effect, other

behaviors may show much higher degrees of change

and more dynamic situations.

To establish a baseline parameter conﬁguration

at which real-world observations are imitated by the

model, a calibration using BehaviorSearch, a tool in-

tegrated in NetLogo, was performed. In this study,

a distribution of 29% average smokers in Germany

was expected to be achieved, based on available em-

pirical data (Starker et al., 2022). In prior trial experi-

ments, the model showed that agents tend to build sta-

ble, homogenous groups after approximately 30 steps,

each representing a situation in which social pressure

might have been experienced. Thus, the goal of cali-

bration is to minimize the deviation of the number of

smokers from the target value at the end of the simu-

lation, once equilibrium has been reached. Addition-

ally, the composition of agent types should add up to

100%. Further, a system in which the initial number

of smokers and non-smokers matches statistical data

and starts in a stable equilibrium is not conducive to

the creation of a realistic system, which is why static

behavior (do nothing) is among the observations to

be minimized. As a result, Equation 5.1 was used as

minimization function for BehaviorSearch, with the

weights chosen to adjust the priority of the different

conditions according to their importance.

ﬁtness =





Smokers

− 0.3



∗ 1000



(

∑

do nothing ∗ 0.1)+ (11)

((

Conformists

Hedonists

Security

) − 1) ∗ 10

The results of the calibration are included in the

last column in table 1. To better understand how the

parameters interact with each other, however, an in-

depth sensitivity analysis is required. Still, a prior

calibration as a ﬁrst step towards model veriﬁcation

is necessary to ensure that a model can achieve valid

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure

321

Table 1: Parameters, ranges and the recommended default

values based on calibration. Population = 100 was set as

ﬁxed parameter to restrict the search space.

Parameter Description Range Value

n people Number of agents int 100

Conformists % of conformists 0 - 1.0 0.23

Hedonists % of hedonists 0 - 1.0 0.64

Security % of health-aware 0 - 1.0 0.18

pressure mod. Strength of perceived pressure 0 - 1.5 1.25

contact term.

Allowed difference between

attitude and perceived pressure

before changing environment

0 - 1 0.4

min stay

Min. number of actions in new

env. before changing is possible

0 - 10 3

moving distance move forward set number of units 0 - 10 2

number friends Number of peers each agent has 0 - 10 2

change rate Rate of change of attitude 0 - 0.1 0.01

results. In this case, while competing solutions exist,

one possible conﬁguration observes a dominant pro-

portion of hedonistic agents, with security-oriented

people and conformists in similar numbers. The pres-

sure modiﬁer is slightly above the intended default

of 1, amplifying the intended effect. At 0.4, agents

are rather tolerant to differences in opinions, and will

only change their environment in the case of severe

conﬂict of attitudes. With small spatial movements

and a small social circle of two close friends as the

default, agents are thus more likely to hold on to

their environment than to change it, further conﬁrmed

by the low change rate which favours slow, gradual

changes over abrupt shifts.

When testing viable parameter combinations, it

was observed that the system typically reaches a sta-

ble state after approximately 30 simulation steps, with

agents forming homogenous groups similar to various

segregation models (Goles et al., 2011).

5.2 Simulation Results and Discussion

Once the feasibility of realistic model behavior was

conﬁrmed using calibration, a sensitivity analysis

was performed based on the example provided by

Jaxa-Rozen and Kwakkel (Jaxa-Rozen and Kwakkel,

2018). To show how the different design decisions

in the model impact the overall behavior, we an-

alyze the ﬁrst- and second-order sensitivity indices

of the parameters in regard to the observation vari-

ables. A major ﬁnding was the initial distribution of

character types (Hedonists, Con f ormists and Sel f −

Protectors) dominating the model behavior, which

aligns with empirical data that conﬁrms most will re-

tain their initial attitude, e.g. those who are smok-

ers and hedonists at the same time will continue to

smoke, while those who are non-smokers and health-

conscious are unlikely to start smoking (Høie et al.,

2010).

To analyze the effects of other variables on the

different observations, the calibrated values for the

population distribution were used and the input vari-

able was removed from the sensitivity analysis. The

change rate (c r), describing the amplitude of changes

agents make and how strongly they perceive the pres-

sure, is the major determinant for model behavior, ac-

counting for approximately 70% of the variance in

most observations. The pressure modiﬁer, adjusting

the strength of pressure agents experience, offered in-

sight into the role that peer pressure plays in the appli-

cation of PMT in this model. Figure 4 summarizes the

ﬁndings regarding the impact of the pressure modiﬁer,

and thus the social pressure altogether, with respect to

the different observation variables.

Social pressure contributes towards the distribu-

tion of behavioral types in the population. The low

inﬂuence on the number of habitual smokers, in com-

bination with the number of non-smokers and casual

smokers being more sensitive to social pressure, in-

dicates that people with moderate to low attitudes are

more strongly affected by external inﬂuences. In that

regard, the population shows a strong leaning towards

persuading agents to give up smoking. Strong interac-

tions between the pressure modiﬁer and other param-

eters, such as the number of friends and contact ter-

mination rates, indicate that social pressure is a major

contributor towards reduction of smoking, contrary to

the initial expectation that social pressure might be a

driving factor in previously non-smoking agents start-

ing to smoke. This gradual reduction of smoking at-

titude does indeed align with the global trend of a de-

creasing number of smokers (WHO, 2021).

Moreover, social pressure is strongly related to

agents deciding to be inactive — its absence or pres-

ence is the main explanator for the frequency of in-

action. While the desire for conformity is barely af-

fected by pressure, the average desire for hedonism

and security are dominated by the pressure modiﬁer

— without social pressure, agents will retain their ini-

tial beliefs. Assuming that agents will, at the begin-

ning of the simulation, already perform the behavior

that fulﬁlls their desires, they will have no incentive

to change their attitudes or perform any adaptive or

maladaptive behaviors due to the lack of external in-

ﬂuences. In this regard, this study comes to the same

conclusions as (Kurchyna et al., 2022), which noted

stagnation in a system with purely cognitive processes

that do not receive any external inputs in the form of

social inﬂuence.

Such ﬁndings indicate that the combination of

external/social and internal/cognitive components is

well-suited for autonomous systems of continuous

simulation without clearly deﬁned end states. In this

example, agents pursue the continuous goal of a bal-

anced needs satisfaction in a mixed social environ-

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

322

Figure 4: First (S1) and total-order (ST) indices for pressure sensitivity towards different observations.

ment that may hinder or encourage their quest for

homeostasis. Due to the mechanistic design of PMT

and the two-layered approach to contact networks, the

ideas of this model can also be transferred to other ap-

plication areas in which agents act under uncertainty

and conﬂicting goals and options. A good example

for such scenarios is human-AI-interaction and assis-

tance systems, which need to be able to anticipate and

consider inconsistent goals and seemingly irrational

user behavior.

6 CONCLUSION

The study of the decision-making processes leading to

people’s behavior and large-scale emergent phenom-

ena is an important research topic in ABSS. To exam-

ine how psychological theories can be used to model

decision-making processes of agents in continuous

simulations, an implementation of PMT and social

pressure for the use case of smoking was presented.

While PMT provides the cognitive process leading

to decisions within agents, the concepts of cognitive

dissonance and social pressure ﬁll the model with

life. This proof-of-concept implementation shows

that without social pressure, cognitive concepts risk a

vacuum in which agents lose incentive to act, consis-

tent with previous ﬁndings in similar types of studies;

cf. e.g., (Kurchyna et al., 2022).

In the broader context of simulations in which

agents interact under uncertainty and with potentially

conﬂicting goals, concepts such as PMT may be well

suited to include a variety of factors into decision-

making without having the large scope of alterna-

tive approaches such as the theory of planned behav-

ior (Ajzen, 1991), which may be more suited for use

cases which require more complex behaviors and de-

liberate planning. The insights from the model will

enable AI assistance systems to support individuals

in their decision-making processes and motivate them

to pursue more desirable actions when faced with con-

ﬂicting objectives and perspectives on each action cat-

egory. Additionally, this use case demonstrated two-

layered networks of contacts exerting potentially con-

tradictory inﬂuences. The insights provided by this

model and its evaluation are promising in regard to

various dimensions of future works.

Such future work therefore involves validating the

model using empirical data, as well as testing its

transferability to other health-related application ar-

eas. Furthermore, the authors plan on improving the

model with the aim of adjustments such as more in-

dividualized behavior. For example, in the current

model, adjustments in attitude and social network are

determined by the change rate and the contact termi-

nation threshhold. However, these variables are set

to a global static value. According to psychologi-

cal theory, a person’s emotional autonomy deﬁnes an

individual bound for each, which determines when

perceived social pressure will contribute to behavior

adjustment (Savarimuthu et al., 2008). Additionally,

the similarity of the person to its surrounding net-

work may play a role in the adaptation of behavior.

If the person does not identify enough with the group,

the social pressure exerted will have no inﬂuence on

the person’s behavior (Terry and Hogg, 1996). In the

same vein, a more elaborate model of social pressure,

which includes detailed perception of other’s opin-

ions, would provide beneﬁts to the realistic represen-

tation of the agent’s decision making process. The

importance of other characteristics (such as demo-

graphic variables) should be examined as additional

variables, since, e.g., a person’s age may contribute

to willingness to change and thus adopt protective be-

haviors (Badham and Gilbert, 2015).

In general, the authors intend to extend and adapt

the agent architecture based on the BDI approach

(Rao and Georgeff, 1995) to achieve a more differen-

tiated model of individual desires as well as complex

decision-making, including planning steps towards

a behavior and the formation of habits (see, e.g.,

(Kurchyna et al., 2022)). Finally, a future step ex-

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure

323

tends the model towards the possibility of testing dif-

ferent intervention strategies, for example, different

communication strategies (c.f. (Haer et al., 2016)),

in order to lead agents to beneﬁcial behavior such as

quitting smoking.

ACKNOWLEDGEMENTS

This submission is a result of the work in the con-

text of the SEMSAI Project, funded by the German

Federal Ministry of Education and Research under the

grant number 031L0295A.

REFERENCES

Abdulkareem, S. A., Augustijn, E.-W., Mustafa, Y. T., and

Filatova, T. (2018). Intelligent judgements over health

risks in a spatial agent-based model. International

journal of health geographics, 17(1):1–19.

Ajzen, I. (1991). The theory of planned behavior. In Orga-

nizational behavior and human decision processes.

Badham, J. and Gilbert, N. (2014). Personal protective

behaviour during an epidemic. In Social Simulation

Conference-SSC 2014.

Badham, J. and Gilbert, N. (2015). Tell me design: Protec-

tive behaviour during an epidemic. CRESS.

Berndt, J. O., Rodermund, S. C., and Timm, I. J. (2018). So-

cial contagion of fertility: an agent-based simulation

study. In 2018 Winter Simulation Conference. IEEE.

Casetta, B., Videla, A. J., Bardach, A., Morello, P., Soto,

N., Lee, K., Camacho, P. A., Hermoza Moquillaza,

R. V., and Ciapponi, A. (2017). Association between

cigarette smoking prevalence and income level: a sys-

tematic review and meta-analysis. Nicotine & To-

bacco Research, 19(12):1401–1407.

Castelfranchi, C., Conte, R., Paolucci, M., et al. (1998).

Normative reputation and the costs of compliance.

Journal of Artiﬁcial Societies and Social Simulation,

1(3):3.

Davidsson, P. (2002). Agent based social simulation: A

computer science view. Journal of artiﬁcial societies

and social simulation, 5(1).

Dechesne, F., Di Tosto, G., Dignum, V., and Dignum, F.

(2013). No smoking here: values, norms and culture

in multi-agent systems. In AI and Law.

Festinger, L. (1957). A theory of cognitive dissonance.

Stanford University Press.

Floyd, D. L., Prentice-Dunn, S., and Rogers, R. W. (2000).

A meta-analysis of research on protection motivation

theory. In Journal of applied social psychology.

Ganley, B. J. and Rosario, D. I. (2013). The smoking atti-

tudes, knowledge, intent, and behaviors of adolescents

and young adults: Implications for nursing practice.

Journal of Nursing Education and Practice, 3(1):40.

Goles, D. N., Goles, E., and Rica, S. (2011). Dynamics

and complexity of the schelling segregation model. In

Phys. Rev. E. 83. American Physical Society.

Haer, T., Botzen, W., and Aerts, J. (2016). The effective-

ness of ﬂood risk communication strategies and the

inﬂuence of social networks—insights from an agent-

based model. In Environmental Science & Policy.

Hashimoto, T. and Egashira, S. (2001). Formation of social

norms in communicating agents with cognitive frame-

works. Journal of Systems Science and Complexity.

Hedayati, S., Damghanian, H., Farhadinejad, M., and Rast-

gar, A. A. (2023). Meta-analysis on application of

protection motivation theory in preventive behaviors

against covid-19. International Journal of Disaster

Risk Reduction, page 103758.

Heidari, S., Jensen, M., and Dignum, F. (2020). Simulations

with values. In Advances in Social Simulation: Look-

ing in the Mirror. Springer International Publishing.

Huang, C. Y. and Wen, T. H. (2014). A novel private at-

titude and public opinion dynamics model for simu-

lating pluralistic ignorance and minority inﬂuence. In

Journal of Artiﬁcial Societies and Social Simulation.

Høie, M., J., M. I., and Rise, J. (2010). An extended version

of the theory of planned behavour: Prediction of inten-

tions to quit smoking using past behaviour as moder-

ator. In Addiction Research & Theory.

Jaxa-Rozen, M. and Kwakkel, J. H. (2018). Pynetlogo:

Linking netlogo with python. In Journal of Artiﬁcial

Societies and Social Simulation.

Jovanovic, M., Mitrov, G., Zdravevski, E., Lameski, P.,

Colantonio, S., Kampel, M., Tellioglu, H., and Florez-

Revuelta, F. (2022). Ambient assisted living: scoping

review of artiﬁcial intelligence models, domains, tech-

nology, and concerns. Journal of Medical Internet Re-

search, 24(11):e36553.

Kothe, E. J., Ling, M., North, M., Klas, A., Mullan, B. A.,

and Novoradovskaya, L. (2019). Protection motiva-

tion theory and pro-environmental behaviour: A sys-

tematic mapping review. Australian Journal of Psy-

chology, 71(4):411–432.

omker, D., Eierdanz, F., and Stolberg, A. (2008). Who

is susceptible and why? an agent-based approach to

assessing vulnerability to drought. Regional Environ-

mental Change, 8:173–185.

Kurchyna, V., Rodermund, S., Berndt, J. O., Spaderna, H.,

and Timm, I. J. (2022). Health and habit: An agent-

based approach. In KI 2022: Advances in Artiﬁcial

Intelligence: 45th German Conference on AI Proceed-

ings. Springer-Verlag Berlin, Heidelberg.

Li, L., Wang, S., and Lin, Y. (2022). The casual effect of

relational mobility on integration of social networks:

An agent-based modeling approach. In Current Psy-

chology.

opez y Lop

ez, F., Luck, M., and d’Inverno, M. (2002).

Constraining autonomy through norms. In Proceed-

ings of the ﬁrst international joint conference on Au-

tonomous agents and multiagent systems: part 2.

Macy, M. W. and Willer, R. (2002). From factors to actors:

Computational sociology and agent-based modeling.

In Annial review of sociology.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

324

Maddux, J. and Rogers, R. (1983). Protection motivation

and self-efﬁcacy: A revised theory of fear appeals and

attitude change. In Experimental Social Psychology.

Martin-Lapoirie, D., D’onofrio, A., McColl, K., and Raude,

J. (2023). Testing a simple and frugal model of health

protective behaviour in epidemic times. Epidemics,

42:100658.

McEligot, K., Brouse, P., and Crooks, A. (2019). Sea bright,

new jersey reconstructed: Agent-based protection the-

ory model responses to hurricane sandy. In 2019

Winter Simulation Conference (WSC), pages 251–262.

IEEE.

Mou, J., Cohen, J. F., Bhattacherjee, A., and Kim, J. (2022).

A test of protection motivation theory in the infor-

mation security literature: A meta-analytic structural

equation modeling approach. Journal of the Associa-

tion for Information Systems, 23(1):196–236.

Nardin, L. G., Balke-Visser, T., Ajmeri, N., Kalia, A. K.,

Sichman, J. S., and Singh, M. P. (2016). Classifying

sanctions and designing a conceptual sanctioning pro-

cess model for socio-technical systems. The Knowl-

edge Engineering Review, 31(2):142–166.

Perreau de Pinninck, A., Sierra, C., and Schorlemmer, M.

(2010). A multiagent network for peer norm enforce-

ment. In AAMAS 2010.

Qian, K., Zhang, Z., Yamamoto, Y., and Schuller, B. W.

(2021). Artiﬁcial intelligence internet of things for the

elderly: From assisted living to health-care monitor-

ing. IEEE Signal Processing Magazine, 38(4):78–88.

Radtke, R. (2022, accessed on 05 March 2023).

Anteil rauchbedingter Todesf

alle in aus-

gew

ahlten L

andern im Vergleich der Jahre

1990 und 2019. In https://de.statista.com/

statistik/daten/ studie/1172056/ \umfrage/

anteil-rauchbedingter-todesfaelle-nach-laendern/ .

Statista.

Rao, A. S. and Georgeff, M. P. (1995). Bdi agents: from the-

ory to practice. In International Conference on Multi-

Agent Systems.

Rogers, R. (1983). Cognitive and psychological processes

in fear appeals and attitude change: A revised theory

of protection motivation. In Social psychophysiology:

A sourcebook.

Savarimuthu, B., Craneﬁeld, S., Purvis, M., and Purvis, M.

(2008). Role model based mechanism for norm emer-

gence in artiﬁcial agent societies. In Coordination,

Organizations, Institutions, and Norms in Agent Sys-

tems. Springer Berlin Heidelberg.

Savarimuthu, B. T. R., Purvis, M., Purvis, M., and Crane-

ﬁeld, S. (2009). Social norm emergence in virtual

agent societies. In Declarative Agent Languages and

Technologies VI: 6th International Workshop, DALT

2008, Estoril, Portugal, May 12, 2008, Revised Se-

lected and Invited Papers 6, pages 18–28. Springer.

Schwartz, S. H. (1992). Universals in the content and struc-

ture of values: Theoretical advances and empirical

tests in 20 countries. In Advances in experimental so-

cial psychology. Academic Press.

Sheeran, P. and Webb, T. L. (2016). The intention–behavior

gap. Social and personality psychology compass,

10(9):503–518.

Smith, E. and Conrey, F. (2007). Agent-based modeling:

A new approach for theory building in social psychol-

ogy. In Personality and social psychology review.

Squazzoni, F., Jager, W., and Edmonds, B. (2014). Social

simulation in the social sciences: A brief overview. In

Social Science Computer Review.

Starker, A., Kuhnert, R., and Hoebel, J., a. R. A. (2022).

Rauchverhalten und Passivrauchbelastung Erwach-

sener–Ergebnisse . In GEDA 2019/2020-EHIS.

Tapp, L., Kurchyna, V., Nogatz, F., Berndt, J., and Timm,

I. J. (2022). School’s out? simulating schooling strate-

gies during covid-19. In Autonomous Agents and Mul-

tiagent Systems. Best and Visionary Papers: AAMAS

2022 Workshops. Springer.

Terry, D. and Hogg, M. (1996). Group norms and the

attitude-behavior relationship: A role for group iden-

tiﬁcation. In Personality and Social Psychology Bul-

letin.

Tesser, A. (1980). When individual dispositions and social

pressure conﬂict: A catastrophe. In Human Relations.

Wagner, N. and Agrawal, V. (2014). An agent-based sim-

ulation system for concert venue crowd evacuation

modeling in the presence of a ﬁre disaster. In Expert

Systems with Applications.

Wens, M., Veldkamp, T. I., Mwangi, M., Johnson, J. M.,

Lasage, R., Haer, T., and Aerts, J. C. (2020). Sim-

ulating small-scale agricultural adaptation decisions

in response to drought risk: an empirical agent-based

model for semi-arid kenya. Frontiers in water, 2:15.

WHO (2021). Who global report on trends in prevalence

of tobacco use 2000-2025. In Geneva: World Health

Organization. Licence: CC BY-NC-SA 3.0 IGO.

Xu, X., Leung, D., Li, B., Wang, P., and Zhao, Y. (2015).

Smoking-related knowledge, attitude, social pressure,

and environmental constraints among new undergrad-

uates in chongqing, china. In International journal of

environmental research and public health.

APPENDIX

The model’s NetLogo code can be viewed and

executed under http://modelingcommons.org/browse/

one model/7256.

Seeing Through the Smoke: An Agent Architecture for Representing Health Protection Motivation Under Social Pressure

325