-
Roughly, a motivator is derivative if it is explicitly derived
from another motivator by means-ends analysis and this origin is recorded
and plays a role in subsequent processing. A desire to drink when thirsty
would be nonderivative, whereas a desire for money to buy the drink would
be derivative.
The affective reasoner is concerned only with manipulations of the
nonderivative motivators which lead directly to the emotion inference
mechanism.
Among the emotion-reasoning systems, the Affective Reasoner and
ACRES may be characterized as approaches that deal with the representational
issues of emotions proper, and of the processes that generate them. By
contrast OpEd, SIES and AMAL may be seen as more general approaches that
have an emphasis on reasoning within the object domain about
situations that give rise to emotions. BORIS and THUNDER sit somewhere in
the middle. In other words, the Affective Reasoner and ACRES focus on the
nonderivative motivators which lead to emotions whereas the other systems
have varying degrees to which they also concentrate on the nonderivative
motivators. Note also, that while the Affective Reasoner does little
reasoning within the object domain itself, it does have an associated system
for indexing stories in an object domain according to the underlying
initiators of emotions (see [10]).
In this sense it is similar to SIES.
Of these programs, only ACRES gives any treatment at all to the
biological systems involved in the production of emotions, and then only
cursorily. The remainder (the Affective Reasoner included), only deal with
the cognitive aspects of emotions.
The earliest well-known computer model of emotion reasoning was
Kenneth Colby's PARRY [5], a program that mimicked the behavior
of a schizophrenic paranoid personality. A concise account of PARRY is
provided by Ortony [17] who writes,
- Colby's program, PARRY, accepted English-like input and responded as
a patient in a psychiatric interview by attempting to construe inputs as
directly or indirectly indicative of malevolence. If malevolence was
detected, one or more of three affective responses fear, anger, or
mistrust was triggered, depending upon the nature of the construed
malevolence. Physical malevolence induced fear, psychological malevolence
anger, and the induction of either also induced mistrust. The linguistic
output was designed to offset the construed malevolence either through
counterattack or withdrawal. In all cases, the nature of the construals and
of the responses was determined by the constantly updated values of a number
of key variables, including those for fear, anger, and mistrust.
The Affective Reasoner, like PARRY, also models agents who ``have''
emotions. In addition, both systems maintain internal dynamic states, which
affect the way in which the modeled agent(s) construe the perceived world
and manifest the emotions resulting from those construals. Starting with a
psychological model, PARRY attempted to model a true paranoid personality.
By contrast, the Affective Reasoner uses an ad hoc theory of
rudimentary personality representation designed to be broadly applicable
rather than focused on a specific psychological syndrome.
More recently, another system, BORIS [8], analyzed text
for affective responses to goal states and interpersonal relationships. In
this system, Dyer used ACE structures (Affect as a Consequence of Empathy)
to account for a number of inference mechanisms that included
affective content. In particular he developed the idea that understanding
affect can allow one to produce a causally coherent analysis of stories. In
BORIS, Dyer does not make a distinction between principles, goals and
preferences. This is important because the Affective Reasoner uses this
distinction to resolve some ambiguities. For example, a person may turn
himself in after committing a crime because it upholds a standard of
honesty, even though it clearly violates his goal of being free. Such
conflict between goals and principles are a common theme of human emotional
life and give us important leverage for reasoning in the affective
domain.
Dyer uses a six-slot frame structure to represent his ``affects''.
By contrast, the Affective Reasoner uses a hierarchy of loosely-structured
frames to represent interpretation schemas, and a simple nine-attribute
relation (called the Emotion Eliciting Condition Relation) to represent
emotion types. There are a number of reasons this latter approach is better
suited to the type of reasoning we do. In the first place, frame hierarchies
allow for the use of slot inheritance. This is useful when specifying the
interpretation schemas used for appraising emotion eliciting situations.
That is, properties may be shared by many interpretation schemas and yet can
be specified once in an ancestor, and then inherited by children frames. For
example, when a team scores a touchdown during a football game it may elicit
emotions in the fans. However, much of what defines this as an emotion
eliciting situation has to do with the nature of games in general, and at an
even higher level, with entertainment. Interpretation schemas may inherit
slots that specify the associated goals as entertainment
goals instead of health goals, and descriptive slots such as team-that-scored-the-goal, etc. The use of frames allows a great deal of
efficiency in making use of the system's stored knowledge about situations.
The frames used in this flexible representation of an agent's
interpretation schemas should not be confused with the simple relations
representing the emotions. Both BORIS and the Affective Reasoner use such
relations to represent emotions, although interestingly BORIS represents
these relations as frames. Here there is no room for inheritance: each
attribute of the tuple is specified and has an associated value. To give
inheritance to such an emotion specifier is to imply that one emotion is a
child of some other emotion, and, as such, derives value from the parent.
(Such a position is viable, but it is not one taken either by the Affective
Reasoner or BORIS.)
One result that the BORIS work shares with the Affective Reasoner
and the work by Ortony, et al. [18] is in suggesting valid, but
generally unnamed, emotion types. These arise when certain configurations of
the attributes in the Emotion Eliciting Condition Relations (to use the
Affective Reasoner's term) seem to give rise to a recognizable, but unnamed
set of emotional states in the agent. Dyer refers to the case where 
feels negative toward as a result of a goal failure caused by 
and
gives the example of a woman who felt foolish and castigated herself for
having had her purse stolen by a thief [8]. Ortony, et al.
identify the class of emotions that arise when one's fears have been
confirmed. They call these, appropriately, the fears-confirmed type
of emotions. In the current Affective Reasoner research questions are
raised about one's goals being blocked by the upholding of standards (i.e.,
``doing the right thing'' which has a bittersweet, heroic quality), and
having one's goals furthered by the violation of standards.
Related work has been done by Dyer on OpEd [9], which
understands newspaper editorial text. In the OpEd architecture, frames are
used to represent conceptual dependency structures [26]
which form belief networks about the world. Inferencing is guided by
knowledge of affect which is incorporated into the knowledge base. For
example, in trying to understand the meaning of a piece of text, the word
``disappointed'' causes demons to be spawned that resolve pronoun references
and search for blocked goals. This processing all falls into the category of
plausible reasoning within the object domain which, as previously discussed,
is beyond the scope of the Affective Reasoner's focus.
The THUNDER (THematic UNDerstading from Ethical Reasoning) program
of Reeves [21], is a story-understanding system that focuses on
evaluative judgment and ethical reasoning. Schema called Belief
Conflict Patterns are used to represent the different points of view in a
conflict situation. This work is related to that of the Affective Reasoner
in that the kind of moral reasoning needed to understand the stories that
THUNDER takes as input involves the same sorts of knowledge required to
generate the attribution and compound emotions. One of Reeves's
contributions is in the detailed working out of a set of moral patterns at
the level of stories. Reeves develops the idea that evaluators of
stories make many moral judgments about the characters in those stories
using these patterns, and that this moral framework is often necessary for
understanding text. For example, Reeves analyzes a story in which hunters
tie dynamite to a rabbit ``for fun''. The rabbit runs under their car for
cover and when the dynamite blows up it destroys their car. Reeves argues
that to understand the irony in this story the evaluator must know that
killing a rabbit with dynamite for entertainment is reprehensible. This
allows the judgment that the chance destruction of the car is an example of
morally satisfying retribution.
Clearly moral reasoning is closely related to the rise of emotions.
In both THUNDER and the Affective Reasoner, the focus is on the point of
view of individual agents within, or observing, a situation. Reeves writes:
``...using different value systems will produce different ethical
evaluations. For example, the actions of a high school student who killed
himself after failing a test would be judged ethically wrong by a Catholic,
but not by a Samurai...'' ([21], page 23). The Affective
Reasoner has no text understanding mechanism, but it could simulate
the suicide scene with various observers responding with different emotions,
dependent upon their (moral) principles.
One area that has not been focused upon in great depth is the action
generation component of emotion reasoning. Most action schemes in AI
research focus on the logically defined needs of the agent, (i.e., actions
as a function of planning for goals). Since emotions themselves are so
poorly understood and so seldom incorporated into experimental AI
systems, it is understandable that little work is being done that uses them
as motivations for action. One system that does function in this way,
however, is the ACRES program (Artificial Concern REalisation System)
developed by Frijda and Swagerman [12] in the Netherlands.
The main premise of their research is that emotions provide
a way of operating successfully in an uncertain environment (see also
[28]). They approach the design of their system from a
functional standpoint: if an agent is to be able to respond functionally
to its environment, what properties must the subsystem implementing
that functionality have? To this end, in contrast to the
Affective Reasoner's descriptive approach, the primary emphasis is
placed on behavior potentials and the search for reasons to
initiate them. To quote:
-
The major phenomena are: the existence of the feelings of pleasure
and pain, the importance of cognitive or appraisal variables, the
presence of innate, pre-programmed behaviors as well as of complex
constructed plans for achieving emotion goals, and the occurrence of
behavioral interruption, disturbance and impulse-like priority of emotional
goals [12], (page 235).
Their architecture is designed to meet these demands:
- The system properties underlying these phenomena are facilities
for relevance detection of events with regard to the multiple concerns,
availability of relevance signals that can be recognized by the action
system, and facilities for control precedence, or flexible goal priority
ordering and shift (page 235).
The Affective Reasoner also addresses some of these issues, although
the emphasis is different. The Frijda-Swagerman action architecture arises
from issues relative to processing needs, whereas ours arises from a
description of actions. Similarities can be found on a number of fronts.
Both systems incorporate the idea of the cyclic nature of emotion machinery:
a situation arises, it is evaluated, it is either ignored or a response is
initiated which is fed back to the modeled world. As with the
Affective Reasoner, ACRES performs no reasoning about derivative motives.
Instead, emotion triggers are directly wired into the architecture.
The Affective Reasoner and ACRES also both stress the importance of
concern realization as a way of filtering out situations that are
not of interest, and of interpreting those that are. Entities have multiple
concerns and limited resources. To solve this problem both programs focus on
matching as an efficient means of assessing emotion eliciting
situations. To quote Frijda and Swagerman, ``Concerns can be thought of
as embodied in internal representations against which actual conditions are
matched,'' [12] (page 237), which is exactly the conceptual
approach of the Affective Reasoner.
The two systems differ, however, in the way in which they use the
realized concerns. Of primary importance to ACRES (but only of incidental
importance to agents in the Affective Reasoner) is filtering out situations
irrelevant to resource conservation (as most situations are). In addition,
unlike the Affective Reasoner, ACRES stresses the correct interpretation of
situations with respect to urgency (as distinct from importance- see
[28]). To approximate this behavior ACRES uses a system of
interrupts.
There are a number of interesting issues that ACRES does not
address. One of these is the maintaining of expectations with respect to
future goals. The aging of expectations and the elicitation of associated
prospect-based emotions is important for systems representing emotional
behavior, as is the representation of the emotional status of other agents.
ACRES does not maintain internal models of other agents, nor does it
maintain a working memory for the storage of currently active relationships.
Without these features it is not possible to reason about the emotional
lives, and therefore motivations, of other agents with whom it must
interact. A simple test of such a system is to see how well it fares on
modeling a situation such as the following: Tom pities his friend Harry who is sad because his (Harry's) mother has had her garage
burn down. Humans have no trouble with this because not only do we have an
internal representation of our own concerns, but we are also apparently able
to use some similar mechanism for representing the concerns of others, and
thus ``imagine'' how they might feel with respect to some situation. In this
case Tom must not only model Harry, but he must also model
the concerns that Harry is modeling for his mother. (See
section 2.8.3.)
The Affective Reasoner, in contrast to ACRES, incorporates
representations of the concerns of others in its architecture (see section
2.8), and represents relationships between agents. These
representations are incorporated into the emotion-generation process when
initiating fortunes-of-others emotions (see section 2.3) and into
the process that helps agents identify new situations as instances of a
certain emotion type.
The System for Interpretation of Emotional Situations (SIES) of
Rollenhagen [24], and the AMAL system of O'Rorke
[16] share a number of features. They both have taken
on the burden of
having to do non-derivative reasoning about the situations that give rise to
emotions within the object domain. They both use retrospective reports about
emotion-inducing situations as the basis for the structural content analysis
that drives the reasoning systems. Lastly, they both are descriptive in
nature. SIES attempts to better define emotion terms by exploring their
similarities and differences through abstract representations. AMAL attempts
to explain situations as emotion episodes through the use of abductive
reasoning.
By contrast, the character of the Affective Reasoner, as a simulation, is quite different from these two programs. In the Affective
Reasoner the process of representing emotions contributes not only to
developing a language for describing emotion episodes but also to
the building of intelligent agents that ``have'' emotions. In addition,
while the Affective Reasoner uses a logic-based representation for reasoning
about the eliciting conditions for emotions, similar to both SIES and AMAL,
neither of these latter systems has anything like the case-based approach
used in the Affective Reasoner to reason backwards from the manifestations
of emotions to the emotions themselves.
The theoretical underpinnings of AMAL and SIES are quite different,
although in practice this difference is minimized. AMAL (based in part on
the work of Ortony, et al. [18]) is an extension of the cognitive appraisal approach where the antecedents of emotions are
described independently of the particular situations from which they derive.
In this approach it is situation types that give rise to emotions
(e.g., an undesirable event giving rise to distress), rather than relatively
specific situations (such as the death of a loved one). However, since AMAL
is designed to be able to represent such real-world emotion-eliciting
situations as those portrayed in student diaries [31],
it must also be able to do extensive reasoning within the object domain. To
represent these emotion anecdotes a situation calculus is used.
SIES, on the other hand, is an attempt to combine the cognitive approach
with a situational approach in which the antecedents of emotions are
seen as arising in certain real-world situations (e.g., loud noises leading
to fear, or separation from a loved one through death leading to sadness).
In his description of SIES, Rollenhagen gives many rules for emotion
reasoning, but most of these are employed in describing an instance of a
given emotion eliciting condition as embodied in a narrative description -
that is, in giving structure to the derivative motivations of the agents.
SIES should be understood as a system for analyzing emotion
episodes. In this regard, much of the contribution of the system is in
structuring the representation of the data. The program does not map
situation features into specific eliciting conditions. Instead, inferences
tying abstract situations into the personal concerns of the subjects are
made at the time the researcher encodes each emotion eliciting situation.
The representation of emotion eliciting situations in the Affective
Reasoner is from a different perspective, and is structured around the idea
that agents may be seen as carrying internal schemas that match
situations that arise in their world. Here the focus is on frames that
contain enough information to discriminate one situation from another
through matching. Varied interpretations of these situations,
produced by unifying an interpretation frame with a situation frame during
the course of the simulation, are seen as whole units. These interpretation
units are configured in various ways such that they give rise to the various
emotions.
Unlike SIES and the Affective Reasoner, AMAL [16] does do
extensive, automated, plausible reasoning about derivative motivators. To
do this it uses an abduction engine. AMAL uses a logic-based situation
calculus framework to describe actual emotion
episodes[31]. Using abductive reasoning, AMAL is able to
construct plausible explanations for emotion instances. For example AMAL
analyzes plausible relationships in the following story:
One other important difference among the three systems should be
noted. AMAL and SIES typify the approach that takes the logical structure
of the preconditions for emotions to be simply an extension of the structure
of the world (e.g., shouting causes Tom to startle, which causes him to drop
a hammer on his foot, which causes him to feel pain, which causes his friend
to feel sorry for him). Our approach is fundamentally different in that
there is a definite demarcation between the internal strong-theory rules
used to generate emotions for agents, and the object-domain rules used to
map into them. In this approach, since we see emotion eliciting situations
in the object domain as instances that may or may not match internal schemas
representing the concerns of the agents, processing and data-entry tasks are
very different in nature. While each approach can, of course, be reduced to
a logic representation, the two languages for content theory representation
are entirely different, and thus, in practice, so are the systems. In AMAL
and SIES everything is represented as rules, in the Affective Reasoner the
concerns of the users are represented as collections of features.
Toda's work on Solitary Fungus-Eaters [30] attempts
to describe emotions as a situated functional component of society. In this
case, Toda's society is that of simple (fantasy) uranium-mining robots that
operate autonomously on a distant planet and eat native fungus to survive.
The goal of these robots is to collect as much uranium as possible. If they
consistently choose only to travel to locations containing uranium, however,
they may not encounter enough fungus to keep themselves going and in such a
case would ``die''. Society arises out of a need for a ``coalitional
bond'' to help protect the robots against stronger predators, and so forth.
Toda's structural analysis of the human emotion system is based on
his stated belief that emotion was a useful mechanism for a very different,
primitive, society than our current one, and that just as our bodies have
problems dealing with modern-day environmental stresses such as pollution,
so does our emotional system have difficulties with such an (evolutionarily)
unfamiliar environment. To illustrate his ideas about why an emotion system
arose in the first place, giving insight into its nature, Toda details a
system of urges necessary to solve coordination problems in his
simple robot society.
For example, in response to the threat of a larger predator, a
Fungus-eater may have a Fear Urge which causes him to send out a
distress signal. This in turn triggers a Rescue Urge in other robots
who then come to jointly attack the predator. Since a reward is necessary
for this scenario to fit into the ``individual-gain'' paradigm of Toda's
robot society, a Gratitude Urge is developed. Rewards, given out of
gratitude by the rescued robot, may have to be postponed, however, if the
rescued robot is short on resources. Since this delay may cause problems, a
more immediate reward in the form of a Love Urge is developed, and
so forth.
This major difference between this work and that done on the
Affective Reasoner is of course that the latter is actually implemented as a
running program. While Toda has given some level of detail for many aspects
of his system (including vision and planning, not discussed here), there are
nonetheless many gray areas in his description. Vision and planning aside,
such social actions as defecting from one's coalition partner and joining a
more profitable coalition are treated as primitives in the discussion. To
implement even a small number of these primitives of the robot society would
be a major task. In addition Toda's urges, intended to be thought of
tentatively as simple emotions, are presented anecdotally and not as part of
a cohesive, complete theory.
Toda's theoretical perspective shares with both the Affective
Reasoner and ACRES the idea of emotions as mediators for controlled,
immediate responses to situations (see also [27]). To this
end, emotions take a functional role in the automated organism.
From this short survey it should be apparent that comparatively
little work has been done in this area, and that what has
been done is often so different from what has come before that comparisons
are strained. PARRY is a system that attempts to model a diseased
and limited mind. BORIS, OpEd and THUNDER use a content theory of emotions
to aid in the understanding of text. ACRES attempts to model the dynamic
utility of emotions. SIES provides a structure for mapping from the features
of situations into the emotions. AMAL uses abductive reasoning, well suited
to the understanding of emotion episodes, to explain emotion situations.
Clearly these are diverse approaches only
loosely tied together by the generic categorization of the work as ``emotion
research''. One theme that all of these treatments share however,
is that working out theories at the level of detail required for a running
computer program forces the theorist to refine and completely specify all of
the hidden ``gray areas'' of the theory. This is particularly important for
this complex, and little understood subject.
1.4 TaxiWorld
The Affective Reasoner has three primary components: (1) the
affective reasoning component, which is independent of any domain, (2) a
world simulation based on an object-domain theory, and (3) a graphics
interface based on the simulated world. Once an object-domain theory has
been incorporated into the Affective Reasoner, and a graphics interface
customized for that domain, it becomes a domain-specific system. In its
current form the Affective Reasoner primarily manipulates agents that
are intended to be interpreted as operating in a schematic representation
of the Chicago area (figure 1.1). We call this TaxiWorld.
In fact, other partially analogous domains can be simulated if the user
simply reinterprets the meaning of the icons displayed. This how the
football game example discussed in chapter 3 was run.
TaxiWorld has approximately forty-five different sets of simulation
events which produce emotion eliciting situations. Included are events that
represent traffic accidents, rush hour delays, getting speeding tickets,
picking up passengers, getting paid, having to wait a long time for a taxi,
and so forth. In addition, there are approximately forty different
rudimentary ``personalities'' which may be given to each agent that
participates in one of these situations. For example, one taxi driver may
get angry when he is given a small tip whereas another may just figure it is
part of the job. Or, both may get angry but one will be rude to his
passenger, whereas another may just smile and pretend that he does not care.
There are roughly 150 different candidate interpretations of the forty-five
situations. Different interpretations, or sets of interpretations,
can give rise to one or more of twenty-four emotion types, each of which
has about sixty possible action responses associated with it. Used in
different combinations these components can yield tens of thousands of
different emotion episodes.
In any research of this kind the question naturally arises as to
what has actually been implemented and what has been run. In this
dissertation only a few emotion episodes will be discussed in detail so as
to allow us to focus on emotion representation issues. We provide at most
one emotion episode example for each point. However, it should be understood
that many different examples may actually have been run in the process of
addressing any one particular problem. In addition, unless otherwise stated,
illustrative examples in this dissertation are based on code that has been
written and actual simulated episodes.
The researcher using TaxiWorld has two forms of input into the
system. First he or she statically configures the object domain (i.e., the
world of taxi drivers and their passengers) through LISP files. Second, he
or she dynamically manipulates the running simulation through menus and
windows. Configuring the object domain has six parts: (1) Should the
researcher wish to extend the number of situations represented then new
simulation events must be entered into a LISP file, and event handlers
written to process them. (2) Once the object domain (e.g., the world of taxi
drivers) is stable, the researcher may choose to add new interpretations of
the situations that arise in it by creating construal frames (see
chapter 2) with which agents interpret situations. (3)
New emotion manifestations may be created within the object domain. For
example, in the world of taxi drivers we may wish to represent ``cutting
someone off'' as an expression of anger. (4) New personalities may be
created by grouping (old and/or new) construal frames into sets, and by
grouping temperament traits (i.e., tendencies toward certain types
of actions as expressions of emotion) into new sets. (5) Simulation
sets may be constructed by specifying which of the represented situations
are to arise, which types agents are going to be present, and which
personalities those agents will have. Lastly, (6) a content theory for
reasoning about observed actions from cases may be recorded in a data file
or through interaction with the running system. User input macros have been
provided for (2) - (6).
Figure 1.1 shows a portion of TaxiWorld's map display.
Interaction with TaxiWorld is through this interface. Except for stop and go each of the buttons on the control panel leads to a
pop-up menu for manipulating different aspects of the system. This control
mechanism allows the following functions, by button: (1) stop and
go: start the simulation running or pause it, (2) demos:
select a configuration of agents and situations for the simulation, (3) control: change the speed of the simulation, or enable and disable
different features in the simulation such as whether or not agents
negotiate and what types of traffic delays may arise, (4) set: set
rudimentary global moods for the agents, and set the mode for the heuristic
classification system (i.e., set the case-based reasoning system in learn mode, in report-only mode, or off), and (5) queries: look at a scrolling window containing summaries of the
situations that have arisen so far.
Figure 1.1:
The TaxiWorld display.
|
|
Once the user has configured and started the simulation, the
animated icons move around the screen as agents who ``experience'' the
situations in the simulated world. The figure shows a simulation that has
been running for a short time. Four taxi drivers, Tom, Dick, Harry and Sam
are shown. Tom and Sam are traveling between Chicago's O'Hare Airport and
the Loop, Dick is between the ``Junction'' and the Loop, and Harry is on his
way to the Loop from the Museum of Science and Industry (not pictured). One
passenger is waiting for a ride at Northwestern University, and five
passengers are queued up waiting for rides from the Chicago Botanical
Gardens. During rush hours the roads swell with traffic and the agents move
more slowly. This occurs also with accidents. Highway patrol agents may
also be displayed.
Selecting an agent with the mouse provides information about that
agent. One option will bring up a scrolling window containing the current
``physical'' state of the agent: where he is, whether he has someone in the
cab (if he is a taxi driver agent), how much gas he has left, where he is
going, etc. Another option brings up an emotion-history scrolling window.
Figure 1.2 shows one of these, in this case for Harry. Here we
see summaries of two emotion instances, distress and dislike. These have arisen from two different construals of the same
situation. Briefly, a new passenger has gotten into Harry's cab and was
cheerful. Harry, whose ``personality'' type might be roughly described as
depressed grouch has a goal and a preference that apply in such a
case. First, he has no desire to be around happy people because they remind
him of how unhappy he himself is. This leads to distress. Second, he just does
not like happy people, he finds them distasteful, so he dislikes this
passenger.
Figure 1.2:
The emotion history scrolling window.
|
|
For maximum flexibility, the researcher has the option of
communicating with the running system directly through LISP. Basic tools
have been provided for this, in particular to control the asynchronous
processes spawned in the simulation. The user can also interact with the
case-based reasoner through LISP as well.
TaxiWorld is written in Common LISP and runs on a SUN SPARC
station 1 with 40MB of RAM. It uses a simple SOLO graphics
interface. The TaxiWorld code is about fifteen thousand lines in length,
and is combined with an additional fifteen thousand lines of slightly
modified code based on Dvorak's Common LISP version of Protos
[7].
1.5 Two examples from TaxiWorld
Figure 1.3:
Harry learns about Tom's fear.
 |
In this section we look at two examples of the kinds of episodes
that occur in TaxiWorld. Figure 1.3 represents a simple episode
where a taxi driver, Harry, observes another taxi driver, Tom, and learns
something about his fears. This episode makes use of three agents (the third
is Tom's passenger - not pictured), one situation (that arises in the
Chicago area) and a single affective state, in a TaxiWorld
simulation. In this episode we see the following: (1) Tom picks up a
passenger headed for Joliet at the Museum of Science and Industry. (2) The
passenger is seedy-looking, causing Tom to fear that he will be robbed. (3)
This fear is expressed as a flushed expression on Tom's face, a stutter in
his voice, and the statement that he does not have enough gas to drive to
Joliet. (4) Another taxi driver, Harry, watches this episode. He does not
know that Tom is inclined to be afraid of seedy-looking passengers whose
destinations are Joliet. He has no representation of how Tom might construe
such an event. However, (5) Harry has seen a case where another agent
had a flushed face and a stutter in his voice. The previous case was known
to be an example of fear. (6) Harry now reasons to come up with a
good explanation: if Tom is experiencing fear, why might this be so? He
``imagines'' some possible interpretations of this event that might cause
someone to be afraid and reasons that if Tom had the goal of retaining his
money and/or maintaining his personal safety, then given that seedy-looking
people have been known to rob cab drivers - thus taking away that money
and perhaps causing harm, Tom might experience fear. Since this explanation
fits, Harry drops his search. (7) He updates his internal representation of
Tom as someone for whom retaining money is important, and who is afraid of
unsavory passengers. (8) Harry also now has a new case which he may wish to
save. The ``I don't have enough gas'' ploy, which was not a feature of the
previous case of fear, may be explained to him as a problem-reduction strategy, or it may be left unexplained but associated,
depending on whether the system was being trained, or was running in
automatic learning mode.
Harry now believes he has learned something new about Tom. If this
knowledge is correct he will be better able to explain Tom's future actions
in response to such a situation without resorting to a case search (i.e., he
knows how Tom will interpret the situation). In addition he will be
able to predict how Tom might react to such an event, even in the
absence of any empirical evidence. On the other hand, if at a later time the
knowledge proves to be erroneous it will be discarded and a search for an
alternate explanation will be initiated.
Figure 1.4:
Negotiating about who takes the seedy-looking passenger to Joliet.
 |
The episode pictured in Figure 1.4 expands on the previous
example. Here we see that some time later Harry has come to know Tom quite
well.![[*]](footnote.png)
He has met a new agent, Dick, who seems to respond to events the way Tom
does, although he (Harry) has never seen Dick pick up a seedy-looking
passenger headed for Joliet. Until he knows differently, Harry makes the
assumption that Dick is just like Tom (i.e., with respect to the way he
interprets situations). At one point, Harry (who has the goal of making
money and likes to go to Joliet because, as a long trip, this furthers his
goal), sees an opportunity. Here is the sequence of events: (1) Dick and
Harry arrive at the Botanic Gardens in that order, giving Dick the right to
the first passenger. (2) A seedy-looking passenger arrives at the cab stand
and wants to go to Joliet. (3) Harry matches the passenger-arrival
situation against an opportunity prototype.He sees an opportunity to make a request of Dick, but ``thinks it through''
first to see if it is feasible. Using his representation of how Tom sees
the world as a default for Dick, Harry believes that Dick, like Tom, has
fear about going to Joliet with the seedy-looking passenger. On the other hand
he believes that he will, himself, be happy about going on such a trip. (4)
Harry proposes to Dick that Harry, instead, take the fare to Joliet. (5)
For his part, Dick, considering this proposal, ``imagines'' that Harry gets
to take the passenger. He has no emotional response to this. Next he
``imagines'' taking the passenger himself. As it happens, Harry's schema
for Dick is incorrect: while Dick is indeed very much like Tom in many ways,
he is unlike him in that he is not afraid of seedy passengers who want to go
to Joliet. Consequently Dick does not see a threat to his goals, but rather
only that if he does make the trip to Joliet he will make some
money. In other words, he reasons that if he makes the trip he will be
happy, and that if he does not make the trip he will simply experience a
lack of emotion. (6) He therefore does not agree to Harry's proposal and
instead takes the passenger himself. (7) Harry believes he has learned
something new about Dick. Apparently Dick does not construe seedy
passengers going to Joliet as threatening. Harry splits off his internal
representation for the goals, standards and preferences of Dick from the
representation of those for Tom and removes the offending interpretation
frame from the latter.
In this example we see that Tom and Dick both have distinct
emotional lives, and that Harry is able to reason about them. Tom has a goal of making (keeping) money and is able to experience fear as a
result of believing that this goal may be threatened. Similarly, Dick also
has a goal of making money and is capable of feeling hope over
the prospect of taking a passenger on a long trip.
On the other hand, Harry, who is observing Tom and Dick, draws on
his experience, and his internal schemas for the other agents, to explain
their actions. In the first case, he has no clue as to how Tom might
interpret the situation. He looks through some cases to see if he might make
a guess as to what stuttering, flushing and the out-of-gas ploy
might indicate. He has seen something similar when someone else was afraid,
and he can explain the fear as a fear of getting robbed. Since this
explanation of the situation is workable he assumes that this is part of how
Tom interprets the world, and he now adds this to his representation of
Tom. Until he has cause to do differently he will now always make this
assumption about Tom.
In the future, Harry will test this knowledge whenever a similar
situation arises by asking this question, Given a similar emotion
eliciting situation (seedy passenger going to Joliet), does the observed
agent respond in a manner compatible with an emotion to which the assumed
interpretation leads? If the agent does, then Harry will be more
confident that his assumed interpretation is correct; if he does not, then
Harry will have to search for another explanation (i.e., one which leads to
a different emotion).
In addition, Harry's representation of Tom is one he can draw on as
a default personality type, for reasoning about a new agent, Dick. As
long as Dick's actions can be explained in terms of Tom's personality, the
default personality suffices and he uses it to ``see the world through
Dick's eyes''. When an opportunity arises he believes that it is worth
pursuing because he is able to ``imagine'' how Dick will perceive his offer.
However, when Dick turns him down, Harry reasons that he has made an
incorrect assumption about Dick and removes this supposition from his
representation of how Dick sees the world. But this means that Dick no
longer looks just like Tom, so Harry must split the two schemas apart. The
old one still represents Tom, and the new one now represents Dick.
The rudimentary negotiation, the simple opportunity recognition, and
the rudimentary default personality reasoning, all open research questions
in their own right, are not what is important in these examples. What is of
interest is the idea that there are units of appraisal, here represented as
frames, which can be used as filters for the interpretation of situations,
and that these appraisal units can be mixed and matched as necessary to
construct rudimentary personalities for agents (i.e., to give them primitive
affective life). Of additional interest is the idea that these appraisal
units can also be used by an observing agent to construct an internal
representation of the observed agents for generating explanations of
their actions, and for predicting their emotional responses to future
similar situations. Lastly, it is important to consider that agents
manifest their emotions in ways that can be understood by observers: agents
often communicate their emotions through their actions, or their lack
thereof.
In the following chapters we attempt to illustrate some concerns
that have to be addressed by emotion reasoning systems, how our
representation addresses these concerns, and how it can be used to model a
number of different multi-agent interactions.
2 Overview
In this chapter we provide an overview of the Affective Reasoner.
The discussion focuses on the functional role of the various modules in the
system. Two of these modules (construal, and action-generation), are quite
elaborate, having important theoretical content, and so are only introduced
here since they are discussed more fully in later chapters. The other
modules are less elaborate, and are treated fully in this overview.
The sections are organized around the various stages the system goes
through in processing a simulated situation. They are introduced in the
order of the different processing stages. These stages, and the static
representations which they use, are illustrated in figure 2.1.
Processing flows from some initiating simulation event through emotion and
action generation to the final stages simulating observation by other
agents. Roughly speaking, this may be interpreted as follows: Something
happens in the modeled ``world'',
creating a situation. The agents that populate this world interpret the
situation in terms of their individual concerns. Each interpretation is
reduced to a nine-attribute relation, and some variable bindings. Specific
configurations of the relation lead to the production of specific emotional
states in the agents (i.e., their emotional responses to the situation).
These emotional states, in turn, are manifested as action responses. Agents
can observe each other's responses to situations and attempt to explain
them in terms of emotions the other agent may have been
experiencing. An observing agent then finds an explanation for that emotion
based on varied interpretations of the original situation. Once this is
done, the observer places the schema that lead to the successful
interpretation (i.e., the interpretation that lead to what was believed to
be the other agent's emotion) into a database representing the concerns of
the other agent. This Concerns-of-Other database may then be used to predict
and explain the other agent's responses to future similar situations.
Figure 2.1:
Processing stages and related representations
 |
2.1 Fundamental concepts
This research is at the juncture of Artificial Intelligence and
Cognitive Science. Its interdisciplinary nature can lead to confusion over
the use of terms, such as personality. Accordingly we will, in this
section, make explicit our intended meanings. In addition, we also discuss
some central concepts, such as the basic emotion eliciting condition theory
from
[18] on which the construal process is based.
Rudimentary personalities. To computer scientists personality means something like an aggregate of qualities that
distinguish one as a person. (Webster). To psychologists however, personality means something rather different, something like co-occurring classes of trans-situational stable traits.
In the Affective Reasoner, the agents in the system have rudimentary
``personalities'' (at least computer scientists might consider this to be
so) that distinguish them from one another. These rudimentary personalities
are divided into two parts: the disposition agents have for interpreting
situations in their world with respect to their concerns, and the
temperament that leads them to express their emotions in certain ways. We
have chosen the terms interpretive personality component and manifestative personality component to denote these two parts of an agent's
unique makeup. By the former we mean the rudimentary
``personality'' which gives agents individuality with respect to their
interpretations of situations (i.e., their uniquely individual concerns).
By the latter we mean the rudimentary ``personality'' which
gives agents individuality with respect to the way they express or manifest
their emotions.
Situations that lead to emotions. In the Affective Reasoner,
some simulation events create situations that can initiate emotion
processing on the part of the agents involved. These we call emotion
eliciting situations or simply, eliciting situations. An example of
such an eliciting situation is the conceptual ``arrival'' of some
agent at a location, which might give rise to an emotion such as relief or
distress. Eliciting situations are not to be confused with eliciting conditions which derive from the emotion eliciting condition
theory of [18] and are discussed below.
Goals. This term has a very broad meaning. Here we use it in
its simplest sense: a state of affairs that an agent desires to have come
about. In this dissertation, unless otherwise specified, most goals will be
considered to be equivalent in structure.Clearly this is not actually the case. Some goals are preservation goals,
some are never satisfied (i.e., living a good life), some can be partially
achieved by achieving some other goal (i.e., saving another $100 towards a
house), and so forth. In general, however, these distinctions have more to
do with goal generation, interaction and retirement than they do with the
way goals fit into our underlying cognitive theory. Consequently, for most
of the discussion it will suffice to define the term goal to mean
simply, a desired state of affairs that, should it obtain, would be
assessed as somehow beneficial to the agent.
Object domain. Emotion reasoning can be performed as well by
one person as by another. Furthermore, people can have emotions about almost
anything and in almost any circumstance. People have emotions about
something important like the state of their finances, but they might also
have them about something as apparently inconsequential as the state of
their shoelaces. For this reason emotions may be considered an abstract domain that operates within object domains. The language of the
emotion domain is abstract: goals, standards, preferences, and so forth. The
language of an object domain is specific: money, shoelaces, etc. The
Affective Reasoner can be used to do abstract emotion reasoning in any
object domain that can be modeled. One such object domain is the world of
taxi drivers, as represented by the TaxiWorld version of the Affective
Reasoner. By object domain then we mean, that domain in which
the eliciting situations for emotions are described, and in which
emotion-based actions are manifested.
The emotion eliciting condition theory. The emotion eliciting
condition rules we use for the strong-theory reasoning component of the
Affective Reasoner are based on the work of [18]. Ortony, et al.
specify twenty-two emotion types based on valenced reactions to
situations construed either as being either goal-relevant events,
acts of accountable agents, or attractive and unattractive objects. The extended and adapted twenty-four emotion-type
version of the emotion eliciting condition theory that we have used in the
Affective Reasoner is outlined in table 2.1, based on
the work of [16]. Each of these twenty-four emotion states has a
set of eliciting conditions. When the eliciting conditions are met, and
various thresholds have been crossed, corresponding emotions result. A key
element of the theory is that the way emotion eliciting situations map into
these eliciting conditions depends on the interpretations of the
individual agent.
basket at the buzzer, and your team loses. I may experience joy at the
event, whereas you may experience distress. In both cases we share the
same sets of eliciting conditions for our emotions and the emotion eliciting
situation is the same (i.e., the ball went in the basket just before the
buzzer); it is only the interpretation or construal of the situation
which is different.
Table 2.1:
Emotion Types
| Group |
Specification |
Name and Emotion Type |
| Well-Being |
appraisal of a situa- |
joy: pleased about an event |
| |
tion as an event |
distress: displeased about an event |
| Fortunes-of- |
presumed value of |
happy-for: pleased about an event |
| Others |
a situation as an |
desirable for another |
| |
event affecting |
gloating: pleased about an event |
| |
another |
undesirable for another |
| |
|
resentment: displeased about an event |
| |
|
desirable for another |
| |
|
sorry-for: displeased about an event |
| |
|
undesirable for another |
| Prospect- |
appraisal of a situa- |
hope: pleased about a prospective |
| based |
tion as a prospec- |
desirable event |
| |
tive event |
fear: displeased about a prospective |
| |
|
undesirable event |
| Confirma- |
appraisal of a situa- |
satisfaction: pleased about |
| tion |
tion as confirming |
a confirmed desirable event |
| |
or disconfirming an |
relief: pleased about a disconfirmed |
| |
expectation |
undesirable event |
| |
|
fears-confirmed: displeased about |
| |
|
a confirmed undesirable event |
| |
|
disappointment: displeased about |
| |
|
a disconfirmed desirable event |
| Attribution |
appraisal of a situa- |
pride: approving of one's own act |
| |
tion as an account- |
admiration: approving of another's act |
| |
able act of some |
shame: disapproving of one's own act |
| |
agent |
reproach: disapproving of another's act |
| Attraction |
appraisal of a situa- |
liking: finding an object appealing |
| |
tion as containing |
disliking: finding an object unappealing |
| |
an attractive or |
|
| |
unattractive object |
|
| Well-being / |
compound |
gratitude: admiration + joy |
| Attribution |
emotions |
anger: reproach + distress |
| |
|
gratification: pride + joy |
| |
|
remorse: shame + distress |
| Attraction / |
compound emotion |
love: admiration + liking |
| Attribution |
extensions |
hate: reproach + disliking |
The emotion types are simply categorizations of selected patterns of
emotion eliciting conditions. They have been given English names roughly
corresponding to an intensity-neutral label for the type of
emotions represented by the specific configurations of the emotion
eliciting conditions. It is important to note that these names (e.g., joy and anger) given to the emotion types are not to be
mistaken for the specific emotions to which they usually refer. For
example, annoyance is one of the anger type emotions, as is
rage, because they both follow from interpretations of a situation
as an undesirable event coming about as a result of someone else's
blameworthy act.
Emotion eliciting conditions leading to emotions
fall into four major categories: those rooted in the effect of events on the goals of an agent, those rooted in the standards
and principles invoked by an accountable act of some agent, those
rooted in tastes and preferences with respect to objects
(including other agents treated as objects), and lastly, selected combinations
of the other three categories. Another way to view these categories is as
being rooted in an agent's assessment of the desirability of some
event, of the praiseworthiness of some act, of the attractiveness of some object, or of selected combinations of these
assessments.
Emotion types. As discussed above, when we speak of an
emotion generated by the system we are really talking about an emotion of
that type. It is usually not convenient (i.e., it obscures the
meaning of the text) to talk about the ``emotion type of anger'' and so
forth. However, it should be understood that throughout this text that when
we refer to some emotion, as though by name, we are actually referring to
some unspecified emotion characterized by that named emotion type.
Goals, standards and preferences databases. These
are referred to in the text as GPSs. They are the hierarchical frame
databases used to represent an agent's concern structure.
They are organized around the
three categories of an agent's concerns. They hold most of the information
used to define an agent's interpretive personality component. When
an agent is created, the GSP database must be filled in to give it a
unique set of concerns. When an eliciting situation is interpreted using
this database, attributes of the eliciting condition theory are bound to
features in the situation. See table 2.1, and
chapter 3 for discussion.
Representing concerns of others. An agent may represent the
concerns of some other agent by keeping a partial GSP database for that
other agent, and filling it in as knowledge about the agent is acquired.
Since these databases represent the concerns of other agents, with respect
to an observing agent, they are known as Concerns-of-Others,
or COO, databases.
Since they are generally learned by the agent, COOs are usually incomplete,
and may contain erroneous interpretation schemas as well. See
section 2.8 for a discussion.
Emotion Eliciting Condition Relations. The process that
matches frames in the GSP (and COO) databases against an eliciting situation
will, if the match is successful, reduce the eliciting situation to a set of
bindings. Some of the bindings represent values for two or more of the nine
attributes of a special relation known as the emotion eliciting
condition relation.
Different patterns of bindings for the attributes in this relation, and
different values, give rise to different emotions. In the text these are
abbreviated as EEC relations. See section 2.3 for
discussion.
Action response categories. Once an agent is in an emotional
state, it will manifest this emotion in one way or another. Some of these
manifestations naturally fall into one functional category, while others
fall into another. For example, trembling and breaking into a
cold sweat may be categorized by the somatic characteristics they
share. We have specified approximately twenty different groups (there is
some variation between the emotions) to which the various action responses
belong. See figure 4.1 for a complete listing of the action
response categories for gloating, and section 2.6.1 and
chapter 4 for discussion.
Frame types. Construal frames, which interpret emotion
eliciting situations (also represented as frames) must be retrieved from an
agent's GSP database at the time the situations arise. The retrieval of
candidate construal forms is a feature-indexing problem. A satisfactory
solution to this problem is beyond the scope of the current research. We
sidestep it in the Affective Reasoner by simply giving each eliciting
situation a type, such as an arrival-at-destination frame
type, or a passenger-pays-driver frame type. Construal frames of a
particular type are candidates for interpreting eliciting situations of the
same type.
The structure of simulation events. The Affective Reasoner is
designed around a simulation engine. Initial simulation events are
placed in the system's event queue. When the simulation is started these
simulation events are processed, and spawn further simulation
events which are then placed in the event queue, ad infinitum. In this
manner, once the simulation is started, it continues until it is halted.
Some of the simulation events simply drive the system, moving icons around
the screen and so forth, and are not of theoretical interest. Other
simulation events have eliciting situation frames attached to them
which, when matched against the concerns of agents, can initiate emotion
processing.
2.2 The construal process
The construal process is covered in detail in
chapter 2.2. Here we give a brief introduction to the
emotion eliciting condition theory from [18], and place the
construal process in the context of that theory. After this introduction we
discuss how a matcher is used to identify and instantiate internal schemas
called construal frames which are used to interpret eliciting
situations in the simulated world for the simulated agents.
When a simulation event occurs which has an attached eliciting
situation frame, agents appraise the situation's relevance to their
concerns by trying to match it against internal schemas representing the
eliciting conditions for the various emotions. Has a goal been achieved?
If so, is the goal important enough to generate an emotion? Has a standard
been violated? Who is responsible for the blameworthy act? The answers to
such questions result in an interpretation of the event with respect
to the eliciting conditions for emotions in terms of the observing agent's
concerns. If the interpretation is such that the eliciting conditions have
been met, and certain feature thresholds have been reached (e.g., was enough money lost...), then an emotion will result.
The internal schemas which map emotion eliciting situations into the
eliciting conditions for an emotion for some agent are represented as
frames, known as construal frames. These construal frames are in
an inheritance hierarchy. Leaf node frames inherit slots from ancestors, so
that many attributes need only be specified once, in some ancestor.
The frame system is based on XRL [4] but has been extended so that
slots may contain pattern-matching variables and attached procedures. Since
variables allow us to create generalized frames, many different instances of
a certain situation type may match the same construal frame, albeit with a
different set of resultant bindings for the variables produced by the match.
Figure 2.2 shows how inherited slots are used to match
eliciting situation frames. In this figure the slots 
through 
come
from three different frames. Frame 3 is the leaf node construal frame,
and should be thought of, conceptually, as containing all five slots. These
five slots match the slots for the eliciting situation, pictured at the
bottom of the illustration. Note that the hierarchical nature of the GSP
database is purely for the convenience of capturing an agent's concerns in
frames. There is no functional reason for the hierarchy (i.e., there is no
run-time processing dependent upon the hierarchy); the hierarchies of frames
could be compiled by collapsing them into ``flat'' representations of the
leaf node frames, so that inherited slots would be propagated down and
included in each of the containing frame's leaf-node descendants.
Figure 2.3:
Inheritance of slots in a construal frame
 |
Once candidate frames have been selected as potential matches for an
eliciting situation (using the situation frame type as discussed above), the
matcher comes in to play. This matcher uses a specialized unification
algorithm, and is discussed in chapter 3. If a match
between the emotion eliciting situation frame and a particular construal
frame succeeds then the eliciting situation is considered to be of concern
to the agent, as interpreted by the construal frame. At the same time, a
set of bindings is produced from the unification of the pattern-matching
variables and the features of the situation, and any attached procedures
(see below). The feature values contained in these bindings, and
specifications within the construal frame, are used to determine the exact
nature of the eliciting situation with respect to the agent's concerns. In
addition, these bindings are later used to specify the
possible responses as well. Output from the match process is either an
indication that the match has failed, or an instantiated construal frame
containing bindings for eliciting condition attributes such as desirability
for the agent, praiseworthiness of the act, attractiveness of the object,
and so on.
In figure 2.3, Frame 2 is shown with a
procedure attached to slot 
. Such procedures allow working memory
values to be incorporated into the match process. They also allow
fine tuning of the match process by calling predicate functions
which may base decisions on the current set of bindings. Finally,
these attached procedures may also contribute to the current set
of bindings, since the (new) set of bindings is returned from the
procedure calls.
Frame 1, slot A and Frame 3, slot D are shown as
sharing a common variable ?z. In the eliciting situation, this
feature is instantiated as 
in slots A' and D'. During
the match, since unifies with , this portion of the
unification process succeeds and a binding of the variable ?z to the
constant is created and added to the bindings list.
An important point to note is that within this context, no
event has meaning to an agent until after it has been filtered through the
concerns of that agent. Without going into the philosophical
foundations of this argument
(but see [22] for a discussion of this) it should be evident
that people work this way: extraneous cognitive and perceptual information
is filtered out of the input stream, and the relevant situations that do
pass this filter are sent on, along with the interpretation of why
they are relevant. Not all interpretations are relevant to the human
affective machinery (i.e., the information that a room is dark and that the
light needs to be turned on is not likely to cause an emotional response),
but a significant amount of binding to affective inference structures occurs
at the time eliciting situations are assessed. For example, affective
states are intertwined with expectations. To form a match between a stored
expectation and some current situation for the purpose of (dis)confirming
the expectation, one must bind features of the new situation to the expected
facilitation or blocking of the stored goals, to the expected upholding or
violation of the stored standards, and so forth.
2.3 Emotion Eliciting Condition relations
Once an initial interpretation has been made, Emotion Eliciting
Condition (EEC) relations are constructed. These are, essentially, a set
of features derived from eliciting situations and their interpretation
which, taken as a whole, may meet the eliciting conditions for one or more
emotions. Different patterns of features lead to different emotions. The
source of these features varies. Some features, such as the names of agents,
are directly derived from the situation. Other features, such as the desirability of the event, are derived from matching the situation against
the agent's concern structure (i.e., they are contained in the bindings
produced when a construal frame from the agent's GSPs is matched against the
situation frame); still other features, namely pleasingness and status, are dependent upon dynamic information and so must be derived
partially from working memory. The complete set of features comprising the
EEC relation is shown in figure 2.4, and is discussed
below:
Figure 2.4:
The Emotion Eliciting Condition Relation
| self |
other |
desire-self |
desire-other |
pleasingness |
status |
evaluation |
responsible agent |
apealingness |
| (*) |
(*) |
(d/u) |
(d/u) |
(p/d) |
(u/c/d) |
(p/b) |
(*) |
(a/u) |
| Key to attribute values |
| abbreviation |
meaning |
* |
some agent's name |
d/u |
desirable or undesirable (event) |
p/d |
pleased or displeased about another's fortunes (event) |
p/b |
praiseworthy or blameworthy (act) |
a/u |
appealing or unappealing (object) |
u/c/d |
unconfirmed / confirmed, or disconfirmed |
- Self. This attribute represents the identity of the
agent experiencing the emotion. The value for this attribute is derived from
the situation. It is always bound to an agent's name. If two agents are
involved in some situation, and it is relevant to the concerns of each of
them, then two sets of emotion eliciting condition relations will be
produced, with self bound to a different agent in each of the
different sets. Three agents will yield three sets, and so on.
- Other. This attribute represents the identity of some
other agent about whose fortunes the self agent may have an emotion.
Such an emotion will result only when pleasingness (for self
- see below) is also bound, on the basis of a relationship stored in
working memory. The value for this attribute is derived from the situation.
When the attribute is bound, it is always bound to an agent's name, but it
is never the same value as self.
- Desire-self.
This attribute represents the self agent's assessment of the
eliciting situation as an event. If he interprets the situation as one
where a goal of his is blocked then the value of this attribute is undesirable; if he interprets it as one where a goal is achieved
the value is desirable. If he does not interpret the eliciting
situation as relevant to its goals then this attribute is not bound. The
value for this attribute is derived from the agent's construal of the
eliciting situation.
- Desire-other.
This attribute represents the self agent's assessment of the eliciting
situation as an event relevant to the desires of another agent. If a self agent has a Concerns-of-Other representation for some other agent
(either specific or default - see section 2.8), he can interpret
that agent's reaction to a situation by ``imagining'' what it is like for
the other agent. If it is perceived that a goal of the other agent is
blocked the value of this attribute is undesirable, if it is
perceived as having been achieved, the value is desirable, otherwise
the attribute is not bound. The value for this attribute is derived from the
agent's construal of the eliciting situation.
- Pleased.
This attribute represents the valence of the self agent's
response to the emotions of another agent.
It is used strictly in situations in which two conditions hold: (1) an
eliciting situation perceived as a goal-relevant event ``happens'' to
another agent (bound to the attribute other), and (2) the agent
bound to the attribute self is related to the agent bound to other either by friendship or animosity. Once the value
of other has been determined from the eliciting situation, the value
of desire-other has been determined from a Concerns-of-Other
representation (chapter 2.8), and the relationship (friendship or
animosity) has been determined from working memory, the value of pleased can be determined. For example, if an agent has a friend, and the
agent observes some situation in which he perceives his friend to have
achieved one of his (the friend's) goals, the agent can be pleased about it.
This attribute is derived from the desire-other attribute in
conjunction with working memory.
- Status.
When bound, this attribute represents the status of an expectation. When
a situation perceived as an event takes place, in most cases it has no status attribute associated with it (i.e., the status attribute is not
bound). The question of expectations does not arise - an event just happens
and that is the end of it. On occasion, however, events may lead to
expectations on the part of agents. They may also confirm or
disconfirm prior expectations. The representation of such expectations is
problematic and is treated in depth in section 3.3.
Briefly, a situation perceived as an event may be unconfirmed, meaning
that the self agent perceives it as possible, but as not yet
actually having taken place. Once an agent has had this perception, a schema
representing the expected outcome is stored in the
agent's expectation database. At this point, if some situation matching this
schema comes about, then the situation will be interpreted as having been
confirmed or disconfirmed depending on the nature of the
confirming situation. Some of the values for this attribute are taken from
the situation; some are derived from prior expectations.
- Evaluation.
This attribute represents the self agent's assessment of the
eliciting situation as containing a praiseworthy or blameworthy act. If the
agent perceives that such an act has been performed by some (not necessarily
other) agent then he may evaluate this act as in accordance with, or perhaps
in conflict with, his principles. If it is in accordance with his principles
the evaluation attribute is bound to praiseworthy. If it is
in violation of his principles the attribute is bound to blameworthy. Otherwise, the attribute is not bound. The value for this
attribute is derived from the agent's construal of the eliciting situation.
- Responsible Agent.
This attribute represents the agent that the self agent holds
responsible for a perceived praiseworthy or blameworthy act. If it is bound,
it is always bound to some agent's name. It may be bound to the same name as
the self attribute. The value for this attribute is taken from the
agent's construal of the situation.
- Appealingness.
This attribute represents the self agent's assessment of the
eliciting situation as containing an attractive or repulsive object. If
this attribute is bound its value is either appealing or unappealing, according to the agent's tastes. The value for this
attribute is taken from the agent's construal of the situation. Note that
the object itself is not relevant to the generation of emotions and so is
passed in the bindings list rather than in the emotion eliciting condition
relation.
We now present two simple examples. In the first we suppose that Tom has the
goal of making it through the day without getting a speeding ticket. If he
achieves his goal then this situation is construed by him as
desirable, and if he fails to achieve the goal it will be construed
as undesirable. (We assume the simple case where has no
expectations one way or the other.) Let us suppose that the latter case
obtains, and that Tom is stopped by a policeman and given a traffic
ticket. The EEC relation derived from construing this as an undesirable
situation would have ?self bound to Tom and
?desire-self bound to undesirable, and is shown in
figure 2.5. This configuration of
features is necessary, and in this case sufficient, to meet the eliciting
conditions for a distress emotion (see section 2.1). The features for which no value is shown
may be thought of as bound to the value none, although this is not
strictly necessary. In general, each different interpretation of an
eliciting situation will produce a different EEC relation. Thus, for
example, interpretations of an eliciting situation regarding the
attractiveness of an object, which involve determination of the
appeal attribute, will not be mixed with interpretations of the
situation as containing an praiseworthy or blameworthy act, which involves
determination of the evaluation and responsible-agent
attributes. This is discussed later in more detail in the sections on
compound and multiple emotions (sections 2.4.1 and 2.4.4).
Figure 2.5:
An Emotion Eliciting Condition relation for distress.
| self |
other |
desire-self |
desire-other |
pleasingness |
status |
evaluation |
responsible agent |
apealingness |
| Tom |
|
u |
|
|
|
|
|
|
In the second example, Tom again gets a speeding ticket.
This time however, he has a friend, Harry, who observes the
situation. It is Harry's point of view with which we are concerned:
he is displeased about the bad fortune of his friend Tom,
and feels sorry for him. To represent this second point of view, we need a
second Emotion Eliciting Condition relation. Such a relation is shown in
figure 2.6 In this case ?self is bound to Harry,
?other is bound to Tom, ?desire-other is bound to undesirable, and ?pleasingness is bound to displeased. This
configuration is necessary (but not sufficient - see
section 2.4.3) to meet the eliciting conditions for pity.
Figure 2.6:
An Emotion Eliciting Condition relation for pity.
| self |
other |
desire-self |
desire-other |
pleasingness |
status |
evaluation |
responsible agent |
apealingness |
| Harry |
Tom |
|
u |
d |
|
|
|
|
Note that it is not necessary or even likely that Harry has goals
with respect to Tom not getting a ticket. He is only interested in his
friend Tom's general welfare. Probably this eliciting situation is not of
direct interest to Harry at all, although it is possible that he has, for
example, made a bet about Tom getting speeding tickets.In any case Harry's own direct personal goals can be seen as
distinct from the (possibly conflicting) goals he may have with regard to
Tom's fortunes.
Many goals, standards and preferences may be specified in a
hierarchy so that features may be inherited. The interpretation schema of a
getting-speeding-ticket situation might be represented as a set of
goals. In this approach the situation itself is considered to block a
low-level goal, getting no speeding tickets. This in turn may be a
subgoal of a retain money goal which may be a subgoal of,
respectively, retain resources, increase profits, be wealthy and be secure, which is a high-level goal. Inherited features might be that
money is quantifiable, that amount of money lost/found is
important in calculating thresholds, that money goals can have both
positive and negative outcomes, and so forth. The concepts represented in
these slots, both local and inherited, are used to interpret an eliciting
situation and to reduce it to the nine attributes of the Emotion Eliciting
Condition relation.
One final point is important. The number of EEC relations and the
number of different emotion instances following an eliciting situation can
both vary. Sometimes one eliciting situation can generate more than one
EEC relation, even for the same agent, and sometimes one EEC relation can
generate more than one emotion. These issues are discussed in depth in the
following section.
2.4 Generating emotions
Once a set of EEC relations has been generated for an agent,
the relations are used for generating emotions. This
section discusses how this is done within the context of the underlying
cognitive theory. There are a number of complex issues. The first of
these has to do with the generation of mixed and multiple emotions. Since
the discussion is lengthy it will have a section devoted to it. Following
this we show how reasoning about multiple emotions, and other
complicating issues, are dealt with to produce actual emotion instances from
the set of eliciting conditions.
Compound emotions are generated when an agent is seen as being
accountable for
some blameworthy or praiseworthy act that has like-valenced consequences
with respect to the goals of some, not necessarily different, agent. The
four possible compound emotions are shown in table 2.7.
Figure 2.7:
Compound Emotions
 |
As an example, consider again the eliciting situation where
Tom gets a speeding ticket. Suppose that Tom blames the policeman
for giving him a ticket by invoking a principle that says, in effect, that
Policemen should not stop motorists for speeding when they are
traveling at the same speed as everyone else. Taken separately, the
blocking of the get-no-speeding-tickets goal leads to distress, as
discussed above, and the violation of the principle leads to reproach. Together, however, this combination represents the eliciting
conditions for anger, and this emotion replaces the other two.
In the Affective Reasoner, a construal leading to a compound emotion
is represented as a single compound EEC relation derived from combining two
individual EEC relations into one. In the speeding ticket example, this
would mean replacing the EEC relation derived from the construal of a goal
being blocked and the EEC relation derived from the construal of a standard
being violated with a single EEC relation representing both. This EEC
relation is shown in figure 2.8.
Figure 2.8:
An Emotion Eliciting Condition relation for anger.
 |
2.4.2 Subsumption of constituent emotions
Does anger really subsume distress? Do compound emotions always
subsume their constituent emotions? That is, in feeling anger does a person
also feel distress and reproach? This is a difficult question.
Unfortunately, since we are implementing a platform that generates discrete
instances of emotions, we cannot finesse this issue. Either they do or they
do not. There can be no middle ground until the eliciting condition theory
is extended, and the EEC relations extended. For our purposes here we have
made the arbitrary decision to replace the constituent emotions with
the compound emotion.
There are some technical details to be considered with respect to
this issue. Previously we stated that the EEC relation shown in
figure 2.4 was necessary, but not sufficient to meet the
conditions for distress. All of the attributes necessary
for distress (i.e., ?self and ?desirability) are bound,
but if, in addition, the attributes for ?evaluation and ?responsible-agent are bound then the generation of distress may be
blocked. For this reason, the presence of the first two attributes is not
enough to guarantee the generation of the distress emotion.
Similarly, the presence of bindings for ?self evaluation and ?responsible-agent is necessary for the generation of the standards-based
emotions, but not sufficient since the elicitation of
these emotions may be blocked by the presence of a
binding for ?desire. In general, necessary conditions are those
specified above, sufficient conditions further require that ?evaluation
not be bound to a similarly-valenced value for the goal-based emotions, and
that ?desire not be bound to a similarly-valenced value for the
attribution-based emotions.
Except for these two cases the bindings for the remaining attributes
are ignored. For example, a binding for appealingness in an EEC
relation has no bearing on the generation of either the goal-based or
attribution-based emotions, or the compound emotions. The EEC relations in
figure 2.4 and figure 2.5 have, respectively,
seven of nine and five of nine attributes with no bindings shown. In most
cases these attributes will, in fact, have no bindings. This is because
construal frames tend to calculate and bind only those values necessary for
the particular type of interpretation they represent. For example, a
construal frame for attraction interpretations is not likely to create
bindings for responsible-agent, and a construal frame for
fortunes-of-others interpretations is not likely to have bindings for desire-self. However, sometimes bindings for these unrelated attributes do
show up in an EEC relation (a common unrelated attribute is the one for ?other). When they do they are simply ignored.
2.4.3 Multiple emotions
An agent may have many concerns relevant to a single type of
situation. Situations seen as events may facilitate or interfere with
several goals at once or may achieve one goal while interfering with
another. The situations perceived as containing a praiseworthy or
blameworthy act may uphold or violate several principles at once, or may
both uphold and violate different principles at the same time.
Situations seen as containing objects may both attract and repulse at the
same time for different reasons. These different, effectively simultaneous,
construals lead to multiple emotions.
There are no restrictions placed on which construal frames may be
placed in an agent's GSPs. Any construal frame may coexist with any other
construal frame. This means that for any given situation EEC relations may