Emotion Contagion among Affective Agents

Issues and Discussion

Mara Pudane

, Michael A.Radin

and Bernard Brooks

Department of Artificial Intelligence and Systems Engineering, Riga Technical University, Kalku 1, Riga LV–1658, Latvia

Rochester Institute of Technology, Rochester, New York 14623, U.S.A.

Keywords: Agent based Simulation, Affective Agents, Emotion Contagion.

Abstract: Emotional contagion is a mechanism which results in transferring an emotion from one person to another; in

fact, it has been proven to be one of the key factors in successful crowd managing and preserving an

amiable working atmosphere. However, believable simulation of a group of people with social links

amongst them requires not only complex interaction models but also complex internal models as well. This

paper describes the progress of an on-going research that explores and simulates various types of emotions

while they are being transmitted amongst affective intelligent agents that are connected in simulated social

structure. The internal mechanism of agents is based on affective agent architectures while the contagion

and its rules are being modelled by using tools from graph theory.

1 INTRODUCTION

Emotions are one of the currently trending topics in

various research fields, i.e., psychology, sociology

as well as computer science. Previously there has

been a long discussion on whether emotions are

good, harmful or if they are needed at all; however,

lately the consensus amongst the scientists is that

emotions are a very crucial part of rational thinking.

It is considered that emotions provide humans with

necessary adaptation mechanisms as well as allows

to make extremely complex decisions with very

limited resources.

In the light of these findings, series of research

on various internal and external emotion-related

mechanisms have emerged. Especially, in 1997, the

affective computing was defined (Picard, 1997). It

explores how artificial units (such as intelligent

agent) and systems can benefit from implementation

of affect (broad term for variety of emotion related

concepts - emotion, mood, personality etc.) on

various complexity levels. Affective computing

mostly deals with system's internal structure as well

as emotion acquisition and expression thus enabling

creation of systems that behave, think or appear to

think similarly as a human. However, apart from

that, another direction is external mechanisms

including emotional interactions among emotional

units.

One of such mechanisms is an emotion

contagion: the ability to "catch" other people’s

emotions using the body as emotion elicitor

(Hatfield et al., 1993). Emotion contagion is based

on bodily feeling theories (see i.e., Damasio, 1994)

that proposes emotions that are generated by

responding to certain body poses and mimicry.

A proper emotion contagion model would be

beneficial not only for practical use but also for

research purposes - e.g., it could lead to better

understanding of how various higher level affects,

such as mood, impact emotional states of group

individuals, as well as understanding how the

atmosphere of human group develops.

Agent based models are ideal for expressing the

knowledge of subject experts in mathematical model

and have already been used to model contagion

(Bosse et al., 2015), as well as rumour flow on

networks (Brooks et al., 2013).

This paper will present preliminary stages, ideas

and observations of the research that aims at fully

simulating emotional contagion mechanisms.

Section 2 presents related work; Section 3 focuses

on affective basis of emotion research. Section 4

describes microstructure – i.e., the internal structure

of agents, while Section 5 focuses on details of

macrostructure – topography of network and transfer

rules. Section 6 provides some final notes as well as

conclusions.

328

Pudane M., Radin M. and Brooks B.

Emotion Contagion among Affective Agents - Issues and Discussion.

DOI: 10.5220/0006252603280334

In Proceedings of the 9th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2017), pages 328-334

ISBN: 978-989-758-219-6

2 RELATED WORK

A certain amount of attention has been turned to

emotion contagion within the past several years as

human resource management and research of crowds

have become increasingly important. Emotional

contagion can be separated into two various

mechanisms, namely, primitive emotional contagion

which includes mimicking of emotions

subconsciously and secondary emotional contagion

that involves cognition (Barsade, 2002). The

cognition involvement into emotional contagion

makes it more complex as it includes social relations

and empathy.

The primitive emotion contagion directly refers

to group of people that do not have strongly defined

structure, i.e., crowd. For this reason, there are some

developments that are focused on emotion contagion

in crowd for various purposes, i.e., researching how

to prevent riots (Bassi, 2006), or to create a learning

system - i.e., teaching soldiers how to prevent crowd

from becoming unpredictable and uncontrollable by

simulation (Aydt et al., 2011).

It has also been proven that an emotional

contagion refers to a web environment. Thus this

phenomenon has been researched in large scale

social networks. Controversial research was

conducted by analysing an emotional level in

Facebook news feeds (Kramer et al., 2014).

Similarly, (AlSagri and Ykhlef, 2016) also focuses

on emotion contagion research in online

communities; their focus is to reduce contagion of

negative emotions.

Despite conducting much research regarding

social networks, emotion contagion greatly differs

depending on how an emotion is expressed. Thus,

the contagion method that works successfully for

emotions that are expressed and acquired via written

text will substantially differ from the one that works

with emotions acquired from facial or vocal

expressions. Also, the emotion contagion for the

group with practically no structure (i.e., crowd) will

differ from a social group where people have links

with various weights amongst them. Thus, most of

these models cannot be directly applied for

modelling a group of people who know each other.

There have also been general models of emotion

contagion, i.e., (Bosse et al., 2015). This particular

research includes a mathematical model of the spiral

effect (i.e., the property of emotion amplifying) in

group dynamics. The model of separate group

members is also very elaborate and includes not only

some personality factors of group members (e.g.,

expressiveness) but also multi-weighted

relationships amongst the people (Bosse et al.,

2015). The model also provides mathematical

analysis of an emotional contagion. As the authors

themselves notice though, the model focuses on only

one type of emotion.

Another general framework was developed by

(Bispo and Paiva, 2009). It models five emotions:

fear, sadness, joy, anger and love. The choice of

these emotions comes from theory of Emotional

Contagion Scale (Doherty, 1997). The model

includes expressiveness and energy of various

emotions as well as considering various personalities

as a minimum and maximum variable for each

emotion.

Although some of the models that allow

simulating emotional contagion are elaborate, they

still lack internal mechanisms of agents that would

enhance the system’s believability.

3 AFFECTIVE BASIS

There are several mechanisms of the system that

must be considered and pertinent questions that must

be answered from the perspective of psychology and

sociology to simulate the emotional contagion.

 Personality of involved people. How to

properly represent the personality of people?

What personality traits impact the emotional

contagion?

 Contagion of various emotions. What

emotions to model? How various emotions

would interact with rational agent? What

should be the internal mechanism of agent to

consider secondary emotional contagion?

 Generation of believable network structure.

What should be the network structure to be

considered believable?

 Emotion contagion patterns. What are

mechanisms or patterns that are involved into

emotion contagion in a group?

3.1 Personality Impact

The personality of involved people influences the

emotional contagion in general as well as emotional

intensity level of one individual. Personality in

primitive emotion contagion impacts two things,

namely, how fast does the emotion spread (i.e., the

expressiveness of emotion) and how deep is the

impact of an emotional contagion (i.e.,

susceptibility) (Barsade, 2002). Although some of

the related works consider personality as

expressiveness and susceptibility variable (e.g.,

Emotion Contagion among Affective Agents - Issues and Discussion

329

Bosse et al., 2009), these models do not explain

what types of personalities have high or low

susceptibility, as well as what kind of personality

traits impact these factors. The Big Five model is

currently the most used and best verified model that

allows modelling personality as a combination of

five traits: Openness, Conscientiousness,

Extroversion, Agreeableness, and Neuroticism

(McCrae and Costa, 2003). The usage of such

psychologically grounded model would enable better

exploration of group of humans, as there are

researches that have focused, e.g., on how these

traits impact interaction (Pease and Lewis, 2015).

3.2 Types of Emotions

Another vital question to consider is what types of

emotions to model. This question has been

addressed in affective computing literature as it is

one of the basic questions along with how to

combine various types of emotion (Hudlicka, 2015).

In case of primitive emotional contagion, the process

depends on how the emotion is expressed and

whether the receiver of emotion has the same bodily

feeling. For this reason, it makes sense to model 6

basic emotions identified by (Ekman, 1992): fear,

surprise, anger, joy, sadness and disgust. These

emotions have been found to be universal amongst

the various ages and cultures (Ekman, 1992), thus

triggering same emotional responses in various

people. Basic emotions also correspond to Damasio's

primary emotions (Damasio, 1994).

The secondary emotional contagion, on the other

hand, involves not only imitation mechanisms but

also social and cognitive evaluation of other’s

emotion (Barsade, 2002). One option for modelling

secondary emotional contagion would be to model

same emotions as for primitive contagion, however

it does not provide intended diversity of the model.

To simulate such mechanisms, appropriate

psychological background is required. The view on

emotions separates theories of emotion into three

groups: categorical theories (such as Ekman’s theory

of six basic emotions), dimensional theories and

appraisal theories (Lisetti and Hudlicka, 2015).

Appraisal theories reflect agent’s cognitive

evaluation of the world state in terms of its goals,

beliefs, behavioural capabilities and available

resources (Lisetti and Hudlicka, 2015). One of the

most used models (Lisetti and Hudlicka, 2015) is

OCC model (Ortony et al., 1988) which groups

emotions into three categories one of them being

feelings towards other agent’s actions. OCC allows

to describe the emotional relations amongst the

agents, however, these emotions are related to

agent’s own emotional state thus is not suitable for

direct representation of an emotional contagion. The

appraisal theories, however, could be used for an

agent to deduce what the other is feeling.

Dimensional theories are theories that explain

emotions as a value in multi-dimensional space. One

of the most used examples of that is PAD space

(Russel and Mehrabian, 1977) that models emotions

in Pleasure-Arousal-Dominance space. The research

has been conducted based on PAD model the result

of which was PAD values for 151 emotions (Russel

and Mehrabian, 1977). As there is such

formalisation available for this model, theoretically

any of these emotions could be modelled, although it

would make model unnecessary complex. Russel

and Mehrabian offer eight mood types depending on

positive or negative values of PAD. These moods

are used e.g. in (Gebhard, 2005) to create believable

virtual agent. Modelling these moods would make

an agent more flexible and would allow to orient

itself in a large space of possible PAD value

combinations.

3.3 Structure of Network

As mentioned previously, this research focuses on

real-life person group. Although such a graph can be

obtained by analysing an existing structure (e.g.,

research lab), we have chosen to generate graphs

artificially, elaborated more in section 5. By

generating artificial graphs that replicate real graphs

we can repeat the Monte Carlo experiments and our

results will not be only a function of the few real

graphs’ topology. There are some things to be

considered, regarding relationships amongst the

people. The emotional contagion is stronger amongst

people that are more connected (Barsade, 2002), the

weight can either represent a relationship, or the

frequency of the contact during the average day.

The weight of a link would impact, first, whether

agent gets emotion at all. Secondly, it would impact

intensity of experienced emotion.

3.4 Mechanisms of Emotional

Contagion

Although primitive emotional contagion happens in

direct interaction amongst two people, there are

some more complex mechanisms associated with

emotional contagion modelling. One of such

mechanisms is so called spiral effect – when mood

of entire group becomes worse or better than that of

one individual (Barsade, 2002). In (Bosse et al.,

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

330

2015) it is modelled as depending on individuals,

i.e., each unit in a model has tendency to turn

emotions either up or down. These tendencies can be

associated with personality in terms of OCEAN

model to create more believable models that can also

be analysed from point of view of sociology.

We do not consider patterns where an agent is

feeling opposite emotions than the other (e.g.,

gloating if another agent is angry or sad). Although

such patterns appear in real-life networks as well as

in emotion related literature (e.g. in OCC), it is

unclear whether such “negative contagion” can be

modelled with the same mechanisms or can be

considered as a contagion at all.

4 MICRO LEVEL

Based on issues discussed before, this and next

chapter is focused on micro and macro levels of the

system (Wooldridge, 2009). The micro level

concerns one unit’s, in this case, the agent’s,

architecture and functionality. The macro level

focuses on how these agents should interact in terms

of interaction protocols and other technical issues

that concern architecture of the entire system.

4.1 The Architecture of an Agent

The internal architecture of an affective agent

consists of two abstract interrelated parts –

emotional computation model and rational processes

model (Marsella and Gratch, 2009) as well as

defines relations amongst these two parts. Emotional

computation model uses one or more emotional

theories, some of which are described in Section 3.2.

One promising way for modelling affective

processes among other options is vertically layered

architectures (Wooldridge, 1999). This type of

architectures enable organizing rational reasoning

components as well as emotion computation

components into layers where only some are

connected to external environment directly. Such

architectures have been successfully applied in multi

agent simulations, i.e. in (Tavakoli et al., 2014).

The architecture that we will apply in this

research is described in (Pudane et al., 2016). It

consists of three interconnected layers: primary

emotion layer, secondary (cognitive) emotion layer

and tertiary (self-reflection and social) emotion layer

(see Figure 1).

The used architecture allows not only rapid

stimulus processing on the primary level (i.e.,

startling from something) but also cognitive and

social processing of stimulus on secondary and

tertiary levels.

Figure 1: The internal structure of agent.

In an emotional contagion, the primary layer

generates primitive emotional contagion responses.

The secondary emotional contagion would appear on

higher levels, i.e., secondary level that compares the

world state to agents Desires and tertiary level that

allows to self-reflect and contains social Believes.

4.2. Affective Mechanism of Agent

Agent’s affective mechanism of emotional contagion

consists of three parts: personality calculation, rapid

(primitive) emotional state evaluation and

expression as well as performing secondary

emotional contagion.

First some processing of agent’s personality is

needed. Initially it is defined as a set of OCEAN

model values. Then those values are transformed

into agent’s default mood, defined in PAD space –

Emotion Contagion among Affective Agents - Issues and Discussion

331

as done in (Petrovica and Pudane, 2016). This

transformation allows associating the OCEAN

model with personality functions.

The default mood further impacts parameters of

the four functions: the activation function and decay

function for emotional state generation (for more

information see Petrovica and Pudane, 2016) as well

as susceptibility threshold and expression function

for emotional contagion. All of the functions are

being modelled separately for each of the basic

emotions.

Activation function maps objective intensity of

irritation to subjective intensity. Sigmoid was chosen

as a type of activation function, as it allows more

believable activation of emotions by enabling

emotion saturation and synergy properties (Picard,

1997). People also do not stay at high emotional

state for a very long time: they eventually go back to

their default mood thus the second function is

exponential Decay function.

Similarly, for each personality there is a level of

emotional intensity at which the emotion is started to

display (modelled as a threshold Susceptibility

function). The Expression function, similarly as

Activation function, is a sigmoid and determines

how actively the emotion will be expressed.

To implement the second level contagion, higher

levels of architecture are used. In the three-layered

architecture described before moving to higher

levels happens when the action in lower level cannot

be found (Pudane et al., 2016). Similarly, in the case

of an emotional contagion, if the emotion expression

cannot be started because the intensity of an emotion

is not above the susceptibility level, the agent still

starts processing it. The rational processing is based

on both, social and personal, Believes and Goals.

Believes also contain relationship weights,

depending on which, agent turns the intensity of

emotion up (thus emotion may reach the

susceptibility threshold and can be shown) or down

(thus fastening decay). Social Believe set also

include the agent's believes of what are other agent's

personalities. For example, if agent knows that

another agent is very expressive, he might not feel as

upset when another agent expresses sad emotion.

Similarly, the Believe about which of the 8 types of

moods another agent has, might lead to change in

emotional intensity.

The entire mechanism of calculating emotions is

shown in Figure 2. After the irritation comes, the

objective strength of emotional response is

determined (Petrovica and Pudane, 2016). In case of

an emotional contagion, the objective strength of

emotion is equal to output of first agent's expression

function value.

Figure 2: The process of emotional contagion.

The objective value is then applied to current

agent's activation and decay functions (which in turn

depend on personality) thus calculating the

subjective value.

If subjective value is above this threshold, the

expression strength is calculated, if not, emotion is

still passed to cognitive processing. The output of

cognitive processing is new value of emotional

intensity which may either fasten decay causing it to

drop or reach susceptibility threshold allowing to

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

332

express emotion. Otherwise the contagion stops.

If two emotions are above the susceptibility

threshold, then the relatively highest value is chosen,

i.e., if max happiness value of an agent is 5 and max

fear value is 3, and current value for both emotions

is 2, fear will be expressed instead of happiness.

If there is more than one interaction in a short

time period, each type of emotion is processed in its

own thread (thus finding the most intensive). If

emotions are of the same type, they are processed

one by one as shown in Figure 2.

5 MACRO LEVEL

Macro level of a system concerns two issues – how

the structure graph is modelled and how the dynamic

part of the simulation is executed. The structure of

network represents the interaction link amongst the

people – i.e., if they know each other and impact

each other. The dynamic part of the network runs a

simulation based on the structure, representing

interactions during specific period of time.

5.1 Structure of Network

The graphs (networks) on which the Monte Carlo

experiments will be run will need to be calibrated to

real networks. Thus, network metrics such as degree

distribution, clustering coefficient and importance

measures should match those of real networks. Real

networks often display a degree distribution that

follows a power law with a few people have many

connections and most people having only a few

connections.

To randomly generate a network the Barabasi

Albert algorithm (Barabasi and Albert, 1999) will be

used. This method of randomly creating a network

begins with a small graph and adds people (nodes)

one at a time until the desired total number of people

are connected to the network. Each new node to be

added to the network is connected to 2 already

connected nodes. The 2 already connected nodes are

chosen with a probability proportional to their

degree. Hence people who already have many

connections have a high probability of acquiring

new connections; the rich get richer.

The Barabasi Albert algorithm can be modified

in order to better match the particular parameters of

target networks. In addition, the people in the

network can be considered as identical and the

network homogeneous or the network can be

comprised of different types of people. It has been

shown that the distribution of the different types of

people on a network can affect the flow of a

contagion-like phenomenon over a network.

(Brooks, 2013)

5.2 Simulation in a Network

The agents in this model will alter and transmit their

emotional contagions according to axioms grounded

in the psychological literature. Monte Carlo

experiments will show how these contagions flow

across the network.

The simulations will be based on transmitting

rules that will depend on agents' personality and

weight on edges in network structure.

There are two possible options how weight can

impact transmitting rules.

 The weights represent the average value of

interaction frequency amongst the people. In

this case the weights would increase or

decrease the probability that the people would

“meet” in simulation.

 The weights represent close relationship

amongst the people. In this case the rules will

be implemented by using Believe set in agent

architecture.

All the rest contagion rules depend on functions

described in Section 4.2. If agent does not express

emotion, contagion stops.

6 CONCLUSIONS

The paper discusses technical issues as well as

affective mechanisms needed for full simulation

model of emotional contagion. The model will allow

implementing both primitive and secondary

(cognitive) emotional contagion mechanisms to

simulate real-life group of people.

Such system would enable a lot of features

currently simplified in existing models or eliminated

at all - such as multiple emotion integration.

The main success of the proposed system is dual

focus on both – micro- and macro- levels. It

implements rich model for internal agent structure

thus enabling agents with full-fledged emotions.

Furthermore, the mechanisms described here would

enable simulating groups of agents that ensures

believable interaction with the user.

This is an on-going research; the future work

includes architecture as well as dynamic simulation

implementations. Mathematical analysis will be

performed as well.

There are some mechanisms that so far are not

planned to implement, however it would be of

Emotion Contagion among Affective Agents - Issues and Discussion

333

paramount interest to be considered. One such thing

would be the simulation of a full working day that

would include various interaction events, i.e.,

meetings. Another issue was already mentioned

above – it would be interesting to see how a

“negative emotion contagion” works.

ACKNOWLEDGEMENTS

This work is partly funded by Faculty of Computer

Science and Information Technology (Riga

Technical University) assigned Doctoral grant

DOK.DITF/16 and Latvian National Research

Program SOPHIS grant No.10-4/VPP-4/11.

REFERENCES

AlSagri, H., Ykhlef, M., 2016. A Framework for

Analyzing and Detracting Negative Emotional

Contagion in Online Social Networks. In Proceedings

of ICICS 2016.

Aydt, H., Lees, M., Luo, L., Cai, W., Hean Low, M.Y.,

Kadirvelen, S.K., 2011. A Computational Model of

Emotions for Agent-based Crowds in Serious Games.

In Proceedings of IEEE/WIC/ACM International

Conferences on WI-IAT, vol. 2, pp. 72-80.

Barabasi, A.L., Albert, R., 1999. Emergence of scaling in

random networks. In: Science, vol. 286, pp. 509–512.

Barsade, S.G., 2002. The Ripple Effect: Emotional

Contagion and Its Influence on Group Behavior. In

Administrative Science Quarterly, 47, pp. 644–675.

Bassi, B., 2006. Computer Simulation of Crowd Dynamics

and Destructive Crowd Behavior. Honors Scholar

Theses.

Bispo, J., Paiva, A., 2009. A Model for Emotional

Contagion Based on the Emotional Contagion Scale.

In Proceedings of ACII 2009, pp. 1-6.

Bosse, T., Duell, R., Memon, Z.A., Treur, J., and Wal,

C.N. van der, 2015. Agent-Based Modelling of

Emotion Contagion in Groups. In Cognitive

Computation Journal. In press.

Brooks, B.P, 2013. Rumour Propagation on Social

Networks as a Function of Diversity. In Advanced

Dynamic Modeling of Economic and Social Systems:

Studies in Computational Intelligence, vol. 448/2013,

pp. 49-60.

Brooks, B.P., DiFonzo, N., Ross, D., 2013. GBN-

Dialogue Model of Outgroup-Negative Rumor

Transmission: Group Membership, Belief, and

Novelty. In Nonlinear Dynamics, Psychology and Life

Sciences, vol. 17, no. 2, pp. 269-293.

Damasio, A., 1994. Descartes’ Error, Emotion Reason

and the Human Brain, Avon Books. New York.

Doherty, R.W., 1997. The Emotional Contagion Scale:

Measure of Individual Differences. In Journal of

Nonverbal Behavior, vol. 21, pp. 131-154.

Ekman, P., 1992. Are there basic emotions? In

Psychological Review, vol. 99, no. 3, pp. 550-553.

Gebhard, P., 2005. ALMA - A Layered Model of Affect.

In Proceedings of AAMAS'05, pp. 29-36.

Hatfield, E., Cacioppo, J. L. and Rapson, R. L., 1993.

Emotional contagion. In Current Directions in

Psychological Sciences, vol. 2, pp. 96-99.

Hudlicka, E., 2015. Computational Analytical Framework

for Affective Modeling: Towards Guidelines for

Designing Computational Models of Emotions. In

Handbook of Research on Synthesizing Human

Emotion in Intelligent Systems and Robotics. IGI

Global, U.S.A.; pp. 1-62.

Kramer, A., Guilory,J. E., Hancock, J.T., 2014.

Experimental evidence of massive-scale emotional

contagion through social networks. In PNAS, vol. 111

no. 24.

Lisetti, C., Hudlicka, E., 2015. Why and How to Build

Emotion-Based Agent Architectures. In The Oxford

Handbook of Affective Computing

. Oxford University

Press, UK.

Marsella, S., Gratch, J., 2009. EMA: A computational

model of appraisal dynamics. In Cognitive Systems

Research, vol 10. no. 1, pp. 70-90.

McCrae, R.R., Costa P.T. Jr., 2003. Personality in

Adulthood: A Five-Factor Theory Perspective, New

York: Guilford, 2003.

Ortony, A., Clores, G.L., Collins, A., 1988. The Cognitive

Structure of Emotions, Cambridge University Press,

Cambridge.

Pease, C.R., Lewis, G.J., 2015, Personality links to anger:

Evidence for trait interaction and differentiation across

expression style. In Personality and Individual

Differences, Vol. 74, pp. 159–164.

Petrovica, S., Pudane, M., 2016. Emotion Modeling for

Simulation of Affective Student-Tutor Interaction:

Personality Matching. In International Journal of

Education and Information Technologies. Vol.10, pp.

159-167.

Picard, W., 1997. Affective Computing, MIT Press.

Cambridge.

Pudane, M., Lavendelis, E., Radin, M.A., 2016. Human

emotional behavior simulation in intelligent agents:

processes and architecture. In Procedia Computer

Science, in press.

Russell, J.A., Mehrabian, A., 1977. Evidence for a three-

factor theory of emotions. In Journal of Research in

Personality, vol. 11, pp. 273-294.

Tavakoli, M., Palhang, M., Kazemifard, M., 2014. Three

levels of information processing: improvement using

personality. In Proceedings of ICIS 2014, pp. 1 – 6.

Wooldridge, M., 1999. Intelligent agents, In Multiagent

Systems: A Modern Approach to Distributed Artificial

Intelligence. MIT Press, Cambridge, pp. 27-78.

Wooldridge, M., 2009. An Introduction to MultiAgent

Systems. John Wiley & Sons.

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

334