Building Emotional Agents for Strategic Decision Making

Bexy Alfonso

, Emilio Vivancos

, Vicente Botti

and Pen´elope Hern´andez

Departamento de Sistemas Inform´aticos y Computaci´on, Universitat Polit`ecnica de Val`encia, Valencia, Spain

ERI-CES and Departamento de An´alisis Econ´omico, Universidad de Valencia, Valencia, Spain

Keywords:

Agents, Emotional Architecture, Simulation, Prisoner Dilemma, Trust Game.

Abstract:

Experimental economics has many works that demonstrate the inﬂuence of emotions and affective issues on

the process of human strategic decision making. Personality, emotions and mood produce biases on what

would be considered the strategic solution (Nash equilibrium) to many games. Thus considering these issues

on simulations of human behavior may produce results more aligned with real situations. We think that

computational agents are a suitable technology to simulate such phenomena. We propose to use O3A, an

Open Affective Agent Architecture to model rational and affective agents, in order to perform simulations

where agents must take decisions as close as possible to humans. The approach evaluation is performed

trough the classical ‘prisoner dilemma’ and ‘trust’ games.

1 INTRODUCTION

There is enough evidence about the behavioral issues

of the decision-making process (Lerner et al., 2004;

Elbanna, 2006; Seo et al., 2010). Phenomena like the

individual “risk aversion” (how tolerant is an individ-

ual to risk) (Loewenstein et al., 2001), or the “illusion

of control” (overestimationof individualsof their abil-

ity to control events) (Schwenk, 1984), also involve

the affect of the decision-maker. Thus, for example,

an owner of a property that has beneﬁted from it for

many years, and has affective bonds with it, would try

to sell the property at a higher price than an investor

that has not relation with the property. Bearing all

this in mind one may assume that when performing

experiments with human that must make strategic de-

cisions, it is very likely to get individual biased action

proﬁles regarding to what should be a rational solu-

tion. Therefore, when simulating human behavior in

scenarios of strategic decision-making, the representa-

tion of the individual affect may become an important

parameter in order to get results better aligned with a

real situation. Computational agents could be a suit-

able technology for creating such simulations. This

ﬁeld has evolved enough to build rational entities with

a human-like practical reasoning (Weiss, 1999). The

BDI architecture is a widely accepted conceptualiza-

tion for agents that includes on its deﬁnition represen-

tative elements of human characteristics. Neverthe-

less one of the biggest challenges in this area is to pro-

vide the agents with the necessary structures in order

to reach an accurate representation of the human affec-

tive side. Some steps have been taken inspired on psy-

chological and/or neurological grounds (Marsella and

Gratch, 2009; Becker-Asano and Wachsmuth, 2010),

nevertheless a bigger effort must be done to offer less

disperse approaches and to follow an incremental line

of research (Marsella et al., 2010).

With this work we want to address some impor-

tant questions: is it possible to build a scenario with

entities representing humans that can take decisions

as humans do? How can an affective component bias

results from the rationally optimal solution? How to

model this to properly simulate the strategic decision-

making process in humans? We want to give an an-

swer to these and other questions. Our aim is to model

entities able to behave as humans do by integrating ra-

tional and emotional components, and offer them as

what-if tools than can behave as and interact with hu-

man in situations of strategic decision making. We

propose an instantiation of the Open AffectiveAgents’

Architecture. Our approach is grounded on strong

psychological and neurological fundaments.

In order to validate our proposal we have imple-

mented it, and we have used classical games of exper-

imental economy (as the prisoner’s dilemma, dictator,

ultimatum and trust games). This what-if tool allows

to tune human affective characteristics in multi-agent

systems and analyze how they inﬂuence the decision-

making process in classical games.

390

Alfonso B., Vivancos E., Botti V. and Hernández P..

Building Emotional Agents for Strategic Decision Making.

DOI: 10.5220/0005204703900397

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 390-397

ISBN: 978-989-758-074-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

The rest of the paper is organized as follows. A re-

view of the related literature is presented in section 2.

The main components of the supporting architecture

of the approach are described in section 3. Assump-

tions, considerations as well as the main architecture

components are described in section 4. In section 5

the experiments performed as well as their results are

offered. Finally, section 6 offers a discussion and the

conclusions of the work.

2 RELATED WORK

2.1 Psychological and Neurological

Background

Psychology uses various concepts to describe the af-

fective characteristics of humans. At ﬁrst there are

individual intrinsic traits that inﬂuence motivations

and behavior known as personality (Ryckman, 2007).

Then we found approaches about emotions whose

deﬁnitions vary depending on researchers and disci-

plines, but researchers generally agree on the notion

of reactions as a consequence of events, actions of

other agents, and/or objects (Ortony et al., 1988). An-

other issue considered in psychological literature is

that, regardless of the emotions experienced after an

event, humans keep a kind of temperament or mood,

which has less intensity than emotions, isalso an expe-

riential component but lasts longer and is not necessar-

ily associated to a cause (Mehrabian, 1997). Personal-

ity and emotions have been addressed from different

perspectivesby important psychologists and scientists

(Cornelius, 2000; Ryckman, 2007). The cognitive

perspective stands out among the others for compu-

tational applications. On the other hand neurological

experiments use advanced technology to ﬁnd out the

way the brain works on linking the human emotions

mechanism and cognition processes (Pessoa, 2008).

A. Dam´asio and J. E. Ledoux works have been very

relevant on the study of the emotions and the brain

(LeDoux, 1998; Dam´asio, 1994). Dam´asio argued

that in some situations affective factors, as power-

ful heuristics of the brain to solve complex prob-

lems, may get better decisions than rational factors

(Bechara et al., 1997), in that they alleviate individ-

uals of the overloads that may come by using only

cognitive processes when facing complex choices.

2.2 Computational Approaches

In this section we review some approaches that ad-

dress the main processes related to the affective side

of individuals from a cognitive perspective. These

processes include the appraisal, the affect internal dy-

namics and the affect consequences. The way they

dial with the different processes as well as the main

psychological concepts related to affect is often par-

tial and domain speciﬁc. There are interdisciplinary

approaches that are continuously used (Ortony et al.,

1988; Mehrabian and Russell, 1974; Ekman, 1999b).

They have become widely accepted alternatives due

to their suitability for computational applications.

The BDI architecture has commonly been a start-

ing point to model rational agents. This is the case of

(Jiang et al., 2007). The authors propose a practical

reasoning separated from the emotions mechanism,

and primary (infant-like emotions such as “angry”,

“happy” or “surprised”) as well as secondary emo-

tions (prospect-based and directly related to expecta-

tions and past experiences) receive a differentiated at-

tention when they inﬂuence the process of decision-

making of a traditional BDI architecture. The emo-

tional state of the agent is represented through a set of

emotion intensities. This state inﬂuences the way be-

liefs are acquired from communication or contempla-

tion and also the prioritization of the agents desires.

All emotions, beliefs, desires and intentions are as-

signed a priority. Emotions prioritize desires and they

also help to decide intentions. In this architecture in-

dividual differences associated to emotions are deter-

mined by the speciﬁc model used on each agent to

deal with emotions. The dynamic change of the affec-

tive state is not considered and there is no feedback

from previous situation that allows to learn from emo-

tional states.

S. C. Marsella and J. Gratch created EMA, which

stands for “EMotion and Adaption” (Marsella and

Gratch, 2009). They describe a computational model

for the dynamic of emotional appraisal, and provide

a framework based on a domain independent archi-

tecture for emotional agents. In EMA a computa-

tional model of appraisal uses the interpretation of

a person-environment relationship (causal interpreta-

tion), and this interpretation is done in terms of a set

of appraised variables and is altered by a set of “cop-

ing strategies” (processes that manipulate this repre-

sentation to respond to the appraised interpretation).

The appraised variables have some values for each

proposition extracted from the environment which

are stored in the Appraisal frames. These variables

are: relevance, perspective, desirability, likelihood,

expectedness, causal attribution, controllability and

changeability. A two-level notion of emotional state

is modeled: appraisal and mood. The ﬁrst determines

the agent coping response, and the second has an in-

direct effect on appraisal in that it is applied a mood

BuildingEmotionalAgentsforStrategicDecisionMaking

391

adjustment to individual appraisal frames. Symbolic

labels of emotions (like fear, joy or hope) are assigned

to appraisal frames, but the agent coping responses

are determined by the appraisal variables. Mood is

represented through a set of emotion labels with some

intensity. Again in this approach individual differ-

ences are not clearly modeled and the emotions com-

ponent cannot be easily detached from the practical

reasoning of the agent to allow its improvement or

modiﬁcation.

C. Becker-Asano and I. Wachsmuth built

WASABI (Becker-Asano and Wachsmuth, 2010).

WASABI is an architecture for affect simulation

whose model of core affect is based on the PAD

theory (Mehrabian, 1997), and its appraisal process

is inspired by Scherer’s sequential-checking theory

(Scherer, 2001). This model distinguishes primary

and secondary emotions. It is based on (Dam´asio,

1994), (LeDoux, 1998), (Ekman, 1999a) and (Ortony

et al., 2005) theories. Mood is modeled as a back-

ground state whose value moves in a bipolar scale

of positive versus negative. Other issues added to

this proposal are the memory and emotion dynamics

components. This work was used for simulating

emotions capabilities in a virtual player. The sig-

niﬁcance of primary and secondary emotions was

evaluated. Speciﬁc structures for dealing with each

virtual character personality and individual traits are

not explicitly deﬁned in the architecture. Its range of

domain applications moves around virtual characters

and robotics, since it has embedded elements for

physical behavior like facial expressions as manifes-

tation of emotions. Nevertheless in order to achieve

complex cognitive behaviors its integration in other

domains like multi-agent simulations is difﬁcult.

R. Santos et al. describe a group decision-support

system that combines personality, emotion and mood.

The approach is based on the Five Factor Model

(FFM) to represent the personality (McCrae and John,

1992), and the PAD space to model mood (Mehra-

bian, 1996b). The emotional system uses the Ortony’s

improved version of the OCC model of emotions

(Ortony, 2003). Emotions then inﬂuence the argu-

mentation process after being mapped to the PAD

space (following (Gebhard, 2005)), and also person-

ality is considered in the argumentation phase. Al-

though many affective concepts are considered in this

architecture it is speciﬁc for a scenario of negotiation

with argumentation in a group and it is difﬁcult to sep-

arate the overall practical reasoning of the agent.

A. Fagundes et al. develop an architecture for

emotional agents. Again the affective components

emotions, personality and mood are integrated in a

BDI architecture. The results of the ALMA project

Personality

Current mood

Emotion

reactive

Mood

component

Coping

component

Results

Beliefs

Percepts

influenced

by mood

Emotion

deliberative

Appraisal

Intentions

Options

Beliefs

component

Primary

emotions

Secondary

emotions

Figure 1: Main components of the O3A architecture.

(Gebhard, 2005) are used, and hence the OCC, PAD

and the FFM theories. Although the focus of the ap-

proach is on the cognitive state and it is easy to be

applied to various situations, it makes strong assump-

tions such as the “discarding” of percepts or elements

of the memory if they don’t rise intense emotional

states or if they don’t lead to a desirable emotional

state. Also there is not a possibility for adapting the

way emotions and mood inﬂuence the cognitive pro-

cesses for particular requirements.

In (Alfonso et al., 2014) we propose “O3A” (an

Open Affective Agent Architecture). It mainly pur-

sues to be open enough to allow different implemen-

tations of its components in order to be adapted to dif-

ferent situations. It also integrates the most important

concepts addressed on psychological theories about

affect and emotions on humans in a BDI architecture.

This way the practical reasoning and affective issues

converge in a single representation where the respon-

sibilities of each component are well deﬁned and in-

teract to produce diverse, adaptive, and believable be-

haviors on agents. It will be better described in sec-

tion 3.

3 SUPPORTING ARCHITECTURE

Based on the appraisal theory, O3A combines men-

tal, cognitiveand motivational issues with an affective

component (Alfonso et al., 2014). O3A is build over

a BDI architecture, what allows it to reuse all the ma-

chinery of the widely accepted BDI structure. More-

over O3A brings some of the most important and ac-

cepted theories in psychology and neurology related

to emotions.

3.1 Main Components

The O3A basic structure is depicted in ﬁgure 1. Four

main components are in charge of regulating the emo-

tional processes: appraisal, emotions dynamics, and

the inﬂuence of emotions on beliefs and on intentions.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

392

Appraisal Component. Controls how emotions are

derived from the environment and also from the cur-

rent agent state. This task is performed by two sub-

components: the emotion reactive component and

the emotion deliberative component. The former

derives primary emotions, which follows the idea

of “onto-genetically earlier emotions” (Becker-Asano

and Wachsmuth, 2010) like those experienced by an

infant. The later determines secondary emotions con-

sidering that they are the result of more complex

chains of reasoning.

Beliefs Component. Depending on the agent cur-

rent mood, percepts may trigger different effects. The

judging of the value and importance of objects and

events is inﬂuenced by the affective valence and

arousal; for example the likelihood of seeing a poten-

tial trait increases when experiencing fear (Zadra and

Clore, 2011).

Mood Component. It is in charge of deriving a

global temperament or mood on the basis of the per-

ceived emotions and the previous mood. Personality

determines which is the mood of the agent when it is

in a neutral or “equilibrium” state. This component

also controls the way mood returns to this “equilib-

rium” state.

Coping Component. It has a close relation with the

agent intentions since it determines how the current

mood ﬁnally inﬂuences the selection of the next ac-

tion to perform through the prioritization of its inten-

tions.

4 COMPONENTS DESIGN AND

CONSIDERATIONS

4.1 Components Design

O3A is open on its deﬁnition which allows to use dif-

ferent theories in order to determine how each one of

its processes must be performed. Hence, for character-

izing the way in which some of the architecture com-

ponents speciﬁcally works, we were inspired by pre-

vious works that ﬁt to each component deﬁnition, and

simultaneously we offer our own vision of the not cov-

ered issues. The main theories used to instantiate the

O3A Components are based on the works of (Ortony

et al., 1988), (Mehrabian, 1996b) and (Marsella and

Gratch, 2009). These works have proven to produce

relevant results in the area and are versatile enough to

be reused in new approaches.

In our design the emotion reactive component as-

sumes that percepts are labeled with the “more com-

mon emotions” to be experienced given these percept.

So for example the percept “hurricane” may have an

associated “fear” label that becomes the correspond-

ing primary emotion after been processed by the emo-

tion reactive component. The emotion deliberative

component, as stated on its own deﬁnition, performs

a more complex processing and considers more fac-

tors in order to derive secondary emotions. In our

design secondary emotions are derived from events

of any nature. They can be internal events (like the

beginning of a new intention) or external (a percept).

Events are characterized according to ﬁve variables:

desirability, likelihood, expectedness, causal attribu-

tion, and controllability. The deﬁnition of each one of

these characteristics is similar to the variables of the

appraisal frames presented in (Marsella and Gratch,

2009). From this work it was also used the mapping

from the appraisal pattern to emotion labels, which

are secondary emotions in our design. The variables

characterizing the appraisal frames are linked to the

agent mental state and they are:

desirability: linked to the agent general standards or

preferences of the event consequences. For example

if an agent is given a possibility to play in one of two

lotteries and the minimum to pay in both is higher that

the prize, the desirability of the event “time to play”

will be very low. On the other hand, even if the prize

is lower, but the collected money goes to charity, and

it is signiﬁcant for the agent, the desirability will be

high. We have called this property “personal beneﬁt”,

and is applicable to each possible option to carry out

the intention triggered by the event.

likelihood: this variable is linked to the likelihood of

outcomes. It considers the existence of past or future

states. Currently our implementation of the emotion

deliberative component only considers present propo-

sitions so the value for this variable is always 1.

expectedness: linked to the agent expectations asso-

ciated to the event. Whether the expectations are ful-

ﬁlled or violated may produce emotive reactions and

hence changes in the agent mood.

causal attribution: whether the event was produced

by the agent or by other agent or source.

controllability: linked to the capacity of the agent to

react in some way to the event. If there are no actions

deﬁned to respond to the event its controllability will

be low.

On the other hand, the Mood component is in

charge of keeping the “current mood” updated. We

used the PAD dimensional approach (Mehrabian and

Russell, 1974) for representing the mood. The initial

mood is deﬁned by the agent personality (it is spec-

iﬁed in the agent deﬁnition). For representing the

personality we used the Five Factor Model (McCrae

and John, 1992). The FFM is able to accurately de-

BuildingEmotionalAgentsforStrategicDecisionMaking

393

scribe individual traits through ﬁve dimensions (open-

ness, conscientiousness, extraversion, agreeableness,

and neuroticism). The current mood will be updated

as soon as the emotions are appraised in such a way

that with a positive mood, the agent is less likely to ex-

perience a negative mood and vice versa. In particular

the mapping from the agent ﬁve dimensions of per-

sonality to the three dimensions of the PAD space to

establish the initial mood is done according to Mehra-

bian’s work (Mehrabian, 1996a). On the other hand,

the transformation of the set of emotions appraised

into the PAD mood follows the result of P. Gebhard in

(Gebhard, 2005). The Coping component then feeds

on the current mood in order to establish a priority

for the agent intentions. This requires to represent a

proper measure for each intention indicating beneﬁts,

risks and other factors that may be biased by the agent

current mood. Currently we limit this measure to the

intention risk, that is, a measure of possible losses or

undesirable states for the agent, having in mind the

widely treated issue of risk aversion in the behavioral

economic literature (Loewenstein et al., 2001; Harri-

son and Rutstr¨om, 2008; Demaree et al., 2009). In

particular some works demonstrate that the inﬂuence

of the trait Dominance of the PAD space is a signif-

icant indicator of the risk of the decision (Demaree

et al., 2009).

5 EXPERIMENTS AND RESULTS

5.1 General Design and Integration in

Jason

We have used the BDI algorithm of Jason (Bordini

et al., 2007) to integrate the O3A emotional compo-

nents in a BDI architecture. This algorithm offers all

the necessary elements to carry out the proposed de-

sign. Table 1 shows the main elements of the model

and how they are inserted in the Jason inner structure.

Current mood, and primary and secondary emotions

are modeled as internal objects of the Jason architec-

ture. The agent personality is speciﬁed in the agent

deﬁnition and it is stored on its correspondinginternal

structure. On the other hand, to perform an intention,

each plan must have a set of necessary “personal ben-

eﬁt arguments” in order to evaluate the “beneﬁts” of

each possible option of the agent according to its pref-

erences and standards. The “associated risk” is in line

with the global preferability of each possible option.

The process of derivation of the primary emotions

uses perceptions from the ‘Perceive’ process of Jason.

Secondary emotions are determined through the GET-

SECONDARY-EM process and takes the event to be

processed on each reasoning cycle deriving, if it is ap-

propriate, the secondary emotions. This process feeds

on Beliefs and on the Desires derived in the Jason

‘Get

option’ process. The initial mood is set on the

process of agent initialization and the current mood is

updated each time new emotions are appraised. After

selecting an event to process, the possible options for

an agent are the Jason “applicable plans”. Then the Ja-

son ‘Filter’ process is customized in order to consider

also the agent current mood.

Table 1: Integration of the affective components in the Jason

platform.

VARIABLES

Description Jason element

Current Mood (M) Internal Object

Primary Emotions (PEm) Internal Object

Secondary Emotions (SEm) Internal Object

Personality (P) Agent deﬁnition

Plans attributes

Associated risk Plan annotation

Personal Beneﬁt args. Agent rule

PROCESSES

Description BDI integration

GET-PRIMARY-EM Perceive

input: Labeled perc. | output: PEm

GET-SECONDARY-EM Get option

input: Events | output: SEm Beliefs

SET-INITIAL-MOOD Initialization

UPDATE-MOOD

FILTER-APPLICABLE-PLANS Filter

5.2 Two Classical Games and Results

Strategic games are an important tool in experimental

economy to evaluate how individuals behave in situ-

ations where a single decision must be taken (Swope

et al., 2008). In many of these games there are gen-

erally two players involved, and there is also a stable

equilibrium concept called “Nash equilibrium”: situa-

tion where no player can obtain more beneﬁts by tak-

ing a different decision. Nevertheless, what happens

in real situations is that individuals take decisions bi-

ased by emotions. Our proposal aims to model the

behavior of individuals in real situations where deci-

sions are taken considering rational and emotional as-

pects. We have chosen the Prisoner dilemma and the

Trust Game in order to evaluate if the decisions inﬂu-

enced by emotions in our system are similar to the

decisions of individuals in a real context.

Prisoner Dilemma (PD)

This game offers a model of cooperation that empha-

sizes how individual and collective interests coexist

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

394

(Axelrod and Hamilton, 1981). The participants rep-

resent two prisoners. Each one has the opportunity to

declare that the other committed the crime and hence

to betray him, or otherwise cooperate with him and

remain silent. For each possible situation the conse-

quences are (Tucker, 1983): i) both prisoners betray

the other: 2 years in prison for both, ii) prisoner 1

betrays prisoner 2 and prisoner 2 remains silent: pris-

oner 1 gets free and prisoner 2 gets 3 years of prison,

iii) both prisoners remain silent (cooperate): both get

1 year of prison. The most rational option according

to economic theory would be that they both betray

each other (Sewell, 2010), nevertheless a systematic

behavior in humans is to have a silent participant side

what is an evidence of cooperation between them.

The code of ﬁgure 2 is an extract of the agentA

implementation. It expects that agentB is not going

to betray (line 1). The expectation is represented

through a rule with an “EventFunctor” (the event that

triggers the expectation veriﬁcation), a fulﬁllment

condition (that belief that makes the expectation

to become fulﬁlled) and a violation condition (a

belief that is the opposite of the fulﬁllment belief)

The personal beneﬁt of the agent for each option

is determined trough the rule of line 2. The agent

basically looks for an option that reports to him the

lower value of the maximum possible years in prison

after selecting the corresponding option. This value

is stored on the ‘PB’ variable and ‘Max’ and ‘Min’

are the maximum and minimum possible values for

‘PB’. The fourth parameter of the rule is a list with the

properties (in form of literals) of the plan to consider

in the calculus. Lines 4 and 5 contain the possible

options after the event, that is the plans with this

triggering event. The plans annotations correspond

to the parameters to determine the “personal beneﬁt”

and the risk associated. Risk values are two numbers

arbitrarily selected. What is important is the relative

relation between plans risks (the ‘p

silence’ plan is

more risky considering the best solution according to

Nash equilibrium). Another issue to add is that the

triggering event (‘+decide’), is acquired as a percept:

decide(cooperationpayoff(2), defectpayoff(1),

delatorpayoff(0),temptationpayoff(3) [emotions

(fear,hope,joy)]

that has the most common

emotions we have considered for this situation. The

agentB is similar to agentA, but in this case he doesn’t

have expectations.

Both the fulﬁllment and violation conditions are ex-

plicitly modeled on expectations because we take an “open

world” assumption so if a negation of a belief is not explic-

itly declared, then the agent has no information about if it is

true or false.

Values extracted from (Ashlock and Rogers, 2008)

expect__("silence",FullfilCond ,

ViolateCond ,EventFunctor):-

FullfilCond =silence(decision)

[source(agentB)] &

ViolateCond =betray(decision)[

source(agentB)] & EventFunctor="+

decide".

pb_max(PB, Max, Min, [max_years(MY)])

:- (not .ground(MY)|MY>0) & Max=3

& Min=0 & PB=Max/MY.

@p_betrayal[max_years(2),risk(0.3)]

+!decide (...) <- (...)

@p_silence[max years(3),risk(0.5)]

+!decide (...) <- (...)

Figure 2: Appearance of the Jason implementation of the

agentA agent in the prisoner dilemma.

The values for the parameters of the personal-

ity proﬁles were deﬁned according to (Santos et al.,

2010). A set of four combination of values were

used. These four personality types are: social, trou-

bleshooter, negotiator, and realistic. Some combina-

tion of these personality proﬁles produced the follow-

ing results:

A social, B troubleshooter: A silence, B betrayal

A social, B negotiator: A silence, B silence

A social, B realistic: A silence, B betrayal

A social, B social: A silence, B silence

Coherently with the personality types chosen, in

this experiment the social agent tends to be coopera-

tive (keeping silence). The troubleshooter and realis-

tic tend to be more pragmatic betraying the other pris-

oner and the negotiator takes risks looking for beneﬁts

and remains in silence (cooperates).

Trust Game (TG)

Another important sequential game is the the trust

game (Berg et al., 1995). In this game a proposer (or

trustor) is being given an amount of endowment to be

split and shared with the trustee. The amount shared

is multiplied by a factor (usually three) and then the

trustee decides to give back any amount. This way

the trustor hopes to receive something back related to

it’s initial offer trusting in the intentions of the other.

In real situations both players tend to share more than

what would be predicted in the game perfect equilib-

rium: “no trust” (Bracht and Feltovich, 2008).

The game was implemented similar

to PD. In this case, the trustor perceives

decide(endowment(10))[emotions(hope)]

from

the environment, expects a fair offer from the trustee

(around the half) and his personal beneﬁt is max-

BuildingEmotionalAgentsforStrategicDecisionMaking

395

1 Trustor

2 @p 1st [ r e s t ( 75) , r i s k (0 .3 ) ]+! decide ( . . . )

3 @p 2nd[ r e s t ( 60) , r i s k (0 .4 ) ]+! decide ( . . . )

4 @p 3rd [ r e s t ( 4 5 ) , r i s k (0 .5 ) ]+! decide ( . . . )

5 @p 4th [ r e s t ( 30) , r i s k (0 .5 ) ]+! decide ( . . . )

6 Trustee

7 @p 1st [ o f f e r (2 0 ) , r i s k (0 .1 ) ]+! o f f e r ( . . . )

8 @p 2nd[ o f f e r ( 30) , r i s k (0 .2 ) ]+! o f f e r ( . . . )

9 @p 3rd [ of f er (40) , r i s k (0 .4 ) ]+! o f f e r ( . . . )

10 @p 4th [ o f f e r (5 0 ) , r i s k (0 .6 ) ]+! o f f e r ( . . . )

Figure 3: Appearance of the Jason implementation of the

trustor and trustee agent in a trust game.

imized when he gives a fair offer. Moreover the

trustee has no labeled percepts or expectations, and

his personal beneﬁt is maximized if he offers as less

as possible. The possible options for both players are

shown in ﬁgure 3. The parameters for the trustor are

what he keeps in percent after sharing (rest) and risk,

and for the trustee are what he offers from the triplied

endowment received in percent, and risk. Again the

values have no real meaning but the relative relation

between them. The combination of some personality

proﬁles and the conﬁguration described previously

produced the following results:

A social, B troubleshooter: A 55%, B 40%

A social, B negotiator: A 55%, B 40%

A social, B realistic: A 55%, B 40%

A social, B social: A 55%, B 50%

A troubleshooter, B social : A 35%, B 50%

A negotiator, B social: A 55%, B 50%

A negotiator, B realistic: A 55%, B 40%

A troubleshooter, B realistic: A 35%, B 30%

Similarly to the PD, in the TG the social agent be-

haves in a cooperative way and offers a high quantity

(55% as trustor or 50% as trustee). The troubleshooter

and realistic are prudent showing a pragmatic behav-

ior (35-40% and 30-40% respectively), and the nego-

tiator takes risks offering always more than the mini-

mum allowed (55-40%), but remaining prudent.

6 DISCUSSION AND

CONCLUSIONS

In this work we used an open agent affective archi-

tecture to build emotional agents that must simulate

situations of strategic decision-making inﬂuenced by

affective issues. We have used several widely ac-

cepted supporting psychological theories in order to

contribute to an incremental line of research in the

ﬁeld of emotional computing. We have also made our

own propositions and assumptions to ﬁll uncovered

gaps overlooking a ﬁnal implementation. These ini-

tial results demonstrate in ﬁrst place that there are ev-

idences that we can reach a diverse, believable and

closer to humans behavior when including an emo-

tion mechanism in an agent. The behavior of O3A

agents in some classical games is more like the behav-

ior of humans in real situations compared to agents

that doesn’t include emotions. Between the affective

issues modeled, personality seems to produce more

variability on results, what is quite consistent with a

real situation. Nevertheless using iterated versions of

the games may change the result and would help to

better tune the overall design. In the proposed de-

sign the inﬂuence of agents interactions on emotions

is given by the use of ‘expectations about others’ and

‘personal beneﬁts’ related to others, which produces

emotional reactions associated to interactions. Never-

theless we aim to improve existing structures and to

create new ones in order to offer more realistic simu-

lations of the human decision process.

ACKNOWLEDGEMENTS

This work was supported by the Spanish government

grant MINECO/FEDER TIN2012-36586-C03-01and

HUMBACE (Human and Social Behavior Models

for Agent-Based Computational Economics) project.

Thanks also to the Research and Development Sup-

port Programme of the Universidad Polit´ecnica de Va-

lencia PAID-06-11 Project.

REFERENCES

Alfonso, B., Vivancos, E., and Botti, V. J. (2014). An Open

Architecture for Affective Traits in a BDI Agent. In In

Proceedings of the International Conference on Evolu-

tionary Computation Theory and Applications (ECTA-

2014), pages 320–325.

Ashlock, D. and Rogers, N. (2008). A model of emotion in

the prisoner’s dilemma. In IEEE Symposium on Com-

putational Intelligence in Bioinformatics and Compu-

tational Biology, 2008. CIBCB’08, pages 272–279.

Axelrod, R. and Hamilton, W. D. (1981). The evolution of

cooperation. Science, 211(4489):1390–1396.

Bechara, A., Damasio, H., Tranel, D., and Dam´asio, A. R.

(1997). Deciding advantageously before knowing

the advantageous strategy. Science, 275(5304):1293–

1295.

Becker-Asano, C. and Wachsmuth, I. (2010). Affective

computing with primary and secondary emotions in

a virtual human. AAMAS ’10, 20(1):32–49.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

396

Berg, J., Dickhaut, J., and McCabe, K. (1995). Trust, reci-

procity, and social history. Games and economic be-

havior, 10(1):122–142.

Bordini, R. H., H¨ubner, J. F., and Wooldridge, M. (2007).

Programming multi-agent systems in AgentSpeak us-

ing Jason. Wiley.

Bracht, J. and Feltovich, N. (2008). Efﬁciency in the trust

game: an experimental study of precommitment. In-

ternational Journal of Game Theory, 37(1):39–72.

Cornelius, R. R. (2000). Theoretical approaches to emotion.

In ISCA Tutorial and Research Workshop (ITRW) on

Speech and Emotion.

Dam´asio, A. R. (1994). Descartes’ error: emotion, reason,

and the human brain. Quill.

Demaree, H. A., DeDonno, M. A., Burns, K. J., Feldman,

P., and Everhart, D. E. (2009). Trait dominance pre-

dicts risk-taking. Personality and Individual Differ-

ences, 47(5):419–422.

Ekman, P. (1999a). Basic Emotions, pages 45–60. John

Wiley & Sons, Ltd.

Ekman, P. (1999b). Facial expressions, chapter 16, pages

301–320. The Handbook of Cognition and Emotion.

John Wiley & Sons, Ltd.

Elbanna, S. (2006). Strategic decision-making: Process per-

spectives. International Journal of Management Re-

views, 8(1):1–20.

Gebhard, P. (2005). ALMA: a layered model of affect. In

Proceedings of the fourth international joint confer-

ence on Autonomous agents and multiagent systems,

pages 29–36, NY, USA. ACM.

Harrison, G. W. and Rutstr¨om, E. E. (2008). Risk aversion

in the laboratory. Research inexperimental economics,

12:41–196.

Jiang, H., Vidal, J. M., and Huhns, M. N. (2007). EBDI:

An Architecture for Emotional Agents. In Proceed-

ings of the 6th International Joint Conference on Au-

tonomous Agents and Multiagent Systems, pages 11:1–

11:3. ACM.

LeDoux, J. E. (1998). The emotional brain: the mysterious

underpinnings of emotional life. A Touchstone book.

Simon & Schuster.

Lerner, J. S., Small, D. A., and Loewenstein, G. (2004).

Heart strings and purse strings: Carryover effects of

emotions on economic decisions. Psychological Sci-

ence, 15(5):337–41.

Loewenstein, G., Weber, E. U., Hsee, C. K., and Welch,

N. (2001). Risk as Feelings. Psychological Bulletin,

127(2):267–286.

Marsella, S. C. and Gratch, J. (2009). EMA: A process

model of appraisal dynamics. Cognitive Systems Re-

search, 10(1):70–90.

Marsella, S. C., Gratch, J., and Petta, P. (2010). Computa-

tional models of emotion. In A Blueprint for Affective

Computing: A Sourcebook and Manual, Affective Sci-

ence, pages 21–46. OUP Oxford.

McCrae, R. R. and John, O. P. (1992). An introduction to

the ﬁve-factor model and its applications. Journal of

personality, 60(2):175–215.

Mehrabian, A. (1996a). Analysis of the big-ﬁve personality

factors in terms of the pad temperament model. Aus-

tralian Journal of Psychology, 48(2):86–92.

Mehrabian, A. (1996b). Pleasure-arousal-dominance: A

general framework for describing and measuring indi-

vidual differences in Temperament. Current Psychol-

ogy, 14(4):261–292.

Mehrabian, A. (1997). Comparison of the PAD and PANAS

as models for describing emotions and for differentiat-

ing anxiety from depression. Journal of Psychopathol-

ogy and Behavioral Assessment, 19(4):331–357.

Mehrabian, A. and Russell, J. A. (1974). An approach to

environmental psychology. MIT Press.

Ortony, A. (2003). On Making Believable Emotional

Agents Believable. In Trapple, R. P., Petta, P., and

Payer, S., editors, Emotions in Humans and Artifacts,

chapter 6, pages 189–212. MIT Press.

Ortony, A., Clore, G. L., and Collins, A. (1988). The Cog-

nitive Structure of Emotions. Cambridge University

Press.

Ortony, A., Norman, D. A., and Revelle, W. (2005). Af-

fect and Proto-Affect in Effective Functioning. In Who

needs emotions. The brain meets the machine., pages

173–302. J.M. Fellous & M.A. Arbib.

Pessoa, L. (2008). On the relationship between emotion and

cognition. Nature Reviews Neuroscience, 9(2):148–

158.

Ryckman, R. M. (2007). Theories of Personality. PSY 235

Theories of Personality Series. Thomson/Wadsworth.

Santos, R., Marreiros, G., Ramos, C., Neves, J., and Bulas-

Cruz, J. (2010). Using personality types to support

argumentation. In Argumentation in Multi-Agent Sys-

tems, pages 292–304. Springer.

Scherer, K. R. (2001). Appraisal considered as a process of

multilevel sequential checking. Appraisal processes

in emotion: Theory, methods, research, 92:120.

Schwenk, C. R. (1984). Cognitive simpliﬁcation processes

in strategic decision-making. Strategic management

journal, 5(2):111–128.

Seo, M.-G., Goldfarb, B., and Barrett, L. F. (2010). Af-

fect and the framing effect within individuals over

time: Risk taking in a dynamic investment simulation.

Academy of Management Journal, 53(2):411–431.

Sewell, M. (2010). Emotions Help Solve the Prisoner’s

Dilemma. In Behavioural Finance Working Group

Conference: Fairness, Trust and Emotions in Finance,

London, pages 1–2.

Swope, K. J., Cadigan, J., Schmitt, P. M., and Shupp, R.

(2008). Personality preferences in laboratory eco-

nomics experiments. The Journal of Socio-Economics,

37(3):998–1009.

Tucker, A. (1983). The mathematics of Tucker: A sam-

pler. The Two-Year College Mathematics Journal,

14(3):228–232.

Weiss, G. (1999). Multiagent systems: a modern approach

to distributed artiﬁcial intelligence. MIT press.

Zadra, J. R. and Clore, G. L. (2011). Emotion and percep-

tion: The role of affective information. Wiley Interdis-

ciplinary Reviews: Cognitive Science, 2(6):676–685.

BuildingEmotionalAgentsforStrategicDecisionMaking

397