INTEGRATION OF AN EMOTION MODEL

TO THE BOARD GAME “INTRIGE”

Jan-Marc Ehrmann,

Andreas D. Lattner,

Gabriela Lindemann

and Ingo J. Timm

Goethe-Universit

at Frankfurt am Main, Institute of Computer Science

Information Systems and Simulation, P.O. Box 11 19 32, D-60054 Frankfurt/Main, Germany

Humboldt-Universit

at zu Berlin, Department of Computer Science

Artiﬁcial Intelligence Laboratory, Unter den Linden 6, D-10099 Berlin, Germany

Keywords:

Emotion model, Emotion-based strategy selection, Board game.

Abstract:

One important issue in gaming design is to let the player’s opponents appear as humanlike as possible. The

underlying assumption of our work is that integrating artiﬁcial emotions to computer players will lead to a

higher diversity of game situations and thus, to more interesting and joyful games. In this paper, we integrated

an emotion model to agents playing the board game “Intrige”. The current emotional state inﬂuences the

behavior of the agent. In experiments we have shown that success of agents depends on the used emotional

states and the varying constellations of opponents lead to different distributions of emotional states.

1 INTRODUCTION

The gaming industry aims to design computer games

providing a convincing projection of the reality. One

important issue in gaming design is to let the player’s

opponents appear as humanlike as possible. Emotions

are one important aspect attributed to human beings.

The underlying assumption of our work is that inte-

grating artiﬁcial emotions to computer players will

lead to a higher diversity of game situations and thus,

to more interesting and joyful games. The underlying

research question is if agents should be provided with

emotions at all, and if so, for which kind of applica-

tions it could be useful.

The objectives of this research are to select an

emotional model, integrate it into the behavior de-

cision process of agents, and investigate how these

emotions inﬂuence the results in a given environment.

In order to examine the differences of emotional

agents the players will be playing an implementation

of the existing board game “Intrige”. The investiga-

tion is about to ﬁnd out whether emotional agents

show any different behavior in the interaction with

other players and if there are tendencies in the emo-

tional development in certain combinations. In addi-

tion, the success rates of the different strategies will

be compared.

Although the deﬁnition of emotions still varies

there are generally accepted approaches (Zimbardo,

1995). (Kleinginna and Kleinginna, 1981) deﬁne

emotions as a complex pattern of changes including

psychological arousal, affection, cognitive processes

and behavior. They state that emotions occur when

an individuum percepts a certain situation to be of

importance. A newer and more complex model de-

ﬁnes emotions as a combination of certain tuples in

an Cartesian coordinate system (Russell, 1980). Rus-

sell’s circumplex model of affect distinguishes be-

tween the activation and pleasure axis. He deﬁnes

28 different states of affect, while a modiﬁed ver-

sion of the circumplex model (Schmidt-Atzert, 1996)

uses only eight states which will be described later

on. While criticism claimed that states like tired-

ness cannot be regarded as emotions (Schmidt-Atzert,

1996), the circumplex model allows a fair distinction

between certain emotional states.

2 THE BOARD GAME “INTRIGE”

Intrige (engl.: intrigue) is a German board game

where every player manages a court with four re-

gions of different values. The player’s actions are

twofold: Every court owner has a number of schol-

ars (two priests, two writers, two physicians, and two

547

Ehrmann J., D. Lattner A., Lindemann G. and J. Timm I. (2009).

INTEGRATION OF AN EMOTION MODEL TO THE BOARD GAME "INTRIGE".

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence, pages 547-552

DOI: 10.5220/0001768805470552

 SciTePress

Figure 1: “Intrige” board, sending out two scholars.

natural scientists) he wants to place in the other courts

as well as he has to house scholars from other players

in his four regions. The goal of the game is to earn

as much money as possible. During the game, players

negotiate about which scholar they are willing to ac-

cept or reject on their house positions aiming to earn

the most money at the end by either collecting cred-

its from their placed scholars or from the negotiations

when accepting or rejecting the opponents’ scholars.

Since the game speciﬁcally allows every kind of con-

tract between the players and often forces the player

to break the contracts, it can be considered to possess

the tendency to induce emotions.

Each game lasts ﬁve rounds in which each player

has to do the three following actions if possible:

1. collecting credits from the bank

2. accepting or rejecting pending scholars

3. sending out two of his scholars

The simulation implements almost every rule of the

game. However, the negotiations are currently based

on the amount of the paid bribery only.

Fig. 1 illustrates a situation where player red sends

out two of his scholars. He bribes player green with

7000 credits to get position 6000 and bribes player

yellow with 9000 to get position 10000. Paying bribes

does not guarantee that agreements are kept, i.e., play-

ers green and yellow are free to put the scholars in any

of the regions independent of the negotiation.

The game’s rules additionally have some restric-

tions, e.g., every court can only host one scholar of

each type and each region has only place for one per-

son. For more details we refer to the Intrige game’s

rules (Dorra, 2003).

Figure 2: Emotional states.

3 “EMOTIONAL” AGENTS

The emotional agents designed for this simulation im-

plement Russell’s simpliﬁed circumplex model with

a certain playing strategy in accordance to their game

experience. After experiencing emotionally relevant

incidents their emotional state can vary and therefore

their playing strategy can change. This section illus-

trates the origins of the emotional model and the dif-

ferent strategies, deﬁned for each state.

3.1 Emotion Model

The emotion model is based on the circumplex model

(Schmidt-Atzert, 1996) where eight different emo-

tional states are listed: elated, excited, happy, con-

tent, tired, bored, sad, and angry. In addition a ninth

state called neutral was added knowing as a starting

point to investigate the changes in emotional states.

By differing the two values activation and pleasure

into three categories named low, neutral and high, it

is always possible to determine the emotional state of

the agent. Fig. 2 shows how these different emotional

states result from the different intensity values of acti-

vation and pleasure. Threshold values (±20) indicate

the regions’ borders of the different emotional states.

3.2 Agent Behavior

For the scenario outlined in the previous section we

developed an agent architecture which contains an ex-

plicit model of emotions. The underlying idea is to

enable an agent to select speciﬁc strategies in its be-

havior, i.e., bribing strategy, w.r.t. its emotional state.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

548

Table 1: Action types and substrategies.

Action Type Potential Sub-Strategies

conﬂict handling avoid, invoke, random

sendout position selection high, low, random

bribe value selection high, low, normal

accepting scholar preference old, new, random

The reaction of the other players as well as the suc-

cess of the agent’s actions are used for transforming

the emotional state:

• sendout of scholar

accepted: increase activation

banished: decrease activation

• balance, i.e., earnings minus expenses

increased: increase pleasure

decreased: decrease pleasure

The emotion model as well as the transformation rules

introduced above are used for determining and chang-

ing an emotional state of an agent. The agent se-

lects its behavior in accordance to its emotional state.

The mapping of state to behavior is implemented by a

strategy concept, i.e., one strategy is assigned to each

emotional state. Without consideration of emotions,

an agent has to perform different types of actions. It

has to be decided where to send a scholar, amount of

bribery, where to place foreign scholars, and how to

handle conﬂict situations. Therefore the strategy con-

cept consists of an allocation of one sub-strategy to

each of the four major action types (cf. Table 1).

The ﬁrst three action types deal with sending out

a scholar. In the ﬁrst step, the agent has to select

the destination court. In this process, conﬂicts can

be avoided or provoked; the agent can also select the

court by random. After court selection, it is decided if

the scholar is sent to a region with high, low, or ran-

dom value. The third step is to calculate the bribery

value, namely high, low, or normal. Normal bribery

equals the value of the region. Finally, the accepting

scholar preference addresses the handling of foreign

scholars in the own court. The substrategy old means

that in conﬂicting situations the scholar who entered

ﬁrst is accepted while new accepts the latest.

Table 2 shows the combination of activation and

pleasure values in accordance to the resulting emo-

tional state and corresponding strategy. The strategies

of the agents are divided into three categories:

• upper strategies: agents willing to pay large

amounts of bribe (excited, elated, happy)

• lower strategies: agent paying small amounts of

bribe (content, sad, angry)

• normal strategies: agent paying the position value

(tired, bored, neutral)

3.3 Agent Types

Each emotional state is combined with a certain strat-

egy which is basically connected to the way how a

player is sending out its scholars, how much credits

he is willing to pay, and how he deals with requests

of other players. Altogether the simulation consists

of eleven different types of agents, nine of them play

one certain emotional state, one is playing randomly

(random agent) and one plays all of the emotional

states in accordance to his activation and pleasure val-

ues (emotional agent). Additionally, a human player

has been implemented for testing purposes.

4 EVALUATION

In order to investigate the simulation, different groups

of conﬁgurations have been deﬁned which determine

what agent types participate in one set of thousand

games. The number of players is set to four and for

each position one out of the eleven agent types can

be chosen. Additionally, for the emotional agents, it

is possible to select if the emotional state should be

reset at the end of every game.

In the experiments, we focused on the observa-

tion of one agent in the game setting. Every conﬁg-

uration is described by a notational string of the type

031-1[ti](2)3r-nr-01 where “031” represents the ID of

the conﬁguration “1[ti]” indicates that one emotional

player (in this case the tired player) on Position “(2)”

plays against three random player (“3r”) and the emo-

tion values will not be reset after the game (“nr”) and

the run number “01” has been investigated.

For evaluation purposes, different values are cap-

tured: The number of won games (i.e., highest cred-

its), the credits at the end of the game, and the emo-

tional state (for the emotional agent).

4.1 Calibration

The evaluation of the simulation is based on a cali-

bration. The calibration should guarantee that each

emotional agent can reach as many emotional states

as possible and at the same time being neutral for a

fairly amount of times after many games. In order to

evaluate the simulation the sensible bribe value had

to be re-adjusted by adding the suitable multiplying

values. Fig. 3 shows the distributions of emotional

states before and after an experimental reconﬁgura-

tion was accomplished (without reset of emotional

states). The calibration values have been identiﬁed

empirically. The charts show how often a player is

INTEGRATION OF AN EMOTION MODEL TO THE BOARD GAME "INTRIGE"

549

Table 2: Activation and pleasure of emotional states.

Emotional state Activation Pleasure Conﬂict handling Sendout position Bribe value Accepting scholar

neutral neutral neutral random random normal random

excited high neutral invoke high high random

elated high high invoke high high new

happy neutral high random random high new

content low high avoid low low new

tired neutral low avoid low normal random

bored low low avoid low normal old

sad low neutral random random low old

angry high low invoke high low new

Figure 3: Emotional states before and after calibration.

in a certain emotional state after one game (repeated

1000 times).

4.2 Hypotheses

In this work three different hypotheses have been

claimed:

1. The different strategies vary in their success, i.e.,

there are strategies which are more successful

than others in certain conﬁgurations.

2. Emotional agents are inﬂuenced differently w.r.t.

their emotional state depending on the set of op-

ponents.

3. Given a sufﬁcient number of games, the results

using emotional agents vary from the results using

only random players.

In order to proof the ﬁrst hypothesis a conﬁguration

has to be found where one player has more success

in the sense that he wins the game more often than

another player does in another conﬁguration.

The second hypothesis states that the artiﬁcial

emotional environment leads to different distributions

of emotional states. Given three different strategy

groups (upper, lower, and normal), hypothesis two

will be proven, if there exists a conﬁguration in each

strategy group where the emotional agent is in more

than 50% of the games in a certain emotional state.

The third hypothesis states that the introduction of

emotional agents changes the outcome of the game

compared to conﬁgurations where only random play-

ers participate. While the ﬁrst hypothesis aims to

show that the simulation is a “real” game in the sense

that different strategies lead to different outcome, the

second hypothesis claims that real tendencies of emo-

tional states can be found among the artiﬁcially cre-

ated emotions. The third hypothesis justiﬁes the use

of emotional players by showing the variety arising

from their game results and is proven if

1. The root square deviation of the credits of the in-

vestigated conﬁgurations is bigger than when ran-

dom players compete (sc ≥ 50000)

2. There are conﬁgurations where the amount of vic-

tories is almost equally balanced among the play-

ers (sv ≤ 20)

3. Every emotional state can win in a certain conﬁg-

uration

4. Every position can be victorious

4.3 Results

Among the so-called lower strategies, meaning a

strategy in which the emotional agent is not willing

to pay much bribe, there are two strategies which

were very successful even when running the conﬁg-

uration several times (in this case ten times). Com-

ICAART 2009 - International Conference on Agents and Artificial Intelligence

550

Table 3: Distribution of two successful conﬁgurations.

059-1e(1)3[ha]-nr 052-1e(2)3[ex]-nr

871, 841, 841, 843, 852, 794, 808, 821, 822, 818,

875, 857, 853, 846, 844 819, 828, 822, 821, 821

Table 4: Maximal number of emotional states with corre-

sponding conﬁguration.

conﬁguration state value

132-1e(4)3[n]-r neutral 538

072-1e(2)3[bo]-nr elated 115

070-1e(4)3[ti]-nr excited 660

150-1e(2)3[ti]-r happy 347

051-1e(1)3[el]-nr content 995

137-1e(1)3[an]-r sad 443

161-1e(1)3[an]-r bored 234

164-1e(4)3[an]-r tired 242

078-1e(4)3[sad]-nr angry 375

Table 5: Emotional state distributions for three selected

conﬁgurations.

051-

1e(1)3[el]-nr

070-

1e(4)3[ti]-nr

132-

1e(4)3[n]-r

993, 987, 989, 606, 625, 598, 530, 516, 484,

990, 987, 988, 588, 599, 580, 506, 514, 520,

988, 991, 619, 600, 512, 522,

986, 990 580, 629 507, 525

paring these two successful strategies where the emo-

tional agent on player position 1 archives generally

more than 800 victories in 1000 games, the results of

the students t-test states that the two value sets are sta-

tistically signiﬁcant differently and hence hypothesis

1 is proven. Table 3 shows the distribution of the two

successful conﬁgurations 059 and 052.

Regarding the different emotional states in which

the emotional agent can end up in accordance to

the conﬁguration, Table 4 shows nine conﬁgurations

where (in comparison to all other conﬁgurations) the

player ended up in a certain state more often than any

other conﬁguration.

By choosing conﬁguration 051 of the upper strate-

gies (willing to pay much credits), conﬁguration 070

of the lower conﬁguration group and conﬁguration

132 of the normal strategies (paying normal bribes),

one conﬁguration of each group could be found where

more than half of the games the agent ends up in a

certain state. The execution of the Student’s t-test sta-

tistically proved that these values are signiﬁcantly dif-

ferently proving hypothesis 2. Table 5 shows the ten

different emotional state distributions for all three se-

lected conﬁgurations.

In order to prove that the results of the simulation

Table 6: Distribution of won games.

Adam Betty Carry Dylan RSD

249,45 263,86 248,89 238,39 173,97

vary from the games where only random agents par-

ticipate, certain pre-deﬁned requirements have been

tested. The hypothesis could be veriﬁed, because con-

ﬁguration 4e-r and 4e-nr (4 emotional players com-

pete with and without reset) had a root square devia-

tion (RSD) of the credits sc ≥ 50000 and there were

more than one conﬁguration with with sv ≤ 20 (168-

4[ha]-nr with a value of 14.02). In order to illustrate

the variety of the simulation, Table 6 sums up the

winning distribution among all players (ordered by

their position) and showing the root square deviation

(RSD) of the investigated set. The third requirement

can be deduced by referring to the results from this

table. The fourth requirement has also been proven so

that hypothesis 3 is proved as well.

5 RELATED WORK

Simulation affection in a computer environment was,

for instance, the approach D

orner chose when design-

ing PSI, a learning software agent designed to survive

on an “island” by investigating his surroundings. Al-

though PSI was designed to experience many differ-

ent “needs”, it has yet not been able to interact with

other PSI agents (D

orner, 2002).

The software agent Max, developed at the Univer-

sity of Bielefeld (Boukricha et al., 2007), is even able

to perceive emotions of the other players and there-

fore can be considered to be empathic. In an investi-

gation, Max is playing a card game against a human

player and is differently reacting on the test persons

emotions, giving the test person the opportunity to

evaluate Max reactions. Max is yet not able to change

his game play according to his emotions.

(Reilly and Bates, 1992) introduce the emotion

model Em which has been developed within the Oz

project. The project aimed at the development of tech-

nologies for interactive ﬁction and virtual realities.

The emotion model for the agents in the Oz world is

based on an cognition-based model by Ortony, Clore,

and Collins (OCC). The adapted model used by Reilly

and Bates consists of 14 emotions like joy and dis-

tress. The activation of emotions can have different

causes, e.g., depending on goal success or failure. The

behavior decision process is inﬂuenced by a set of fea-

tures which are controlled by functions based on the

emotional state (Bates et al., 1992).

(Camurri and Coglio, 1998) present an architec-

INTEGRATION OF AN EMOTION MODEL TO THE BOARD GAME "INTRIGE"

551

ture for emotional agents. The authors propose an

architecture including an emotional state that evolves

over time. The emotional changes depend on the input

and create an output possibly inﬂuencing rational or

reactive behavior of the agent. Two two-dimensional

“emotional spaces” are presented where emotional

stimuli move the current position of the emotional

state. While the motivation of this approach is similar

to the one in our approach (trying to create more re-

alistic or entertaining behavior), Camurri and Coglio

aim at an emotional architecture for robots with real-

time interaction.

An approach to modeling emotional BDI agents is

presented by (Pereira et al., 2006). They introduce the

multi-modal logic EBDI where an emotional compo-

nent inﬂuences the interactions of beliefs, desires, and

intentions. In their paper, besides the basic emotional

model they provide a speciﬁcation of a fearful emo-

tional BDI agent. Pereira et al. consider threats and

unpleasant facts triggering fear.

(Steunebrink et al., 2007) introduce a formaliza-

tion of the OCC model of emotions mentioned above.

They introduce a logical language (and its semantics)

which they have used to formalize qualitatively all 22

emotional states of the OCC model.

6 CONCLUSIONS

The work proved that there are conﬁgurations which

are signiﬁcantly less successful in their environment

than others and that the artiﬁcially created emotions

are indeed signiﬁcantly inﬂuencing each other in the

context of the gaming environment. Although not yet

completely determinable the idea of game variety has

been investigated by guaranteeing that the new emo-

tional agents affect the game play severely.

Altogether the work achieved to construct a simu-

lation of artiﬁcial emotions in a gaming environment.

In order to investigate the plausibility of certain emo-

tional reactions a survey on the human player (like

Max) could be interesting. A competition with human

players could lead to interesting results, because it is

yet not sure that lower strategies might loose of their

strength when competing with human players (who

would accept an extremely small bribe?). Another

possible step is to adjust the complexity of the bribing

system by offering the players to bargain much more

as they like to.

Further interesting extensions could address the

integration of promises (in the sense of commitments)

as additional means for negotiations– and, of course,

the satisfaction or violation of such promises. In the

current emotional agents, each agent has only one

general emotional state. It would also be interesting

to introduce “directed” emotions, i.e., emotions w.r.t.

an individual player who has acted in a pleasant or

unkind way.

REFERENCES

Bates, J., Loyall, A. B., and Reilly, W. S. (1992). An ar-

chitecture for action, emotion, and social behavior.

In Selected papers from the 4th European Workshop

on on Modelling Autonomous Agents in a Multi-Agent

World, Artiﬁcial Social Systems, volume 830 of LNCS,

pages 55–68, Berlin. Springer-Verlag.

Boukricha, H., Becker, C., and Wachsmuth, I. (2007). Sim-

ulating empathy for the human max. In 2nd Interna-

tional Workshop on Emotion and Computing, in con-

junction with the german Conference on Artiﬁcial In-

telligence, pages 22–27, Osnabr

uck, Germany.

Camurri, A. and Coglio, A. (1998). An architecture for

emotional agents. IEEE MultiMedia, 5(4):24–33.

orner, D. (2002). Die Mechanik des Seelenwagens. Hans

Huber Verlag, Bern.

Dorra, S. (2003). Intrige. AMIGO Spiel + Freizeit GmbH.

Game’s rules (German).

Kleinginna, P. R. and Kleinginna, A. M. (1981). A catego-

rized list of motivation deﬁnitions with a suggesting

for a consensual deﬁntion. Motivation and Emotion,

5:263–291.

Pereira, D., Oliveira, E., and Moreira, N. (2006). Modelling

emotional BDI agents. In Dunin-Keplicz, B. and

Verbrugge, R., editors, Formal Approaches to Multi-

Agent Systems (ECAI Workshop Proceedings), pages

47–62, Riva del Garda, Italy.

Reilly, W. S. and Bates, J. (1992). Building emotional

agents. Technical Report CMU-CS-92-143, School of

Computer Science, Carnegie Mellon University, Pitts-

burgh, PA.

Russell, J. A. (1980). A circumplex model of affect. In Jour-

nal of Personality and Social Psychology: 39, pages

1161–1178.

Schmidt-Atzert, L. (1996). Lehrbuch der Emotionspsy-

chologie. W.Kohlhammer, Stuttgart.

Steunebrink, B. R., Dastani, M., and Meyer, J.-J. C. (2007).

A logic of emotions for intelligent agents. In Holte, R.

and Howe, A., editors, Proceedings of the 22nd AAAI

Conference on Artiﬁcial Intelligence (AAAI 2007),

pages 142–147, Vancouver, Canada. AAAI Press.

Zimbardo, P. G. (1995). Psychologie. Spektrum Verlag,

Berlin, 6th edition.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

552