With the background reading done, the next step was to choose
an emotion theory which would be heart of the emotions component
in the agent architecture. A recognised and well-structured
theory on emotions was required so that the emotions constituent
of the agent architecture can be designed and implemented with
sufficient justification that emotions really works in this
manner. A emotion theory will also serve as a guide as to how
should they be represented and what are their characteristics.
Two possibilities surfaced at this point, Ortony, Clore and
Collins [1] semantic-based theory and Aaron Sloman's design-based
theory. In order to decide upon a theory to use, a short
discussion on emotions and an analysis of the two theories was
done.
Emotions is one of the most central and pervasive aspects of
human experience. Normal people experience a wide range of
emotions everyday. It is somewhat difficult sometimes to explain
in words what one's current emotional state is. It may even be a
combination of a number of emotional states. There are also no
stereotype for emotion experiences as it varies depending on the
individual and cultural differences.
The effects of emotions are clearly visible and well-recognised.
It not only deepened and enriched the human experience, it can
also cause dramatic disruptions in judgement and performance in
future events be it benign or harmful. Different emotions can
also be generated based on a single incident. For example, two
soccer team feel different emotions when a goal is scored. The
scoring team is happy while the other team is sad. This is
because their construe of the same event is different.
Aaron Sloman's "Attention and Affect" project has a aim to,
"develop a conjecture into a high level theory of the architecture of a human mind, where the conjecture is that design requirements for intelligent agents in a world like ours lead to mechanisms that have these tendencies as "emergent" properties"
as quoted from his paper, A Study of Motive Processing and Attention [20]. In Prolegomena to a Theory of Communication and Affect [16] by Aaron Sloman, he stated that,
"his interest is more on how affective states come about and interact with other states and process"
An affect is define in the Concise Oxford Dictionary as,
feeling, emotion, desire, esp. as leading to action.
The "affect" in Sloman's context is more than just
emotions. It includes moods, desires, attitudes, impulse and
other elements.
He suggested that affective states are dispositional states, both
long and short term, in a control hierarchy that consist of
positive and negative analysis of something and have at least a
inclination to produce motivational states which in turn have a
tendency to yield behaviour.
The control hierarchy consists of affective states with a variety
of time scales and links with behaviour. For example, states A
might be a disposition to produce and maintain states B and C
depending on the situation. This state or a combination of states
can generate a change on one another and might produce behaviours
and it is influenced by time.
In Motives Mechanisms and Emotions [15], emotions are
states where motives responds to and these motives are put
through a number of filters which tends to disturb on-going
activities. Motives is a significant concept in this theory and
it too involves a number of elements such as goals, emotions,
desires.
The emotions theory employed by the Oz Project of the CMU team was one that was adapted from the book, The Cognitive Structure of Emotions by Ortony, Clore and Collins [1]. The theory discussed in this book is more focused upon the meaning of emotional states and its structure. As quoted from the book, what they are trying to doing was,
"to specify the "psychological" structure of emotions in terms of personal and interpersonal situation description"
They are not formulating a theory about emotion words but a theory about the kind of things which are associated to the emotion words. It is also mentioned that they would like to lay a foundation for a computationally tractable model of emotion which principle can be used in an artificial intelligence (AI) system to reason about emotions. The reason for doing this is that they would like to employ AI tools to model the theory. Their aim here is not to create a machine with emotions, as it would be unrealistic to do, but to create a system which could predict and explain human emotions. When the predictions and explanation of the system is able to match those of human, then the theory would be proven to be pointing in the right direction.
The theory proposed by Ortony, Clore and Collins revolved around the notion of an emotion type. As quoted from the book, the definition is, An emotion type is a distinct kind of emotion that can be realised in a variety of recognisably related forms. By concentrating on distinct type of emotions rather than on the vast amount of discriminable emotional states, they managed to reduce the number of emotions to a more realistic proportion. All the emotion types defined are then divided into groups and these groups has two major characteristics. First, emotion types in the same group has eliciting conditions which are structurally related. An experience of an emotion not necessarily must be explicitly stated. It can be inferred from the description of the situation. These descriptions are known as eliciting conditions. Secondly, each distinct emotion type represents a family of closely related emotions. The family of emotions have the same eliciting conditions but may differ in intensity. The intensity here refers to how much of the emotion an agent is experiencing, for example, emotion type "fear" has a family of emotions like "concern", "petrified", "fright" and so on. All of which are different in intensities.
Figure 2.1 shows the organisation structure of the emotion
types as defined. There are altogether 20 emotion types which
falls into three categories, which are in turn divided into
various sub-groups.
The words used in the diagram are purely for technical purposes
and they are not critical. They are defined by the role they play
in the structure.
Labels in upper case represent structural elements and those in
lower case represent emotional states. Individual groups of
emotion types are enclosed in boxes with the names of groups at
the bottom of the groups.
A person's activity is usually supported by a goal or purpose. We rarely engaged in random activities. In this theory, an appraisal structure is used to determine what induces emotions. The three central components are the first three branch as shown in figure 2.1, namely,
I. Goals. The main focus variable is desirability. This covers the first branch of the structure. Desirability refers to the events which leads the satisfaction of a goal. Most of the events leading to the goal should be desirable. As the fulfilment of a goal is very likely to facilitate the fulfilment of a higher goal, The amount of desirability is primary dependent of the value of the higher goal. Three kinds of goals are defined here.
Active-pursuit goals. Goals which a person tries to obtain, for example, obtaining a degree. Interest goals. Goals which a person usually has no or little control over, for example, the success of a friend. Replenishment goals. Example of such a goal is to eat when hungry.
Standards. Praiseworthiness is the key variable. Here the appraisal of an event is not according to goal but towards the actions of other agents, which is the second branch of the structure. To compare other agents' actions with a set of standards to find out if these agents uphold these standards.
Attitudes. The third main variable is Appealingness.
The appealingness of an object is the third branch of the
structure. To determine the agent's attitude towards an object or
other agents. This is closely related to the other two central
variables.Emotion intensity was an important topic discussed in
general length in the book. However, due to the complexity of
this section and my limited time constraint, it was not
implemented in my system. Instead, Elliot's notion [11] of
intensity was explored.
The above had been a general introduction to the theory used by
the CMU team. The detail dynamics of all the branches of the
structure in figure 2.1 can be found in the book [1].
Emotion intensity, as mentioned before is an important aspect
in emotions modelling. A mild emotion intensity level may not
trigger off any changes in the agent but a strong intensity level
might even be felt and affect the emotions of other agents.
Regarding this issue of emotion intensity, Clark Elliot's paper, Emotion
Intensity Factors in "Believable Agents [11] described
how intensity was represented in his simulation, the Affective
Reasoner. Intensity functions are used in his simulation to
calculate the strength of emotion types.
My project is mainly centred upon the semantics-based theory
by Ortony, Clore and Collins [1]. The primary reasons for this
decision is that first, it is a theory developed so that emotions
can be suitably represented in a artificial intelligence system
which fits into my criteria and also it has been tried and
implemented by the CMU team which shows that such an
implementation is possible and relevant. Aaron Sloman's
design-based theory involves much more components than just
emotions. The "affect" component has a much broader
meaning beside just emotions. To breakdown the theory to just
representing emotions would be very difficult if not impossible
and not viable as the entire theory is likely to breakdown
without the other elements. Therefore this option was abandoned.
As for the topic of intensity in emotion, a modified and
simplified version of Elliot's notion [11] was implemented.
With the underlying theory of the emotion component decided, the
full agent architecture can now be designed.
Copyright © 1995 to Hay Yong Joo