Incorporating Emotions and Personality in Artificial Intelligence Software

Keywords: computer, software, artificial, intelligence, human, emotional, personality, characteristics, AI, robots.

The Paramount Importance of Emotions in Understanding and Predicting (or Simulating) Behavior

As a person goes through her daily activities, every action she performs is associated with an underlying emotion. Thus, for instance, when she is asleep, the prevailing emotional condition is one of low arousal (i.e., low physical activity and/or low mental alertness, with slight increases in arousal during REM sleep). While breakfasting, she enjoys jolts of arousal from the coffee or tea or the heat in the food she consumes while also experiencing pleasure (if the breakfast foods are interesting and tasty) and dominance (if she has control over her choices of breakfast items). If she has to take a crowded, noisy, and dirty bus to work, then she experiences discomfort, distress, or even anxiety. When she encounters her best friend as she enters her workplace, she experiences momentary elation until she enters her office and is reminded of an unfinished task that is way overdue which results in feelings of unease.

Emotions are set off by situations (i.e., the setting in which one functions) and constitute an ever-present substrate or foundation to everyday behaviors. Thus, to understand and predict behavior or to have a mechanical robot act in human-like ways, we must have a comprehensive understanding of emotions and how these relate to behavior.

There are added complications as well. Each person brings along his own emotional inclinations (i.e., general emotional biases or temperament) to the various situations he encounters. Biological functions and drugs also influence emotions. Thus, for instance, there is the daily cycle of sleep and wakefulness, there are the physical limiations of how long a person can stay awake and work efficiently at mental tasks, there are patterns of emotion associated with illness, a headache, or the ingestion of caffeine, alcohol, or other drugs.

In short, emotions are:

set off by settings (one's immediate physical and social environment)
biased by one's temperament and mental chemistry
influenced by transitory biological factors (e.g., hunger, tiredness)
by drugs
by behavior.

A general configuration of a mechanical robot to manifest human-like characteristics would thus require inclusion of rules that relate

situational variables to emotions
personality or temperament variables to emotions
biological variables to emotions
drug intake to emotions
behavior-emotion linkages

Behavior-emotion linkages are extremely important in that (a) a specific emotional condition heightens the probability of specific sets of behaviors and (b) acting in a certain way influences how one feels. We need a comprehensive description and measurement of emotions and temperament and a way to "boil down" the innumerable variables that make up our everyday environments. Considering the overwhelmingly large number of variables, the task appears to be almost hopeless. However, as will be seen, the PAD models help reduce the variables and provide a good beginning.

The PAD Emotion Model in Mechanical Systems

Artificial Intelligence models have generally attempted to map cognitive characteristics of humans. Here, I will show how some of the most essential human qualities (i.e., emotions and personality) also can be incorporated in AI models. Importance of emotions and personality should be apparent, because these essentially differentiate humans (and animals) from strictly mechanical and computer-controlled devices.

The PAD Emotion Model has been described in considerable detail

The PAD Emotion Model

and should be reviewed before you proceed. It is an extremely general, yet precise, system for the measurement and description of emotions. Three basic dimensions of emotion are used: Pleasure-Displeasure (P), Arousal-Nonarousal (A) or general level of physical activity and mental alertness, and Dominance-Submissiveness (D) or feelings of control vs. lack of control over one's activities and surroundings. +P and -P can be used as shorthand notations for pleasure and displeasure; +A and -A represent arousal and nonarousal; +D and -D represent dominance and submissiveness, respectively. Thus, for instance, anger is represented by very low pleasure (-P), high arousal (+A), and high dominance (+D). Fear, in contrast, is represented as -P +A -D. It is seen that anger differs mainly from fear, because anger involves greater feelings of dominance (or control) than fear. More precisely,

Anger = -.51*P +.59*A +.25*D
Fear = -.64*P +.60*A -.43*D

showing that fear involves even less pleasure than anger, about the same level of arousal as anger, and considerably less dominance than anger.

Any emotion term can be described similarly using the three PAD dimensions. The descriptions are of course obtained using mini experiments in which participants are given a single emotion term and are asked to describe that feeling using the PAD Emotion Scales. Anywhere from 20 to 50 individuals independently rate a single feeling. Averages of their ratings on the P, A, and D dimensions yield the coefficients in equations such as the two for anger and fear.

An added and extremely powerful feature of the PAD Emotion Model is that it allows us to predict correlations between any pair of emotions. A correlation value can be estimated, for instance, between the feelings of dignity and anger or between the feelings of sadness and anger. Such estimated correlations would be incorporated into the emotional life of a mechanical robot. Thus, based on the moderately positive correlation between dignity and anger, the likelihood that a dignified robot will act in an angry way would be moderately high, whereas, using the negative correlation between sadness and anger, the likelihood that a sad robot will act in an angry way would be quite low.

To show you the power of the PAD Emotion Model, let us assume that we have conducted the necessary mini experiments and have obtained PAD ratings for 200 different emotions (and, indeed, we have PAD ratings for quite a large number of emotions). The correlation matrix among these 200 emotions is a 200 x 200 symmetrical table. Thus,

1/2* (200 x 200) - 200 = 19,800

distinct correlations would be given in such a table and the PAD Emotion Model allows us to compute all these correlations without actually having to do all the necessary experiments to compute those correlations. Availability of (a) the set of equations defining various emotions plus (b) the power of predicting interrelations among those emotions to an AI robot would essentially infuse the robot with life-like quality. From any given emotional starting point, the robot would exhibit emotional reactions that have a meaningful relation to one another. Thus, for instance, if the emotional starting point were one of loneliness, then the robot would also be likely to manifest feelings of depression, sadness, or boredom. Furthermore, as will be seen below, this complex of feelings would be likely to trigger behavior patterns that are compatible with it (e.g., overeating).

The PAD Temperament Model in Mechanical Systems

The PAD Temperament Model can also be incorporated into software that attempts to reproduce human-like qualities in computer-controlled (robotic) systems. This model employs three dimensions that are analogous to the emotion dimensions noted above: Pleasantness-Unpleasantness (P), Arouability (A), Dominance-Submissiveness (D). Pleasantness-Unpleasantness describes an individual's average inclination to be in pleasant or unpleasant emotional states. Arousability describes the ease with which a person is emotionally aroused (i.e., becomes physically active and/or mentally alert) and the duration for which this heightened arousal lasts. Generally, individuals who are aroused more strongly in a given situation (e.g., a near-accident while driving) will maintain the increased arousal for a longer period of time than those who were less aroused to begin with. Dominance-Submissiveness describes the average inclination of a person to view himself as being in control of his actions, relationships, and life circumstances vs. feeling controlled and influenced by others or circumstances.

The PAD Temperament Model is described in considerable detail in a separate section and should be reviewed before you proceed.

The PAD Temperament Model

Any personality trait can be expressed as a function of the PAD dimensions (i.e., in effect be broken down into three basic components of temperament, showing how much of each of the three PAD dimensions is incorporated into that personality dimension). We have data on the PAD compositions of most commonly used personality traits. PAD components for any trait depend of course on the way in which that trait is measured. Thus, for instance, the trait of affiliation, friendliness, or sociability has somewhat different PAD components when measured by Jackson (1967) or by Mehrabian (1970):

Affiliation (Jackson, 1984) = .44*P +.20*A +.26*D
Affiliative Tendency (Mehrabian, 1970) = .47*P +.24*A

The preceding two equations and those following below are taken from Mehrabian's (1996) review of the PAD Temperament Model. There is thus basic agreement in the two measures of affiliation with respect to the P and A components -- friendly people have primarily pleasant and secondarily arousable temperaments. However, Jackson's scale shows affiliative or friendly persons as also being dominant, whereas Mehrabian's scale shows friendly individuals as being neutral with respect to dominance (i.e., they are neither controlling nor are they inclined to be controlled by others in their relationships -- they simply like people and are sociable). You can take your pick of these two approaches, depending on which you consider to be more in line with the general definition of the concept of affiliation or sociability. Incidentally, it should be apparent that the PAD Model allows us to "dissect" a personality measure and examine how it is constituted and whether the characteristics incorporated in the measure make sense in terms of the conventional definition of the trait being measured.

Here are some more examples:

Neuroticism (Eysenck & Eysenck, 1975) = -.26*P +.49*A -.25*D
Trait Anxiety (Spielberger et.al, 1970) = -.43*P +.29*A -.37*D
Binge Eating (Mehrabian, 1996) = -.25*P +.22*A -.20*D

There is a similar pattern in all three traits: neurotic, anxious, and binge-eating persons are likely to have unpleasant, arousable, and submissive temperaments. There are differences among the three, as indicated by the differential weights of P, A, and D in the preceding equations. Thus, for instance, Neuroticism entails more arousability than does Trait Anxiety; Trait Anxiety involves less pleasantness than Binge Eating.

Once again, the PAD Temperament Model allows us to predict the correlation between any two traits for which PAD components have been experimentally identified. Thus, a correlation matrix can be computed for the 60 or so most commonly used personality trait measures for which PAD components are already available. Availability of the PAD equations plus the predicted correlations among traits to an AI robot would infuse the robot with "personality characteristics" that appear to have life-like quality. A robot that is configured to be neurotic would thus manifest related characteristics (e.g., anxiety, proneness to binge eating, depression, or even panic disorder). On the other hand, based on the aforementioned correlation matrix, this neurotic robot would not be likely to exhibit extroverted or nurturing traits.

Situational Effects on Emotions

The major sources of influence on emotions have been mentioned. These include situational variables (i.e., the physical and social characteristics of the environment); temperament or personality; biological factors (e.g., illness, sleepiness, hunger, tiredness); drugs (alcohol, caffeine); and behavior (e.g., running or other exercise). Of these, the situational variables are probably of greatest importance. Throughout the day, we encounter a great variety of situational conditions that in turn temporarily influence how we feel. There are some general principles that we can use to understand effects of these situational conditions on emotions.

The first and most important is a general principle that relates situational "information rate" (meaning the complexity, variability, and/or novelty or unpredictable quality of situations) to arousal levels. Environmental information rate is positively correlated with arousal. Thus, a dark and quiet room has low information rate and is conducive to low arousal and sleep. In contrast, a slightly lit child's bedroom that is housing another noisy child, a clutter of toys, voices of parents talking in the family room, and noise from the street is high in information rate and will explain inability of the children to go to sleep. "Elegant" decor involves greater complexity than "modern" decor and thus we can expect elegant decor to elicit greater arousal than modern decor. Wood paneling or wallpaper is more complex and varied than a wall painted in a solid color. Thus, wood paneled or wallpapered walls will be more arousing. When greater arousal is combined with pleasantness (as in the case of wood paneling), then the effect is "stimulating."

A second general principle relates situational control to dominance. Controlling environments elicit submissiveness; conversely controllable environments elicit dominance.An individual is likely to feel more submissive when dealing with his superiors or with individuals in positions of power and influence than when dealing with subordinates or children. Large and overpowering physical spaces (e.g., the interior of a cathedral, multi-storied lobby of a luxury hotel) elicit feelings of submissiveness. When such submissiveness is combined with pleasure and high arousal, the result is likely to be labeled as "awe," "amazement," "fascination," or "surprise."

It is difficult to formulate a general principle that relates environmental conditions to pleasure-displeasure. Nevertheless, most of the important situational factors that impinge on pleasure-displeasure have been identified. For example, social interaction (i.e., the presence of others) is generally likely to be pleasant. Exceptions are in some crowded conditions where presence of others is experienced as an impediment to one's activities, resulting in feelings of displeasure, high arousal, and submissiveness (i.e., discomfort, irritation, stress). Brighter and more even lighting is more pleasant; colors vary considerably in their emotional effects; noise is unpleasant and high-information (i.e., varied, complex, unpredictable) noise is particularly stressful.

Some of the major situational effects on emotions are reviewed by Mehrabian (1980a; 1980b, chapter 8). It is therefore possible, though not easy, to catalogue major elements of situations (both physical and social) and to translate these coded elements into their emotional counterparts. A reasonably exhaustive catalogue of situations and their emotional correlates would require an enormous amount of work, but it is doable.

Emotions and Behavior

Every emotional state predisposes a person toward certain select sets of behavior. Here are some examples:

a more pleasant emotional state is more conducive to a person acting in a friendly and sociable manner with others; conversely, an unpleasant emotional state (e.g., a headache) tends to heighten chances that the individual will be unfriendly, inconsiderate, or even rude with others
very high arousal is detrimental to concentration on complex tasks and is likely to result in avoidance of, or poor performance on, such tasks
a more pleasant and, for most tasks, a less aroused emotional state enhances one's desire to work (so, if your office is uncomfortably cold or hot, you are less likely to want to do the jobs that await your attention)
sexual desire increases with pleasure, particularly when pleasure is combined with greater arousal and/or dominance; sexual dysfunction (e.g., impotence, inability to experience orgasm) becomes more probable when the individual is experiencing displeasure, particularly with displeasure is combined with greater arousal and/or submissiveness
of all emotional states (i.e., all possible combinations of high vs. low P, A, and D), the pattern of -P-A-D or boredom is the most conducive to nagging desire for food and overeating.

Behaviors, in turn, influence how a person feels. For instance, eating is generally arousing and pleasant and these emotional effects vary depending on the complexity, variety, and quality of the food consumed. Repetitious physical behavior (e.g., rocking in a chair, pacing back and forth in a limited space, walking, jogging, rowing) tends to reduce arousal and increase pleasure and dominance. Thus, repetitious activities are common among unmedicated schizophrenics or autistic children who experience uncontrollably high arousal states. Exercise is also an excellent way in which anyone can counteract anxious or depressed mood. Because exercise reduces arousal, it is even more beneficial in counteracting anxiety than it is in relieving depression.

In general, from the standpoint of simulating human behavior, it is far more important to focus on emotion effects on behavior than on behavior effects on emotion. This is because changing situations throughout the day have continuously powerful effects on emotions and thereby influence action. Action, in turn, tends to reflect the emotional substrate and, in this way, the behavior-emotion linkage is strengthened. There are exceptions, as when a bored person eats (an action that increases pleasure and arousal and thus temporarily counteracts the underlying emotion). So, early AI models can probably perform a reasonably satisfactory simulation by simply focusing on the following chain: Effect of situations on emotions and then the impact of emotions on behavior.

Illustration of a Robot that Incorporates the PAD Models

Imagine interacting via a keyboard with a computer. The conversation proceeds along these lines:

You: "How do you feel?"
Machine: "Where am I?"
You: "You are at home."
Machine: "Describe my home."

At this point the computer screen displays a list of situational qualities and alongside each quality there is a sliding scale that you can use to show if the quality is relevant (i.e., if you have information about the quality) and the intensity of that quality (e.g., "brightness of lighting," "noise level," "people present," "amount of conversation," "temperature,"). Each of these situational qualities has already been coded and translated into its emotional counterparts; thus the machine can compute a weighted average for the pleasure-displeasure (P) values across all the qualities that you indicated were present in various degrees. It separately can compute two more weighted average values for arousal-nonarousal (A) and dominance-submissiveness (D). Once the averaged P, A, and D scores are available, the computer searches for emotion labels that are closest to those three values and outputs one of these labels.

Machine: "I feel vigorous"
You: "How else do you feel?"
Machine: "I feel competent."
You: "Describe yourself."
Machine: "I can't, you need to tell me first what kind of person I am."

At this point the computer screen displays a list of personality traits (e.g., sociable, friendly, achiever, self-confident, empathic, anxious, neurotic) with a sliding scale alongside each so you can use the scale to show the relevance and intensity of some of the items in the scale). P, A, and D values of each of the personality traits are already coded. The machine computes a weighted average (i.e., weighting each trait by the intensity of the trait you have specified) for each of the three P, A, and D values and then finds a personality label that constitutes the single best summary of the PAD values.

Machine: "I am a docile person. Would you now like me to give you a more accurate description of how I feel (in reference to your previous question)?"

You: "yes."

At this point, the machine adjusts the PAD values for "vigorous" or "competent" using the personality-trait PAD values for "docile." This adjustment is done by weighting the situation-induced PAD values by 2/3 and the personality-trait PAD values by 1/3. New weighted PAD values, computed in this way, are again matched against emotion labels and the result is:

"I feel satisfied is a more accurate description of how I feel."

Reason for the shift is as follows: "vigorous" is a pleasant, aroused, and highly dominant state; "docile" is a pleasant, unarousable, submissive personality. Weighting "vigorous" by 2/3 and weighting "docile" by 1/3 yields an emotion condition that is pleasant, slightly aroused, and slightly dominant and is approximated with the label "satisfied."

You: "How else do you feel?"
Machine: "I feel sort of modest."

Note: The qualifier "sort of" is added because "modest" is not quite high enough in pleasure to match the computed P value.

Focused Simulations with Limited Objectivers

Simulation models that have limited applications can be developed with greater ease and are illustrated below.

Excersice Equipment

Exercise equipment, such as a rowing machine, can include a "companion." This companion will be a robot that communicates ("literally talks") to the person exercising by sending messages that are a function of the person's exercise behavior. Specifically, exercise behavior (e.g., high level of exertion, slowing pace) would be translated into emotion terms and these terms would be incorporated into the messages sent to the exerciser by the "companion." For example, high pace of exercise would be coded as involving high arousal and high dominance. This in turn would be translated into a variety of emotion messages from the robot. The emotion terms used by the robot would be consistent with high arousal and high dominance, such as, "This is exciting. I feel powerful!"

Robot generated messages could be made interesting by allowing the user to specify (a) the gender of the "companion" and (b) the personality of the "companion" (e.g., humorous, wise adviser, psychologically savvy friend, exercise expert). Also, a variety of messages would be available to the robot for each input condition (i.e., each category of action by the person exercising to which the robot is programmed to respond). These various messages would be used by the robot in a random order so as to reduce monotony and predictability of the robot's behavior.

Automobile

This application does not deal specifically with emotions and AI as discussed above, but you still may find it of interest. The owner of the automobile would be able to select the gender and personality of the robot "companion." For example, the "psychologically savvy" robot would send off the following messages in stop-and-go traffic: "I know this is getting kind of monotonous (or tiring), so try to keep your mind on other, more pleasant things. What are you going to do for fun when you get home?" In this case, the robot would be acting as an emotionally intelligent frined who cues the driver in emotionally constructive directions. When the automobile engine oil is low, the "expert technician companion" would say something like "Engine is getting too hot and you should be getting worried." If the driver is braking while taking a fast turn, the "expert technician companion" would say "Do your braking and slow down before the turn; not while you are already in the turn." If the driver has left the trunk open and is beginning to drive off, the "humorous companion" (ideally voiced by a comedian, such as, Eddie Murphy) would say "You've left the trunk lid open and I can't think straight with all this wind blowing up my butt!"

Political Campaigns and International Affairs

More involved simulations can be developed to study and predict behaviors of an adversary in a political campaign, the likely reactions of voters, or behavior of another nation's political leader . Simulations of voter choices would involve coding the PAD values associated with various political issues (e.g., amount of taxation, freedom to have abortions, choice of school vouchers). Differential importance of PAD values for various voters would be coded by considering voter demographics. In simulations of a political adversary, the temperament of the adversary would be a key ingredient in the model.

Overview

Developing a complete behavior simulation model is an extremely ambitious task and is tantamount to formulating a very thorough and general psychological theory of behavior. Various facets of the PAD models have been under development over the last 30 years. It is thus possible to summarize many situation-emotion relationships (e.g., What is the effect of light intensity on emotions? What is the effect of color on emotions? How do a person's physical characteristics (e.g., eye size and shape) impact the emotions experienced by someone observing that person?). Also, it is possible to outline some of the most important and common effects of drugs on emotions. More importantly, we also can broadly sketch effects of emotions on behavior. Finally, we can also take into consideration the complicating effects of temperament on emotion-behavior linkages. Developing more delimited and focused simulations are far more doable.

References

Eysenck, H.J., & Eysenck, M.W. (1985). Manual of the Eysenck Personality Questionnaire. London: Hodder & Stoughton.

Jackson, D.N. (1984). Personality Research Form manual. Port Huron, MI: Sigma Assessment Systems.

Mehrabian, A. (1970). The development and validation of measures of affiliative tendency and sensitivity to rejection. Educational and Psychological Measurement, 30, 417-428.

Mehrabian, A. (1980a). Shaping environments that fit your inner landscape. Creative Living, 9 (3), 7-11.

Mehrabian, A. (1980b). Basic dimensions for a general psychological theory: Implications for personality, social, environmental, and developmental studies. Cambridge, MA: Oelgeschlager, Gunn & Hain.

Mehrabian, A. (1996). Pleasure-Arousal-Dominance: A general framework for describing and measuring individual differences in temperament. Current Psychology: Developmental, Learning, Personality, Social, 14, 261-292.

Spielberger, C.D., Gorsuch, R.L., & Lushene, R.E. (1970). Manual for the State-Trait Anxiety Inventory. Palo Alto, CA: Consulting Psychologists Press.