Professional Documents
Culture Documents
Juan D. Velsquez
MIT Artificial Intelligence Laboratory 545 Technology Square, NE43-935 Cambridge, Massachusetts 02139 jvelas@ai.mit.edu Abstract
We present a computational framework for Emotion-Based Control as an alternative approach to the study of affective phenomena. Drawing upon diverse evidence supporting the affect programs theory of emotion, we show how several systems that mediate perception, attention, motivation, emotion, behavior, and motor control can be integrated into specific circuits that implement affect programs. Finally, we suggest how new processing elements such as learning, can be incorporated into these programs to extend and refine emotional responses in ways that may resemble the higher cognitive emotions.
we have developed to demonstrate and exercise the usefulness of the emotion-based control approach. Finally, we conclude with some examples of the systems we have built using this framework, and suggest how extensions to this work could be useful in describing and accounting for more complex emotional phenomena such as the higher cognitive emotions.
Introduction
Computational approaches to emotion, and affect in general, have received increased attention over the last couple of years. Several models and architectures have been proposed for a variety of domains and environments. Some examples include the use of emotions to create synthetic agents with lifelike qualities and personalities [Bates 1994; Kline and Blumberg 1999; Elliott 1992; Reilly 1996; Arafa et al. 1998], systems that reason about emotions in narrative [Elliott et al. 1998], systems that rely on emotional processing to mediate social interactions [Breazeal 1998], architectures that model the influences of emotions in behavior and learning [Caamero 1997; Kitano 1995; Seif El-Naser et al. 1998; Velsquez 1998a], and systems for recognizing the affective state of people [Vyzas and Picard 1998; Tosa and Nakatsu 1996]. The depth and extent to which these models address different affective phenomena vary widely, and so do the theories and taxonomies upon which they are ultimately based. This has led to discussions and arguments for primacy concerning which model or theory of emotion is the right one. Clearly, affect can be studied at many different levels, and all approaches offer different insights into this complex area. Different approaches work best in different domains, and the decision to follow one or the other depends greatly on the specific goals and purposes of these models. Our goal is to understand affect from a computational perspective. In this paper, we present Emotion-Based Control as an alternative approach to the study of affective phenomena. We begin by discussing the affect programs theory of emotion, which we have found to be both compelling and
those mediating fear, anger, sorrow, interest, play, sexual lust, and disgust. But what about the higher cognitive emotions? Are there brain circuits for envy, jealousy, guilt, and even love? Do these concepts fit with the affect program view, or are quite different accounts, such as the cognitive attitude theories or the social constructivist views, needed to study and understand these phenomena? These important issues can only be resolved with further evidence, and at this moment, the best available data, both from evolutionary and biological standpoints, favors the affect programs view.
systems also receive and integrate error signals from drives. The emotional systems assess the affective significance of these different stimuli and bias behavioral responses and future perception accordingly. Finally, relevant behaviors generate and execute appropriate motor actions. In the sections that follow, we describe some of these systems in more detail within the context of a robot control system. A more detailed description of the framework can be found elsewhere [Velsquez 1998b].
Emotion-Based Control
Progress in the neurobiology of affect programs has set up a foundation for the development of computational models that are more concerned with the neural substrates of emotional processing. This paper proposes an alternative approach to the study of emotion that relies on the use of computational frameworks for what we have previously defined as EmotionBased Control (EBC): control of autonomous agents that relies on, and arises from, emotional processing [Velsquez 1998b]. Our current framework, depicted in Figure 1, is an extension of the Cathexis model [Velsquez 1996; Velsquez 1997] and has been used to create a variety of EBC systems that have proven useful in providing insights about the inner workings of emotion and its interaction with other processes.
The EBC framework captures important aspects of emotional processing and integrates these with other systems that mediate perception, attention, motivation, behavior, and motor control. In other words, it essentially implements affect programs by organizing these systems into specific, functional circuits. Affective information processing in these circuits generally works as follows: Perceptual systems obtain information from the world and provide the emotional and behavior systems with stimuli features and objects. These
a set of inputs, an appraisal mechanism and a set of outputs. An important component of the appraisal mechanism consists of what we have called Releasers, which correspond to computational units that filter sensory data and identify special conditions which will provide excitatory (positive) or inhibitory (negative) input to the system they are associated with. Releasers can be innate and hard-wired (Natural Releasers), or they can be learned (Learned Releasers). The latter kind represent stimuli that tend to be associated with and predictive of the occurrence of Natural Releasers. The general form of the response or activation of each basic system is a nonlinear function of its input. Which in turn is proportional to the output of its Releasers and the strengths of their connections, or weights, as described in Equation (1):
A i = f ( R ki W ki ) k
(1)
Where Ai is the activation of System i; Rki is the value of Releaser k, and Wki is its associated weight, where k ranges over the set of releasers for System i, and f is a limiting function such as the standard ramp and logistic functions.
Currently, Yuppy has different perceptual systems based on visual and auditory processing that are capable of providing both stimuli features (e.g., color, shape, motion, intensity of sounds), and objects (e.g., people, specific sound frequencies, styrofoam bones and horses).
Each specific system (i.e., Perceptual System, Emotional System, Behavior System, Motor System, or Drive System), however, may compute its activation value in a slightly different manner as it is described in more detail in [Velsquez 1998b].
is to be issued when the behavior becomes active. Releasers might represent physical objects and specific conditions, such as battery recharger is present, as well as motivational states such as battery level is low and distress is high, which would most likely increase the activity of a Recharge-Batteries behavior. Given its several degrees of freedom and rich sensing abilities, it may be possible for the robot to engage in more than one activity at the same time. We deal with this issue by allowing active, non-conflicting behaviors to be executed simultaneously as long as the sensory and motor systems they depend on are separable. Behaviors may also inhibit or excite each other. For instance, Wag-Tail might inhibit Running and vice-versa. Whereas behaviors such as Play-With-Person might excite lower-level ones like Seek-For-Person. To account for this type of interactions, processing units representing behaviors extend the general form of response described in Equation (1) to include excitatory and inhibitory inputs from other behaviors. Currently, Yuppys behavior systems exist as a distributed network of approximately nineteen different behaviors, directed in most part toward satisfying its needs and interacting with people. Examples of such behaviors include Startle, Cower, Dance, Imitate, Greet-Person, Search-For-Bone, Approach-Bone, Navigate, Wag-Tail, Orient-To-Sound-Location, Wander, Avoid-Obstacle, Approach-Person, and Express-Emotion.
main differences include considering the excitatory (positive) and inhibitory (negative) input from other emotional systems, as well as the addition of a function that controls its temporal decay. Yuppy currently has emotional systems with natural releasers for the set of basic affect programs described before. Some of the most representative releasers and their associated emotional circuits are summarized below: Surprise: The circuits implemented by the surprise emotional system deal with novelty, anticipatory expectancy, and other issues that have been considered essential components of a general attentional system, including orienting to sensory stimuli, executive functions such as the detection of target events, and maintenance of a general alert state [Posner and Badgaiyan 1998]. Typical responses mediated by this system include LookAround, Orient-To-[Stimulus], and Look-At-[Stimulus] behaviors. In addition, the detection of sudden, unexpected stimuli might also mediate more reflex-like responses such as the acoustic startle reflex described below. Fear: The affect programs implemented through the fear emotional system deal mainly with dangerous contingencies. Concerning Yuppy, these include, among other things, situations in which its sensory systems would not work properly (e.g., dark environments for its vision system), and the detection of archetypal predators such as the blue Styrofoam horses (a.k.a the evil pool-ponies). When active, these programs release cowering and retreating responses. Joy/Happiness: This system orchestrates emotional reactions in response to appetitive stimuli, which in our robot include the presence of people and the detection of pink Styrofoam bones and big yellow balls. Typical responses mediated by these affect programs include Greet-Person and Imitate behaviors, as well as Approach[Stimulus] and Play-With-[Stimulus] behaviors. Distress/Sorrow: The affect programs implemented through this system are occupied with responding to certain kinds of stressful stimuli in the robots environment. For instance, all motivations release distress when they are not satisfied, and in extreme cases they may also release anger. Responses to this kind of stimuli typically lead the robot to pursue the needed resources in its environment (e.g., seeking people to interact with, or pink bones and yellow balls to play with). For Yuppy, affect programs serve many different purposes. First, they provide the motivational context that biases the activation of different behaviors and maintains their relevance and coherence. Second, they can facilitate attention by means of the multi-scale activation pathways described before. For instance, the Surprise emotional system modulates the Orient-To-Sound-Location behavior
when an auditory stimulus is present. Activation of that behavior turns Yuppys head towards the approximate sound location. Thus, facilitating Yuppys attention by means of other perceptual systems (e.g., vision). Third, emotions can bias perceptual systems, paying more attention to or neglecting different stimuli. Fourth, they establish appropriate emotional expressions, which are useful in social interactions. And finally, they are adaptive, providing the means by which the robot can learn from past emotional experiences and modify its behavior accordingly.
Emotional Memories. We implement this in our framework by allowing emotional systems to acquire learned releasers, which, as we previously mentioned, correspond to stimuli that tend to be associated with and predictive of natural releasers. In contrast to the fast, primary emotions elicited by natural releasers, learned releasers elicit secondary emotions, which in Damasios view occur only after we begin experiencing feelings and start making orderly associations between objects and situations, and primary emotions. These emotional memories have been modeled with an associative network comparable to Minskys K-lines [Minsky 1986], in which salient stimuli (e.g., features and percepts representing objects and agents) are connected to primary emotions when these have become active throughout the robots interaction with the world. During emotional learning, connections within this network are changed according to a modified Hebbian rule that prevents saturation of the connection weight between the new releaser and the active emotional system.
releaser (presence of the blue pool-pony) generates a fearful response (Flight behavior). Different tone frequencies played with a flute, however, do not produce any activation in the Fear emotional system, thus the Flight behavior does not become active either. If both stimuli are presented simultaneously, the Fear emotional system forms a new cognitive releaser for the sound stimulus. After only one trial, the newly formed releaser for the specific sound frequency that was paired with the blue pool-pony is capable of producing some activation of the Fear emotional system. After several more trials, the connection between the sound frequency releaser and the Fear emotional system is strong enough to produce activation of the Flight behavior, and thus an emotional memory is formed.
Summary
We have proposed the affect programs theory as a well established model and useful strategy for the study of emotions from a computational perspective. Using a robotic instance, we presented a biologically plausible framework that supports Emotion-Based Control. This framework implements affect programs by modeling and integrating important aspects of emotional systems with separate models for perception, attention, motivation, behavior, and motor control. Finally, we suggested how affect programs can lay a foundation for our understanding of a variety of emotional phenomena. Some initial steps toward this end have been presented which we believe will ultimately lead us from affect programs to higher cognitive emotions.
Acknowledgments
Work on the Yuppy robot is carried out in conjunction with Professor Rodney Brooks, Charles Kemp, and Makoto Yoshida.
References
Arafa, Y., et al. 1998. Designing and Building PSAs with Personality. In: First International Workshop on Embodied Conversational Characters.Lake Tahoe. Arkin, R.C. 1990. Integrating behavioral, perceptual, and world knowledge in reactive navigation, Robot. Auton. Syst., 6:105-122. Bates, J. 1994. The Role of Emotion in Believable Agents. Communications of the ACM 37(7):122-125. Blumberg, B. 1994. Action-Selection in Hamsterdam: Lessons from Ethology. In Proceedings of SAB94, 108117. Brighton, England: MIT Press. Brooks, R., 1986. A Robust Layered Control System For A Mobile Robot. Robotics & Automation, RA-2(1):1423. Caamero, D. 1997. Modeling Motivations and Emotions as a Basis for Intelligent Behavior. In: Proceedings of Autonomous Agents97. ACM Press. Damasio, A. 1994. Descartes Error: Emotion, Reason, and the Human Brain. New York: Grosset/Putnam. Darwin, C. [1859] 1998. The Expression of the Emotions in Man and Animals. 3rd Ed. Definitive Edition. Oxford University Press. Ekman, P., and Friesen, W. 1971. Constants across cultures in the face and emotion. In: Journal of Personality and Social Psychology 17(2):124-29. Ekman, P. 1992. An Argument for Basic Emotions. In: Stein, N. L., and Oatley, K. eds. Basic Emotions, 169200. Hove, UK: Lawrence Erlbaum. Elliott, C. 1992. The Affective Reasoner: A Process Model of Emotions in a Multi-Agent System. Ph.D. Thesis, Institute for the Learning Sciences, Northwest-
In another scenario, after locating a pink bone (natural releaser) and approaching it, the robot interacts with the person carrying the bone. Depending on these interactions (e.g., the person pets or disciplines the robot), Yuppy will create positive or negative emotional memories with respect to people, and future selection of behaviors such as approaching or avoiding them will be influenced.
ern University. Elliott, C., et al. 1998. Story-morphing in the Affective Reasoning Paradigm: Generating Stories semi-automatically for use with emotionally intelligent agents. In: Proceedings of Autonomous Agents98. Flynn, J.P. 1967. The neural basis for aggression in cats. In: Neurophysiology and Emotion, ed. C. Glass. New York: Rockefeller University Press. Izard, C. 1971. The Face of Emotion. New York: Appleton, Century, Crofts. Izard, C. 1977. Human Emotions. New York: Plenum. Izard, C. 1993. Four Systems for Emotion Activation: Cognitive and Noncognitive Processes. Psychological Review 100 (1): 68-90. Izard, C. 1994. Innate and Universal Facial Expressions: Evidence from developmental and cross cultural research. Psych. Bulletin. 115:288-299. Johnson-Laird, P. and Oatley, K. 1992. Basic Emotions, Rationality, and Folk Theory. In: Stein, N. L., and Oatley, K. eds. Basic Emotions, 169-200. Lawrence Erlbaum. Kitano, H. 1995. A Model for Hormonal Modulation of Learning. In: Proceedings of IJCAI-95. Montreal. Kline, C., and Blumberg, B. 1999. The Art and Science of Synthetic Character Design. In: AISB 1999 Symposium on AI and Creativity in Entertainment and Visual Art. Edinburgh, Scotland. LeDoux, J. 1996. The Emotional Brain. New York: Simon and Schuster. Maes P. 1991. Situated Agents Can Have Goals. In: Maes, P. ed. Designing Autonomous Agents. MIT Press. Minsky. M. 1986. The Society of Mind. New York: Simon & Schuster. Panksepp, J. 1995. The Emotional Brain and Biological Psychiatry. Advances in Biological Psychiatry, 1, 263286. Panksepp, J. 1998 Affective Neuroscience: Oxford University Press. Pfeifer, R. 1988. Artificial Intelligence Models of Emotion. In: Hamilton, V., Bower, G. H., and Frijda, N. eds. Cognitive Perspectives on Emotion and Motivation, 287-320. Netherlands: Kluwer. Plutchik, R. 1994 The Psychology and Biology of Emotion. New York: HarperCollins. Posner, M, and Badgaiyan, R. 1998. Attention and Neural Networks. In: Parks, R., Levine, D., and Long, D. eds. Fundamentals of Neural Network Modeling, pp. 61-76. Cambridge: MIT Press. Reilly S. 1996. Believable Social and Emotional Agents. Technical Report, CMU-CS-96-138, School of Computer Science, Carnegie Mellon University. Rougeaux, S., and Kuniyoshi, Y. 1998. Robust Tracking by a Humanoid Vision System. In: IARP First Interna-
tional Workshop on Humanoid and Human Friendly Robotics. Tsukuba Science City, Japan. Seif El-Nasr, M. et al. 1998. Learning and Emotional Intelligence in Agents. In: Proceedings of AAAIFS98, Emotional and Intelligent: The Tangled Knot of Cognition. AAAI Press. Staddon, J. 1993. On Rate-Sensitive Habituation. Adaptive Behavior, 1 (4). MIT Press. Tosa, N., and Nakatsu, R. 1996. Lifelike Communication Agent - Emotion Sensing Character MIC and Feeling Sension Character MUSE. In: Proceedings of International Conference on Multimedia Computing and Systems. Tyrell, T. 1994. An Evaluation of Maess Bottom-Up Mechanism for Behavior Selection. In: Adaptive Behavior 2(4):307-348. Velsquez, J. 1996. Cathexis: A Computational model for the Generation of Emotions and Their Influence in the Behavior of Autonomous Agents. S.M. Thesis. Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology. Velsquez, J. 1997. Modeling Emotions and Other Motivations in Synthetic Agents. In: Proceedings of AAAI97. Velsquez, J. 1998(a). When Robots Weep: Emotional Memories and Decision-Making. In: Proceedings of AAAI-98. Velsquez, J. 1998(b). A Computational Framework for Emotion-Based Control. In: Proceedings of SAB98 workshop on Grounding Emotions in Adaptive Systems. Zurich, Switzerland. Vyzas, E., and Picard, R. 1998. Affective Pattern Classification. In: Proceedings of AAAIFS98, Emotional and Intelligent: The Tangled Knot of Cognition. AAAI Press.