Content uploaded by Mannes Poel
Author content
All content in this area was uploaded by Mannes Poel on Jun 30, 2014
Content may be subject to copyright.
Simulation of emotions of agents in virtual
environments using neural networks
Aard-Jan van Kesteren
Rieks op den Akker
Mannes Poel
Anton Nijholt
Department of Computer Science
University of Twente
Enschede, the Netherlands
e-mail : anijholt@cs.utwente.nl
Abstract
A distributed architecture for a system simulating the emotional state of an agent acting
in a virtual environment is presented. The system is an implementation of an event-
appraisal model of emotional behaviour and uses neural networks to learn how the
emotional state should be influenced by the occurrence of environmental and internal
stimuli. A part of the modular system is domain-independent. The system can easily be
adapted for handling different events that influence the emotional state. A first
prototype and a testbed for this architecture are presented.
1 Introduction
This paper presents results of ongoing research on human-computer interaction with intelligent
conversational agents in virtual environments. It reports work in a project aiming at the design and
agent-oriented implementation of a multi-user, multi-agent and multi-modal interactive
environment [Nijholt 1999].
Our interest in emotions and our objective to simulate emotional behaviour finds its primary
motivation in the hypothesis that avatars that somehow show emotions in the way they behave are
more believable than agents that lack these human qualities.
There is another motive for incorporating emotions in the design of synthetic agents; often
mentioned by AI-researchers. A. Damasio came with neurological evidence for the importance of
emotions in human decision processes [Damasio 1994]. The interaction between emotions and the
rational processes that underlie the making of decisions could possibly explain why humans are so
good in making decisions in a context of uncertainty. Answering the question whether machines can
have emotions, M. Minsky even stated: “The question is not whether intelligent machines can have
any emotions, but whether machines can be intelligent without any emotions” [Minsky 1986].
Hence the idea to give synthetic agents emotions in order to make it possible that they perform
whatever tasks they have in a more intelligent way. Whether there is some sense in this depends on
the notion of emotion one has in mind and on the analysis of the aspects involved in the process of
decision-making.
To model and to simulate all of the relevant aspects of emotions in a comprehensive system is a
project for years, so it shouldn’t come as a surprise that every group that’s conducting research on
the topic of emotional agents has made it’s own decisions as to what aspects are built into their
system, and what is left out for (hopefully) future development. Most groups [Reilly & Bates 1992,
El-Nasr & Yen 1988, Velásquez 1997] have taken a wide range of topics related to emotions (e.g.
emotion, facial expressions and emotional behaviour) into account for their emotional systems.
Until now we have only looked at the topic “simulation of emotions”, as we think it’s best to focus
on one topic at a time.
We think of an emotional agent as an agent that has among other things (e.g. believes, desires and
intentions) an emotional state. This emotional state can be altered by stimuli from the environment
or by stimuli from internal elements of an agent (e.g. decisions, future expectations, memory, etc.).
“Simulation of emotions” consists of the elements that appraise the stimuli and the processes that
take care of the dynamics of the emotional state. In principle, every kind of behaviour and internal
process of an agent can make use of the emotional state. For instance, the emotional state can be
used to steer the facial expressions and the decision process. Consequently, our agents don’t show
emotions yet. A user can only get an impression of what the agent “feels”, by looking at a
representation of the emotional state.
For our model of emotions, we have two basic concerns. We want our model to resemble the
emotional processes, as they exist within the human brain, as closely as possible and consequently
we don’t want to diverge too much from the leading theories of emotions. On the other hand we
also want to develop a computational approach, which can be easily used to implement an
emotional agent.
Our research is very much ongoing and some of the elements presented in this paper haven’t been
properly tested yet. This paper will give an overview of how we dealt with the problem of emotions
until now and what our approach for the future will be. Furthermore an architecture for the
simulation of emotions will be introduced.
First the relevant models of emotions will be discussed shortly. After that we will globally describe
what our approach is, followed by a description of the environment that we used as a testbed.
Subsequently our distributed architecture will be introduced, by firstly explaining the architecture
as a whole and secondly by discussing the various subparts in more detail. A description of a first
prototype will be given next to further clarify the architecture and to present some first test results.
The paper will be concluded by a comparison with another model and some conclusive remarks.
2 Theories of emotion
In this section, two theories on emotion will be introduced. How we used these theories for our
model will be explained in the next section.
Event appraisal models
1
have as assumption, that it’s the subjective, cognitive interpretation of
events that explains the experienced emotions.
Event appraisal models mainly focus on the question how events are appraised and in which
direction (and with what velocity) the emotional state is likely to shift. We think of this as the
emotional meaning of an event. Usually event appraisal models don’t model the dynamics of the
emotional state: if an event is appraised, what will then become the new emotional state? For an AI-
model of emotion this question is as important as a natural appraisal of events.
1
For instance [Ortony, Clore & Collins 1988] and [Roseman, Jose & Spindel 1990]
One of the leading event appraisal models is the model of Ortony, Clore and Collins (The OCC
Model) [Ortony, Clore & Collins 1988]. The model only looks at emotion types. Each type contains
a large number of emotional states, which may differ in intensity (e.g. the emotional states: worried,
scared and terrified belong to the emotion type fear), but may also differ because of other reasons,
such as the cause of the emotion (e.g. the emotional state heartache differs from other emotional
states belonging to the emotion type distress). By focusing on emotion types instead of emotions,
the whole field of emotions gets more comprehensible.
In order to appraise a particular event, there are three different aspects that can be focused on. Every
aspect has a different set of emotion types attached to it. We will elaborate a bit on the different
aspects by means of an example. Imagine that someone hears that his sister-in-law has just killed
one of his children. If that someone focuses on the consequences of the event, he’ll probably
experience distress. If he focuses on the action of the agent (the sister-in-law), he’ll probably
experience reproach. And if he looks at the aspects of an object (aspects of the sister-in-law) he’ll
probably experience hate.
Associated with each aspect is a different kind of knowledge representation. If one focuses on the
consequences of an event, goals are of importance. If one focuses on the action of an agent,
standards are of importance and if one focuses on the action of an agent, attitudes are relevant.
A strong point of the OCC model, which makes it particularly useful for designing an AI-model of
emotion, is that the model includes a complete set of variables that influence the intensity of the
emotions. The variables can be global, which means that the variable influences all of the emotion
types, or local, which means that the variable influences only some of the emotion types. We will
shortly introduce the (in our opinion) three most important local variables. The variable desirability
is important if one focuses on the consequences of an event. The variable praiseworthiness is
important if one focuses on the actions of an agent and the variable appealingness is important if
one focuses on the aspects of an object.
Besides the event appraisal models, there are more theories of emotions we found interesting and
useful. One of those is Frijda’s notion of emotion [Frijda 1986]. According to him, emotion is
action readiness change, which refers to the inner disposition (and the absence) for performing
actions and achieving relational change, as it exists at a particular moment. The experience of
emotion largely consists of experienced action readiness or unreadiness: an impulse to flee, strike or
embrace; or lack of impulse, apathy, listlessness. So, emotion is not the same as the feeling of
emotion. A consequence of this is, that an explicit representation of emotions is unlikely to exist.
3 Our approach
Our research goal is to develop a human like agent that shows natural emotional behaviour. Yet, as
we didn’t want to start too ambitiously we are firstly focusing on simple agents in a simulated
environment. The agents and the environment shall be introduced in the next section.
Our model is very much an implementation of the more theoretical OCC model. First of all, only
events can alter the emotional state. Furthermore, the representation of emotion consists of the same
emotion types the OCC model distinguishes, the three aspects of events also play a central role in
our model and the same variables (the properties of events that are important for emotions) as in the
OCC model will be used. Furthermore, the agents will contain a representation of goals, standards
and attitudes.
The current state of research on emotions is, that a lot of work has been done on philosophical
questions like: “What are emotions?”, “What influences emotions?”, “Which emotions can be
distinguished?” and “What is the connection between cognition and emotion?”, but no complete
quantitative models of emotions (besides some AI-models) have been developed yet. Apparently
it’s very difficult to describe the emotional behaviour in a complete and precise manner. On the
other hand we noticed that it’s relatively easy to imagine in practical situations what would be
natural behaviour.
Therefore we chose an approach in which the system has to learn about emotional processes and
rules from examples provided by a trainer. Most other research groups have used a completely
opposite approach. In their systems the designer has to define the emotional processes and rules,
from which natural emotional behaviour has to originate. We’ll use the common terms top-down
approach for our approach and bottom-up approach for the other approach (Figure 1 gives a
schematic representation of the two concepts).
Emotional behaviour
Emotional processes & rules
top-down
bottom-up
Figure 1 A top-down approach vs. a bottom up approach
For our approach trainingsdata is needed, which will be obtained through annotation. We realize
that annotation is a time-consuming and tedious task and therefore we defined the architecture in
such a way, that a minimal amount of trainingsdata is required.
Another advantage of a top-down approach above a bottom-up approach is, that it’s more flexible;
the same system can be used for more applications, which is for instance practical if one wants to
design agents with different personalities. Using a different trainingsset for each personality is an
easy way to do this. Of course the same effect may be obtained in a bottom-up approach by
changing some of the parameters, but we strongly doubt if parameters can be defined in such a way
that it offers the same flexibility as a top-down approach does.
As we want to design a system that only needs a small amount of trainingsdata, it’s especially
important that the system is able to generalize well. This is one of the proven qualities of neural
networks, and accordingly we will use those for our system. For the emotion domain, it isn’t very
important that the system works very accurately in a quantitative way (the chosen numerical
representation of emotions is a rough estimate anyway), as long as it performs well in a qualitative
way and reasonably well in a quantitative way. Therefore relatively small neural networks can be
used, which should enhance the generalization capacities of the system even more.
As Frijda’s theory of emotions clearly suggests, an explicit representation of emotions is unlikely to
exist within the human brain. We do not only think that an explicit representation of emotions is
unlikely from a cognitive point of view, but we also think that a wrongly chosen representation
might have a strong negative effect on the capacities of the system to show a wide range of
emotional phenomena. By using recurrent neural networks, the system can learn an optimal,
implicit representation of emotions itself, which we see as a major advantage of a connectionist
approach.
4 The environment and the agents
In order to develop and test a system that simulates emotions, an environment is needed, which is
inhabited by agents that can have emotions. As we want to model a broad range of emotions, the
environment and the agents must satisfy a number of requirements.
As we make use of the OCC model as basis for our AI-model of emotion, it’s important that the
agents have an explicit or implicit representation of goals, standards and attitudes
1
. Agents should
also be able to translate observations into terms of these three concepts. In practice this last
requirement has the following consequences:
• Agents should be able to see if an event satisfies a particular goal or has a positive or
negative effect on the probability that a particular goal will be satisfied (important for the
emotion types joy and distress).
• Agents must be able to reason about the future and should be able to change their
expectations of the future as a consequence of events. Expectations about goals are
important for emotion types such as hope and fear.
• Agents must have some kind of memory about previous expectations and should be able to
compare new events to these previous expectations (important for the emotion types relief,
fears-confirmed, satisfaction and disappointment).
• The same kind of reasoning the agents must be able to do for themselves, they must be able
to do for other agents also (important for emotion types happy-for, resentment, gloating and
pity).
• Agents must be able to compare actions of themselves or of other agents to their standards
(important for emotion types pride, shame, admiration and reproach).
• Finally, agents must be able to compare aspects of objects or agents to their attitudes
(important for emotion types love and hate).
The future must be partly uncertain for the agents. If it’s precisely known what’s going to happen in
the future, emotions such as hope and fear cannot exist. A way to guarantee this is by making the
environment nondeterministic. Another important property of the environment is, that it contains
multiple agents, as some emotions depend on actions of other agents or on the consequences of
events for other agents.
As the user has to annotate natural emotional behaviour, it’s important that the user is able to put
himself in the position of an agent. A user can only do this, if the behaviour of the agent is
believable enough and if the user has the exact same knowledge as the agent. A final requirement of
the environment is, that the situations are complex enough, such that interesting emotional
behaviour can arise.
We set out to define and implement an environment, which satisfies all the previously stated
constraints. For the domain we were inspired by [Inoue, Kawabata & Kobayashi 1996]. In Figure 2
a picture of the domain can be seen.
1
We refer to [Reilly & Bates 1992] for a good example of an implementation of these concepts.
Figure 2 Picture of the environment
The domain is a gridworld containing grass, water pools that can be dry or contain water, apple
trees that can have apples growing and rocks, with possibly herbs growing on it. The status of the
trees, water pools and rocks constantly changes in a nondeterministic way.
Multiple agents inhabit the gridworld. One of the agents can be seen in the middle of the picture. An
agent can only see a small part of the world (the light part). Currently the agent sees one predator
(directly beneath the agent) and one tree with an apple (directly south-east of the predator). An
agent only knows where the visible agents and predators are and he knows the location of all the
trees, rocks and water pools, but the status is only known of the visible trees, rocks and water pools.
A trainer can only see the things an agent can see; in the light part, a trainer can see everything, just
like an agent, and in the dark part, a trainer sees only the locations of the trees, rocks and water
pools, as this is also knowledge that an agent possesses. Because of this, it’s easier for a trainer to
imagine what an agent feels. A trainer also knows how hungry, thirsty and healthy an agent is,
which events have occurred lately and what his current action is.
Agents need food and water, which can be supplied by apple trees and water pools. There are
predators that can be dangerous for the agents. An attack by a predator affects the health of an
agent. An agent can regain health by eating an herb.
To a large extent, the behaviour of a predator is random. The only exception is when a predator is in
the neighbourhood of food, water or an herb. Then it will stay there because it knows that agents
will come there sooner or later. This means that the presence of a predator is an indication that food,
water or an herb may be nearby. On the other hand the presence of food, water or an herb is an
indication that a predator may be nearby.
Agents are able to make decisions and have all the previously described capacities. There is a social
order between agents and they can choose (dependent on their character) either to follow the leader
of a group or to go their own way. Social grouping is a result of common concerns. An agent knows
how thirsty, hungry and healthy the agents in his group are.
5 The system
5.1 The architecture
In this section, the architecture for our system will be introduced. First a global overview of the
system will be given and after that the various concepts and subparts will be discussed in more
detail. For the system we have chosen a distributed architecture, as depicted in Figure 3. The
motivations for this particular architecture will be given after a brief introduction.
Emotional
State
Calculator
N
E
IV
N
E
I
V
N
E
IV
Event
E
v
e
n
t
I
n
f
o
E
v
e
n
t
I
n
f
o
E
v
e
n
t
I
n
f
o
...
Normalizer 1
Event
Appraiser 1
EIV
Event type 1
Normalizer 2
Event
Appraiser 2
EIV
Event type 2
Normalizer n
Event
Appraiser n
EIV
Event type n
Emotional
State
Figure 3 Schematic representation of the architecture
From a functional perspective, the task of the system is to convert an event with particular
properties into a new emotional state. In order to convert an event to a new emotional state, two
phases are distinguished.
The first phase is to appraise the emotional meaning of the event. This is the task of an Event
Appraiser. Each event belongs to maximally one event type (e.g. the event: An agent sees an apple
at a particular location while he has a particular desire for food belongs to the “Apple_spotted”
event type). For each relevant event type, one event appraiser is defined. The content of the event
information depends on the type of the event and can vary greatly from one event type to another.
To sum up, when an event occurs, first the event type of the event has to be established. Then it’s
exactly known which Event Appraiser has to be used and what kind of information is needed for
appraisal.
In the second phase, the Emotional State Calculator (the ESC) uses a numerical representation of
the emotional meaning of the event (firstly stored in an EIV and secondly stored in an NEIV; both
will be explained in detail later on) to calculate the new emotional state. A history of previous
emotional states and emotional events is implicitly stored within the ESC, which is implemented by
a recurrent neural network. The Normalizers form an interface between the Event Appraisers and
the ESC. The task of a Normalizer is to normalize the data in an EIV, such that the ESC can treat all
of the Event Appraiser in a uniform way. The Normalizers will also be implemented by neural
networks.
This approach has a few qualities:
• Conformity with emotion theory. As mentioned before, emotion theorists are inclined to
see the appraisal of events as a separate problem. In this architecture, the appraisal of events
is explicitly a subpart of the system.
• Conformity with human intuition. In our first annotation attempts, we noticed that if we
had to predict a new emotional state, given an event and the old emotional state, that we
were inclined to first look only at the event and imagine what kind of emotional
consequences this event might have and only after that we looked at the old emotional state
and predicted a new emotional state. Therefore the way the system works conforms at least
to our intuition and maybe also to the intuition of most people. This is an important
property, as this means that the annotation task (which will be explained in the section on
training) will be relatively easy.
• Trainability. The different subparts can be kept relatively small, as a result of which the
various neural networks can be trained more easily and with a smaller amount of
trainingsdata.
• Possibility of incremental development. This approach allows that events are added to the
system one-by-one. This is a big advantage from a software engineering point of view.
• Scalability. If the domain gets more complicated (meaning more types of events), the
amount of neural networks within the system will increase, but the complexity of the neural
networks will remain the same. This means that the scalability of the system shouldn’t be a
problem, if the amount of types of events increases. Whether the system is scalable in other
dimensions also (e.g. the amount of modelled emotions) is one of the open research
questions remaining.
• Reusability of the ESC. The ESC is domain-independent and therefore can be reused for
different agents in different domains.
5.2 Emotion Impulse Vectors
An Emotion Impulse Vector (an EIV) is a data structure especially suited for storing the emotional
meaning of an event. The structure of an EIV has been chosen as follows:
EIV = <ei
1
, ei
2
, ..., ei
n
>
in which every emotion type e
i
has a corresponding emotion impulse ei
i
. The emotion impulse ei
i
indicates how the emotion e
i
is likely to change. For instance a high positive value for ei
i
means that
the EIV probably will have a large positive effect on the emotion e
i
.
A Normalized Emotion Impulse Vector (an NEIV) is a normalized version of an EIV. How the
normalization functions, will be explained later on.
5.3 Event Appraisal
The task of an Event Appraiser is to assess the emotional meaning of an event to construct an EIV.
Every event and domain has its own properties and therefore an Event Appraiser should be
constructed separately for every event type. This way a priori knowledge about the event and
domain can be used in an optimal way. A consequence of this is, that an Event Appraiser can be
implemented by a neural network, but can just as well be implemented by a rule-based system, a
function or something else.
Defining or training an Event Appraiser that can immediately communicate with the ESC in such a
way that the behaviour of the total system is precisely as required, would mean a full understanding
of the ESC and would at least make defining or training the Event Appraiser a very difficult task.
Therefore we introduced the Normalizers. Their task is to convert an EIV to an NEIV such that the
behaviour of the total system works as required. As the Normalizers only have to do scaling
operations, and therefore only have to learn linear or near linear functions, the Normalizers can be
simple feedforward networks with only a few hidden neurons.
A nice property of using Normalizers is that the Normalizers can easily correct quantitative errors
of the Event Appraisers and the ESC. Therefore it’s not very important that the Event Appraisers
and the ESC work well in a quantitative way, just as long as they work well in a qualitative way.
The inputs of an Event Appraiser are the variables defined in the OCC model or properties of an
event that can be used to calculate the variables. The current action of the agent is a special kind of
property, which can be important to appraise an event. For instance, an agent is bound to appraise
the event of spotting a predator differently if he’s attacking than if he’s fleeing.
5.4 The Emotional State Calculator
The task of ESC is to calculate a new emotional state given a NEIV. Although the ESC is one
neural network, conceptually it can be seen as two different parts. The first part, the core of the
ESC, is a recurrent network and models all of the emotional behaviour. The state of the recurrent
network can be seen as an implicit emotional state. We will implement the implicit emotional state
with an Elman network [Elman 1990].
The second part of the ESC is a feed forward network, which has the implicit emotional state as
input. This part, the Emotional state monitor, has as task to translate the implicit emotional state,
which is very hard to understand and use to influence the behaviour, into an explicit emotional state.
The structure of the explicit emotional state has been chosen as follows:
Explicit emotional state = <e
1
, e
2
, …, e
n
>
in which e
i
denotes the intensity of an emotion type. Every e
i
refers to a different emotion type.
A schematic representation of the ESC is depicted in Figure 4.
Emotional state
......
...
NEIV
...
...
Emotional state
monitor
Explicit
emotional state
Implicit
emotional state
Figure 4 Schematic representation of the Emotional State Calculator. A
dashed line represents a time delay of 1.
5.5 Decay
The decay of an emotional state is a topic that has received special attention in most other AI-
models of emotion and is also a topic mentioned in [Ortony, Clore & Collins 1988] as being a
specific point of concern for a computational model of emotion. In our system the decay is modeled
as an event type with a corresponding Event Appraiser and Normalizer. An event of this type occurs
by definition after a fixed interval of time and sees to it that the emotional state eventually returns to
a state of rest. A property that determines the appraisal of this event is among other thing the current
(explicit or implicit) emotional state.
5.6 Training the system
The system is built out of three types of building blocks. For every type of building block a
different training strategy has to be used. The order in which the building blocks have to be trained
or defined is depicted in Figure 5.
Training the
Emotional State Calculator
Constructing an
Event Appraiser
Training a
Normalizer
Adding a new Event
Figure 5 Schematic overview of the trainingsprocess
First the Emotional State Calculator has to be trained. This is the most difficult part of the system to
train, as gathering trainingsdata is difficult. A temporal sequence of (NEIV, explicit emotional
state) pairs is needed to train the Elman network. The NEIVs can be generated randomly, but the
emotional states have to be annotated by hand.
There are two problems with this approach. First of all a reasonable amount of trainingsdata is
needed and secondly the needed data is of an abstract nature, as a result of which annotating even
one sample isn’t an easy task. But we feel that the benefits of this tedious work are large enough to
be profitable. If the Emotional State Calculator has been trained, adding events to the system is easy
and an Emotional State Calculator is domain-independent and therefore only has to be trained once.
The trainingsset doesn’t have to be perfect, as the learning algorithm should be able to filter out the
errors. Therefore the trainer can annotate quickly without worrying too much about errors, which
makes the annotation task a bit easier. Also, it shouldn’t be overlooked that a significant part of the
difficulties of training the ESC are inherent to the complexity of the problem of emotions. A nice
property of our system is that a major part of the complexity of emotions is modelled within a
domain-independent part of the architecture, so that the same problems don’t have to be solved over
and over again for every domain and agent.
After the ESC works satisfactory, event types can be added one-by-one to the system. For a new
event type, an Event Appraiser has to be trained or defined firstly. Again it’s only important that an
Event Appraiser works well in a qualitative way, which makes defining or training an Event
Appraiser relatively easy.
From a procedural perspective, training a Normalizer is a bit less straightforward than training the
ESC or an Event Appraiser. (EIV, NEIV) pairs are needed to train a Normalizer. The user cannot be
expected to annotate NEIVs, because this would mean among other things that the user has to know
exactly how the ESC works, which is not a realistic requirement. Therefore we propose the
following approach.
Instead of (EIV, NEIV) pairs, (Event Information, Explicit emotional state, State of the ESC) triples
have to be gathered. This can be done within the real domain. Every time an event of the relevant
event type occurs, the user has to annotate the new explicit emotional state. The event information
can be extracted from the event and the current state of the ESC can be copied easily. After the
triples have been gathered, the triples have to be converted to the (EIV, NEIV) pairs. It is easy to
convert the event information to an EIV, as the Event Appraiser is already trained or defined. An
NEIV can be found by using the inverse of the ESC. If this inverse is known, the explicit emotional
state (the output of the ESC) and the state of the ESC are enough information to calculate an NEIV.
Of course the inverse is not known, but using an iterative improvement algorithm (e.g. a simulated
annealer), can solve this final problem. This algorithm has to find the NEIV that explains the
explicit emotional state the best. If needed, a second criterion can be that the NEIV has to resemble
the EIV as much as possible.
Using this approach makes training a Normalizer easy for a user. Annotating an explicit emotional
state is easy and as the Normalizers are usually small networks, only a small amount of
trainingsdata is needed.
6 A first prototype
A first prototype of the system has been implemented. In this prototype the various subparts were
kept as simple as possible, such that we could examine whether the subparts are able to cooperate
and whether the system as a whole is capable of simulating the required behaviour. One of the
consequences of this is, that the ESC has not been implemented by the proposed Elman network
yet.
In the first version only two emotion types are modelled, joy and distress, and seven event types.
The two emotion types joy and distress were considered as one emotion type, the joy/distress
emotion type. If the intensity of this emotion type is negative, the agent experiences distress and if
the intensity is positive, the agent experiences joy. So, an EIV, an NEIV and the explicit emotional
state, all consist of only one value.
The ESC has been implemented as a simple linear function (and therefore the inverse is known).
The previous emotional state is one of the variables of the function. Although this simple function
suits us well in this phase of the research, we are convinced that such a simple approach won’t be
sufficient in the future and that the described approach with an Elman network is a more powerful
and a cognitively more credible approach.
The Event Appraisers were kept simple too. We will clarify this part of the architecture a bit more,
by focusing on the “New_predator_spotted” event type in detail. The events belonging to this event
type occur when the agent sees a predator for the first time.
As mentioned before, the variables from the OCC model are the only factors that influence the
intensity of an emotion type. Although more than one variable is defined in the OCC model, that
influence the emotion types joy and distress, only the variable desirability determines the appraisal
value in this version. The event information of the “New_predator_spotted” event type consists of
two properties: The current health of the agent (health) and the amount of predators the agent has
seen previously this turn (#seen_previously). If the health of the agent is low, the desirability of the
event should be lower, than if the health is high, as being near a predator while health is low is very
dangerous. The property #seen_previously should have a negative effect on the desirability of the
event, as for instance seeing a third predator is less desirable than seeing a second predator. Taking
all of this into account, the EIV for this event was defined as follows:
EIV = <desirability(health, #seen_previously)>
= <1 * health – 20 * #seen_previously – 100>
Only qualitative and no quantitative arguments and no were taken into account while drawing up
the formula and as a consequence the 1, 20 and 100 in the formula could for instance just as well
have been 1.5, 30 and 0.
The Normalizer is a simple feed forward network, with one hidden layer, consisting of only 2
neurons. Only 20 trainingsamples were used to train the Normalizer. The other Event Appraisers
and Normalizers were constructed in a similar way.
How to test the quality of the system is a hard question and a research project by itself. This version
was tested in an informal way. We checked whether the emotional behaviour shown by the system
corresponds to our ideas of emotions, which we also used for annotation. This was the case and
therefore we can conclude that the architecture looks promising, but a lot more experiments need to
be conducted before definite conclusions can be made.
7 Comparison with other models
Before defining our own models we studied other models. [Pfeifer 1988] gives a nice overview of
all of the AI-models of emotion until 1988. Since then, more interesting models were put forward.
[Reilly & Bates 1992] also based their model on the OCC model. In [El-Nasr & Yen 1988] fuzzy
logic was used, a very useful technique for the emotion domain we think. A system which we want
to discuss more thoroughly in this section is Cathexis [Velásquez 1997], as it resembles our model
in a couple of ways; it also has a distributed architecture and special care has been taken to allow
for maximum flexibility.
Cathexis is composed of a number of “proto-specialist” agents [Minsky 1986], each representing a
different emotion type. Within each proto-specialist, different sensors are monitoring both external
and internal stimuli for the existence of the appropriate conditions that would elicit the emotion
type represented by that particular proto-specialist. Input from these sensors either increases or
decreases the intensity of the emotion type. Parallel to this process, there’s always a decay process
going on, which sees to it that the intensity of the emotion type eventually returns to a state of rest.
The way the sensors function and how the sensors and the decay function affect the intensity can be
completely defined by a user. Not only the sensors can influence the intensity of an emotion, but
also other proto-specialists agents can, by providing either inhibitory or excitatory input. A
schematic overview of the system is depicted in Figure 6.
Both Cathexis and our approach have distributed architectures. The difference is, that in Cathexis
the emotions are scalable and in our architecture the complexity of the domain is scalable. The
advantage of making the emotions scalable is, that it makes it easier to implement a system within a
simple domain. But, as the domain gets more complex, the approach of Cathexis will experience
scalability problems we think, as all the proto-specialists have to perform increasingly complex task
then, while no provisions have been taken to deal with this complexity.
I
I
n
h
i
b
i
to
r
y
/
E
x
c
i
t
a
to
r
y
I
n
p
u
t
I
n
h
i
b
i
t
o
r
y
/
E
x
c
i
t
a
t
o
r
y
I
n
p
u
t
I I
Inhibitory /
Excitatory
Input
Anger Fear
Distress
Figure 6 The “proto-specialist” agents for anger, fear and distress. (I:
Intensity)
In our approach it’s difficult to train the ESC for a system with multiple emotion types. But after
that part has been done, adding more event types is easy.
The Cathexis system allows for a large degree of flexibility. Nevertheless we don’t think the
framework is powerful enough to model complex temporal emotional phenomena and phenomena
in which more than two emotions participate. Our approach does have these possibilities and
therefore is more powerful than Cathexis.
For Cathexis a bottom-up approach is used to make a system suitable for a particular domain. We
think that it will be very difficult to precisely define all the functions and parameters in such a way
that the system performs well in the case of a complex environment in which a lot of types of events
can occur. In our architecture the generalization capabilities of neural networks should make it
easier to handle complex environments and as a top-down approach is used, a full understanding of
emotional processes and the environment is not needed; the trainer only has to know in practical
cases, what the right emotional behaviour is.
8 Conclusion
An emotion theory based architecture for simulating emotions has been presented in this paper,
which can be used to develop emotional agents acting within complex environments. A first
prototype indicated that the various subparts of the distributed architecture are able to cooperate
well and that the system is able to show natural emotional behaviour.
In this first implementation, the Emotional State Calculator (the ESC) was kept very simple. For the
eventual system we have proposed a recurrent network of which the state of the network can be
seen as an implicit emotional state. A foreseen problem of our approach is training the recurrent
network, as it’s difficult to obtain sufficient trainingsdata. But we feel that this problem is largely
inherent to the complexity of emotions and that a nice property of our architecture is that a major
part of this complexity is concentrated within a domain-independent part of the architecture.
In the future we want to expand our prototype with more emotions and event types. Furthermore we
think that more research is needed on the ESC.
References
Damasio, A.R. (1994). Descartes’ Error: Emotion, Reason, and the Human Brain. New York: G.P.
Putnam.
Elman, J.L. (1990). Finding structure in time. Cognitive Science, 14, pp. 179-211.
El-Nasr, M.S., & Yen, J. (1988). Agents, Emotional Intelligence and Fuzzy Logic. Proceedings of
the IEEE NAFIPS ’98, Fl.
Frijda, N.H. (1986). The Emotions. New York: Cambridge University Press.
Inoue, K., Kawabata, K., & Kobayashi, H. (1996). On a Decision Making System with Emotion.
IEEE International Workshop on Robot and Human Communication, Tokyo, Japan, pp. 461-465.
Minsky, M. (1986). The Society of Mind. New York: Simon and Schuster.
Nijholt, A. (1999). The Twente Virtual Theatre Environment: Agents and Interactions. Proceedings
of the fifteenth Twente Workshop on Language Technology, pp. 147-165.
Ortony, A., Clore, G.L., & Collins, A. (1988). The Cognitive Structure of Emotions. Cambridge,
UK: Cambridge University Press.
Pfeifer, R. (1988). Artifical Intelligence Models of Emotion. In V. Hamilton, G.H. Bower, & N.H.
Frijda (Eds.), Cognitive Perspectives on Emotion and Motivation, Netherlands: Kluwer, pp. 287-
320.
Reilly, W., & Bates, J. (1992). Building emotional Agents, Pittsburgh, PA: Carnegie Mellon
University, Technical Rep. CMU-CS-92.143.
Roseman, I.J., Jose, P.E., & Spindel, M.S. (1990). Appraisal of Emotion-Eliciting Events: Testing a
Theory of Discrete Emotions. Journal of Personality and Social Psychology, 59(5), pp. 899-915.
Velásquez, J. (1997). Modeling Emotions and Other Motivations in Synthetic Agents. In:
Proceedings of the AAAI Conference 1997, Providence, RI, pp. 10-15.
