An Emotion Theory Approach to Artificial Emotion Systems for Robots and Intelligent Systems: Survey and Classification

1 Introduction

Rumbell et al. [20] proposed a classification of artificial emotion software systems using a computer scientist’s approach to classification, based on four questions: “What is the action selection method (arbitration to command fusion)?”; “How is this architected (reactive to deliberative, symbolic to neural, continuous to discrete, hierarchical to distributed)?”, “What roles do emotions serve (e.g., action selection, adaptation)?” and “Which emotional model is used (basic or dimensional)?” While the classification provides a foundation for making architectural decisions for robotic systems, restricting ourselves to such a technologically based classification scheme could limit our ability to predict how such systems correspond to human emotions and, therefore, the extent to which such systems will be convergent with the expectations of humans in an integrated artificial intelligence (AI) – or robot–human system. Such a technologically focused classification perspective may also limit our vision of how systems can take advantage of psychological research into the benefits of human emotions during decision making in complex environments.

We present a psychologically focused extension of the third question posed by Rumbell et al.: “What roles do emotions serve (e.g., action selection, adaptation)?” By classifying architectures using a model similar to how psychologists have classified theories of emotions throughout time, from Aristotle to the current era, constraints on the answers to Rumbell’s other three questions can be identified.

An eight-question model was developed by Power and Dalgleish [18] to classify theories of human emotion. We propose a 10-question model for classification of robots and AI systems using artificial emotions, as shown in Table 1, side by side with the model offered by Power and Dalgleish.

By investigating robotics and AI systems with this Emotion Theory model classification scheme, we can relate the development of artificial emotions directly to a rich literature of more than 2000 years of theory on emotions in human systems. Examining how these systems conform to, or are distinguished from, various descriptions of human emotion may lead to paths for improvement of these systems, may help us understand where such systems may or may not be compatible with future research in these areas in emotion theory and neural sciences, and may enable us to understand more about human emotions by providing a bridge back to using such systems as models for normal emotions or emotional disorders in humans.

Numerous technological methods exist, and new approaches are constantly being introduced. Because this classification scheme can be used for robotic and AI systems at any stage of development from business requirements specification, to system architecture, to deployment and maintenance, we believe it may provide a useful way of characterizing a system using artificial emotions in such a way that architects may more effectively weigh the benefits and disadvantages of existing and emerging technologies. The scheme may make it possible to distinguish technologies that work in general for artificial emotions from those that only work well when artificial emotions are characterized in specific ways.

2 Theories of Human Emotions

Given the importance of emotion to all aspects of human life, it should come as no surprise that the topic was discussed by many of the great philosophers – Plato, Aristotle, Spinoza, Hobbes, and Hume. More recently, emotions have become of interest to cognitive theorists as well. The philosophies of emotion have become connected with many other disciplines, including psychology, evolutionary biology, neuroscience, and robotics. It might be helpful to summarize a few of the theories of emotion as proposed by both the greats of philosophy and modern cognitive scientists to gain a better understanding of how these could be used in the design of artificial emotions. We will use the model of the eight questions proposed by Power and Dalgleish [18] to facilitate comparison of the various theories. This section provides responses to each of the eight questions for five emotion theories, as well as the names of some of the theorists associated with those theories by those authors. We hope that we have not misrepresented the theories in our quest for brevity. A discussion of the possible technological implications of each of these theories is included. These may provide a starting place in the literature for technologists interested in extending their artificial emotion designs.

Feeling Theory: Power and Dalgleish’s eight questions (Plato, Rene Descartes, William James, and, more recently, Damasio and Prinz):

1. For Plato, Descartes, and James, emotions are a part of the soul and not of the body, and are in opposition to rational thought. All feelings (including emotions) are a by-product of the movements of spirits within the pineal gland, or the more modern interpretation provided by Damasio and Prinz of emotions being the result of somatic changes in the body.

2. Emotions are a by-product of movements within the pineal gland. Emotions move us to the extent that the excitation stimulus can be harmful or profitable. Some emotions are mixtures of the six primary passions, which are irreducible.

3. They each have distinct motions of the bodily spirits.

4. An excitation factor (e.g., as seeing a large animal) causes changes in the soul, which results in spirit movements of which we become aware and perceive as emotion (e.g., fear). Emotions are not learned from the environment or experience; rather, they are intrinsic to the soul. Emotions directly cause actions (e.g., moving the legs to run from what is fearful). A later refinement of William James suggests “we feel … afraid because we tremble” [18]. Recent updates to this theory proposed by Damasio and Prinz suggest that somatic changes in the body (such as heart rate, blood pressure, and facial expression) cause changes in mental states that are what we think of as emotions [13].

5. Emotions (or the so-called passions) provide a defensive mechanism in the face of imperfections in the natural world.

6. These are not addressed.

7. There are six so-called primary passions: wonder, joy, sadness, love, hatred, and desire. Descartes “constructed a host of complex ones [emotions] out of these six” [21].

8. Plato saw emotional disorders as the result of lack of rational control or dominion over the emotions. According to Descartes, the will keeps the passions from overwhelming us and causing emotional disorders.

Subscribing to the Feeling Theory could have a number of technological implications. The Feeling Theory suggests that emotions directly cause actions. They are not cognitively processed first. This would mean that the actions of the AI or robotic system would depend directly on the emotional state. In the Feeling Theory, there are six emotions that are irreducible. From a technology perspective, these six emotions are binary. Complex emotions would be characterized by patterns of these binary variables. Emotional control in such systems would be done by suppression of the emotional expression, rather than substitution of expressions that differ from the internal emotional state of the system. Emotional disorders in such systems, if modeled, would be characterized by failure to suppress emotional reactions. Constraints to questions posed in Rumbell et al. include an action selection mechanism on the side of arbitration with emotional reaction defined by the pattern of emotions and rational suppression as an opposing force.

Behaviorist Theories of Emotion: Power and Dalgleish’s eight questions (James Watson, B. F. Skinner, Gilbert Ryle):

1. Emotions are inherited patterns of reactions that involve systemic changes in the body, “particularly of the visceral and glandular systems” [18].

2. Emotions are irreducible.

3. Each emotion has a different systemic pattern.

4. A particular systemic bodily pattern is recognized by the brain. This pattern recognition is innate. This primes the individual experiencing the emotion to associate the particular environmental conditions with that emotion and seek to increase or decrease the probability of replicating the conditions now associated with that emotion.

5. Different emotions (which are inherited reactions to systemic bodily patterns) provide differentiated feedback to operant conditioning.

6. Temperament is an inherited sensitivity to detecting certain systemic bodily patterns. For example, some people are more or less inclined to detect the pattern that indicates fear. Moods are temporary inclinations to detect certain systemic bodily patterns. For example, after a particularly bad day, a person may be more sensitive to fear patterns than love patterns.

7. Fear, rage, and love (more akin to sexual drive).

8. The abnormal emotions would be the result of associations of many or most events with a particular systemic bodily pattern. For example, a person who is afraid of everything may be the result of poor environmental conditions interacting with an increased inherited sensitivity to the fear pattern.

Technologically speaking, each emotion from a Behaviorist point of view would be triggered by specific patterns of sensor readings or information. The presence of these emotional reactions remains constant. Full emotional control would require a system capable of avoidance of, attraction to, suppression of, and amplification of sensing patterns that result in particular emotional states. The fact that this emotional control of attention to sensor information is a function of prior emotional states means that the system must have a means of ignoring, amplifying, or moderating the awareness of sensory data based on their association with prior emotional states. This would suggest that the architecture would need to track emotion over time, associate sensory patterns with them, and adjust the processing of stimuli accordingly. The Behaviorist model of emotions would also suggest that emotions do not need to be learned from the environment; they can be encoded directly into the system.

Emotional disorders could be modeled by increasing the sensitivity of a system to certain emotions while suppressing others, which would result in overexpression of the favored emotional state even when sensor states would normally trigger other emotions. This suggests that systems that implement the Behaviorist model should learn the appropriate preferences for emotional states for a given environment.

Using Rumbell’s model, architectures using the Behaviorist Theory would likely tend toward arbitration, reactive, symbolic, discrete, distributed models with basic emotions. Emotion roles supported would include motivation, attentional focus, alarm mechanisms, and action selection.

Evolutionary Approaches: Power and Dalgleish’s eight questions (Robert Plutchick, Robert Frank, Charles Darwin, Paul Ekman):

1. Emotions are affect programs.

2. Emotions are irreducible; some are innate and universal, while others are learned from the environment.

3. Differences in affect distinguish one emotion from another.

4. Emotions are inherited reactions to specific situations. They are affect programs that prepare the individual to respond in a way most likely to support survival, dominance, mating, and effective affiliations. These behaviors are selected by natural selection. For example, if an organism responded to an attack with fear (running away or freezing in place) or rage (barring teeth), then the reaction most likely to support survival will be the programmed response retained by the species. Therefore, emotional experiences differ between individuals only where these goals will not be adversely affected by the behavior. Naturally occurring variations in the behaviors will be pruned by natural selection to the few that are most likely to promote the goals of survival, dominance, mating, and effective affiliations.

5. Emotional expressions serve evolutionary functions particularly in the roles of survival, dominance, mating, and affiliation.

6. This is not addressed.

7. Generally six are assumed to be universal: happiness, sadness, anger, surprise, and disgust, based on commonality of affect across cultures [9]. All other (infinitely many) emotions are assumed to be learned reactions to the environment and may therefore be different for each person. Plutchick claimed eight basic emotions and claimed that these are found in all organisms: fear/terror, anger/rage, joy/ecstasy, sadness/grief, acceptance/trust, disgust/loathing, expectancy/anticipation, and surprise/astonishment.

8. This is not addressed.

Fully adopting an Evolutionary approach to behaviors would require an architecture capable of generating new reactions to existing emotions. These reactions could be retained or suppressed on the basis of the results of the action. Ideally, additional systems with the same six basic emotions would be able to inherit these retained behaviors from other systems.

Each system would have the six or eight basic emotions; however, individual implementations would have an indefinite number of other emotions that are learned from the environment. Such indefinite behaviors could be modeled by non-semantic networks such as neural networks with stimuli as inputs and actions as outputs, with the specific structures and combination of weights in the network being termed “emotions.” The complexity of these emotions could be modeled by the number of hidden layers or the number of nodes in each layer.

In Rumbell’s model, an Evolutionary approach would suggest that the learned emotions would be neural rather than symbolic, continuous rather than discrete, reactive in real time, but deliberative in the decision to retain or discard emotion–reaction associations.

Functional and Cognitive Theories of Emotion: Power and Dalgleish’s eight questions (Aristotle, Thomas Aquinas, Baruch Spinoza):

1. Emotions are defined by their roles within psychology. Emotions include reactive emotions such as anger, sadness, and fear, as well as the higher cognitive processes for such emotions as jealousy and envy.

2. Emotions are labels that we place on certain collections of action tendencies. In this way, specific action tendencies that make up a pattern of action tendencies could be thought of as the constituents of emotion.

3. The particular pattern of action tendencies for any given emotion for any given person is distinct from the action tendencies for another emotion. This makes the list of possible emotional states indefinite.

4. For Aristotle, emotions are contextual reactions to situations. An excitation object interacts with a current state of mind to produce a stimulus (an appraisal) that gives rise to an emotion, which then becomes an increased or decreased tendency to act in certain ways. For Thomas Aquinas, there is an initial impulse either to approach or avoid a particular stimulus that is a part of what Aristotle would call the state of mind. Some of these impulses (or primary emotions) are inherited (e.g., the impulse to avoid a shark). For Aquinas, there is a secondary cognitive process that can give rise to such emotions as fear or sorrow, which are secondary emotions. For Spinoza, the primary emotions are reflective and non-cognitive, and these are combined with a cognitive process of identification of causation to engender the more complex emotions. For example, love is the assignment of the cause of pleasure with the presence of a particular individual.

5. Emotions give us the capacity to act. They provide the action tendency for certain behaviors.

6. For Aristotle, moods and temperament would be a tendency to imbalance between emotional states.

7. Aristotle presented 10 specific emotions with two valences but did not propose that there were a limited number of emotions. The emotions he listed were as follows: positive valence emotions – calm, friendship, favor, pity – and negative valence emotions – anger, fear, shame, indignation, envy, and jealousy. Aristotle saw some emotions as the opposite of others (e.g., anger is the opposite of calm). Therefore, some emotions cannot co-occur with certain others. Thomas Aquinas would say that there are two primary emotions: attraction or avoidance, and that there are many more cognitively derived emotions. Spinoza would say that there are three primary emotions: desire, pleasure, or pain. For Spinoza then, emotions are anything that is a combination of these three primary emotions and an idea that associates these with a cause. Thus, “tickle” would be an emotion that associates a particular action of one’s fingers with the result of pleasure.

8. For Aristotle, emotional disorders are the result of imbalance between reason and emotion or between emotions. Correction of an emotional disorder is done by rebalancing the experiences of emotion and reason.

Architecture employing a Functional or Cognitive theory of emotion would need to maintain three states: the current state of mind of the system, a state engendered by the pattern of sensory information, and the resulting emotional state. Some emotions would be directly connected to the pattern of sensory information, while others (the cognitive emotions) would be the result of combining the state of mind with these patterns. The system may constrain certain emotions from co-occurrence. Emotional disorders could be modeled by inappropriate suppression of, or emphasis on, the cognitive emotions. This suggests that systems that implement a Functional or Cognitive model of emotions need to incorporate an additional layer that learns the conditions under which emotions should be primarily cognitive.

Constraints for Functional or Cognitive models using the system described by Rumbell et al. could include tending toward designs that include two or three subsystems: one that is reactive, symbolic, discrete, distributed, and has a basic emotional model; a second that is deliberative, neural, continuous, hierarchical, and has a dimensional emotional model; and possibly a third that learns to prioritize certain emotions or to give priority to one or the other of the previous two subsystems proactively, depending on expected environmental conditions.

Appraisal Theories – 20th Century Cognitive Theories of Emotion: Power and Dalgleish’s eight questions (Richard Lazaryus, Nico Frijda, Klaus Scherer, Russell, Magda Arnold, Anthony Kenny, Erroll Bedford, Richard S. Peters, G. Pritcher, Robert Gordon, William Lyons):

1. Emotion is the label used to describe unusual physiological state change caused by particular appraisals.

2. Emotions are labels that we place on certain collections of physiological state changes that occur after particular appraisals. In this way, specific emotions could be thought of as consisting of patterns of physiological states in combination with appraisal patterns that immediately precede these physiological patterns. Some of these theorists have modeled emotions as positions in multidimensional space. Klaus Scherer proposed 18 dimensions. Russell proposed only two or three dimensions, e.g., arousal and valence.

3. Emotions are distinguished by their unique combinations of appraisals and physiological state changes.

4. For William Lyons, an external event results in attention, which causes an appraisal of the situation, which results in physiological change, which then causes a pattern of action tendencies that lead to particular behaviors happening with greater or lesser frequencies. Appraisals may be either conscious or unconscious. In some cases, both will occur, but one or the other may win out in any given situation, such as when a person consciously determines that spiders are no threat yet still unconsciously appraises a spider as dangerous and runs away.

5. Emotions are function in that they are used to select beneficial reactions to deal with specific situations.

6. Moods are short-term, either focused or unfocused, dispositions to particular appraisals. Temperament or personality would be considered a longer-term disposition to particular appraisals.

7. There are indefinitely many emotions. Some may be opposite to each other given that the accompanying physiological changes are a limiting factor (e.g., one cannot both raise and lower one’s blood pressure simultaneously).

8. Emotional disorders are caused by conflicting cognitive appraisal that happen at different levels of the cognitive system. In some cases, one of the conflicting appraisals may be unconscious. One example of emotional disorders caused by this conflict is phobias.

Architectures employing an Appraisal Theory or the 20th Century Cognitive Model are more complex even than those for the Functional Theory. These architectures must maintain separate action selection processes simultaneously. Action can begin to occur on the basis of processes that complete first, but then be changed on the basis of processes that complete at a later time. These architectures can model abortive actions, such as moving toward something and then switching to another action plan. This means that when a new action plan is determined to be “better,” the system must be capable of calculating the path from the current state toward the path intended by the new decision. Appraisal-based architectures could include three or four subsystems: the first three would be similar to those described under the Functional Theory, with the exception that the third system may predetermine action delays for one or the other of the first two subsystems; the fourth subsystem would facilitate reconciliation of partial actions executed with actions intended from later decisions.

3 Review of Published Works on Systems with Artificial Emotions

The published works were selected by using the phrases “robot emotion” and “artificial emotion” in Google Scholar. Initially, more than 50 papers were investigated on the basis of being in the first groups returned from each of these queries. Papers were eliminated if they did not describe implementations, simulations, or architectures of systems using artificial emotions. For each of the remaining works, the system described was classified by answering the 10 questions proposed earlier in this paper in Table 1. The results are included in Tables 2–15. References for additional reading include de Sousa [8], de Frietas and Queiroz [7], and Ziemke [26].

Arkin et al. [1] describe the implementation of Sony’s AIBO and SDR entertainment robotics systems. The authors focus on two goals in the design of AIBO: to examine and model the behavior of creatures in their natural environment in order to model the natural, and therefore generally expected behaviors, and to incorporate a model of “motivational behavior” (e.g., emotions) that supports human interaction with the robotic system. The ethological model provides a range of behaviors from which to select, while the emotional model provides the mechanism that selects the appropriate behavior. The result is a robotic system that interacts with human companions in intuitive and engaging ways.

Kwon et al. [15] describe an intricate system that incorporates a complex model of emotions, implemented in an autonomous robotic platform. They describe an autonomous robot capable of recognizing emotional content in natural language and in touch gestures, e.g., the difference between a gentle and abrupt touch. The robot’s own emotional state can be displayed in any of several modes, including facial expression – displayed on an on-board computer screen – gestures, and generated music. The emphasis in Kwon et al. is on emotional recognition and expression through multiple modalities, offering a wide range of experience both for the robot to recognize emotions in others, and for humans sensitive to different modes of expression to recognize the emotional state of the robot.

Breazeal and Brooks [3] describe the implementation of Kismet, a sociable although non-mobile robot capable of expressing a range of affect through control of facial actuators. Kismet contains a number of cameras and a microphone; views from the cameras are combined to produce a stereoscopic sense of the robot’s environment. Kismet’s primary goal is to encourage humans to interact with it in various natural ways, especially focusing on the robot’s ability to display its emotional state through facial expression and posture. For instance, if the robot is “sad,” its ears will droop. If it is “angry,” its brows will be drawn together, forming a collection of expressions easily interpretable by humans.

Chlebicki et al. [4] also present a system employing an implementation of a model head and conversational system. The authors describe two implementations of the mechanism of expression and a speech module that does not rely on truly synthesized speech, but on recorded voice files, from among which the system selects one or another for playback as part of the conversational process. Their initial implementation dealt favorably with large social situations, interacting with crowds of visitors when presented at conferences in Wrocław and Vienna. The authors note the importance of the robot’s expression of emotion when dealing with large numbers of humans.

Gavrilov [10] proposes a relatively simple model of emotional state determined by a neural network alongside what amounts to a prior knowledge base of fundamentally moral rules. He uses as his example the Three Laws of Robotics proposed by Isaac Asimov as a form of prior knowledge. The system he describes uses emotional state as a training input to a neural network, with the emotional state determined by conformance to the rules; if a situation is perceived as “breaking the rules,” e.g., if a human is in danger, the system expresses this as negative emotion: it feels bad about the situation and uses this as a training input to the neural network that determines actions that will minimize bad feeling and maximize good feeling.

In the EMOBOT, Goerke [11] uses a hierarchical architecture consisting of a sensory upstream, an actuary downstream, and an internal value system/learning action controller. In the paper, he describes the progression of the internal values-based architecture through two simulations (the grid world EMOBOT and the real values EMOBOT) and then into a real-world implementation of a six-wheeled robot equipped with ultrasonic, infrared, and ambient light sensors. In each stage in the development, he has increased both the complexity of the behaviors and the internal value models. Goerke uses the terms “Drives” – for the states that depend directly on the pattern of sensory information – and “Emotions” – to describe states that are interrelated mathematical functions based on the degree of satisfaction (or lack thereof) of the drives. He uses these terms for their metaphorical value; he explicitly states that he makes no attempt to model the related psychological concepts.

Miwa et al. [16] provide a detailed account of the mechanism and emotion of an autonomous robot. Their robot incorporates visual, auditory, tactile, and olfactory senses, and is capable of a large range of expression, incorporating mechanical actuators for the position of various components of its face, and a thin electroluminescent sheet allowing it to change “skin” color slightly, when expressing, for example, anger or embarrassment. The robot’s personality is based on a separable Sensing Personality and Expression Personality, with both personalities coordinating through a mental model that determines overall emotional state as a position within an emotional space determined by axes of pleasantness, uncertainty, and activation.

Ptaszynski et al. [19] focus on the goal of the interaction between a human and a conversational agent. The agent attempts to understand the emotional expressions of its human conversational partner and determine whether those emotions are appropriate to the context; it can then choose whether to be directly supportive of its human partner, or suggest to its partner that the emotion it has expressed is inappropriate to the circumstances. Determination of whether an emotional expression is appropriate, convergent with social norms, is accomplished through a data store of emotional affect terms drawn from “mining” publically available documents found on the Web. Using different reactive models, e.g., task-focused or humor-focused, Ptaszynski and his colleagues were able to show stronger positive reactions in the human conversational partner to some models than to others, e.g., humans reacted more favorable to models that made their point with humor.

Schröder and McKeown [22] describe a European project that is focused on the development of a system with “social and emotional intelligence,” or what they term a Sensitive Artificial Listener. The paper describes the implementation and behaviors of four different personalities that were tested (Poppy, Spike, Obadiah, and Prudence), each with a different affect focus (cheerful, aggressive, gloomy, or pragmatic). The artificial personalities react to listeners with both body language (such as nodding), facial expression, and a narrow set of utterances that are intended to appear consistent with their personality, and designed to steer the user’s affect toward that of the agent. This system’s goals are convergent with those of the conversational agent described by Ptaszynski et al., but place less emphasis on verbal or written language and more on the non-linguistic cues used to affect emotional change.

Steunebrink et al. [23] present a mathematical operation-based model of emotion, emphasizing the role that emotions play in behavior at the level of action tendency; that is, the effect that emotions have on the set of actions from which the agent chooses. Emotions here are presented in a formal mathematical model derived from the work of Ortony et al. [17], describing emotions in terms of objects, actions, and events. While the authors do not describe an implementation of their method, unlike many researchers in the field of artificial emotions, their model incorporates a very wide range of actions in response to emotional conditions, such as idling, the “take a deep breath and count to ten” reaction to negative emotion. The range of actions they model also takes into account reconsidering the agent’s actions in light of new information or changing state, and the option of uncommitting to an action, and replanning, based on feeling differently about the goal.

Suzuki et al. [24] describe a small unobtrusive wheeled robot equipped with camera and ultrasonic sensors that can sense the dominant emotions of the human companions, and is designed to increase their creativity through reflecting the perceived dominant emotional state of its human companions. The authors make explicit reference to using modern psychological theories to inspire their design of self-adaption of competing emotional states resulting in one single dominant expressed emotion at any given time. Detected human gestures and movement are mapped onto emotional states of the robot, which are then expressed as styles of movement, robot behavior, visual media, environmental lighting control, and music. Similar to both Schröder and McKeown and Ptaszynski et al., Suzuki et al. rely on an emotional architecture to influence the internal emotional state of humans. In this case, the authors employ the mobility of robotics and an emotional architecture to inspire human creativity through the establishment of emotionally supportive environments.

Wallach et al. [25] postulate that feelings, rules, and virtues can be used to extend the artificial general intelligence (AGI) model, LIDA, to incorporate higher-order cognitive tasks such as those of Machine Morality. In doing so, the authors provide background on the general state of AGI and the details of the LIDA architecture. While they do not describe a system that has been implemented, they suggest that LIDA, because of its ability to support multiple inputs, may provide an ideal platform for modules that convert sensory information to emotional responses constrained by moral values. Their proposed architecture, with a rich multidimensional emotions that are retained over time, affect future perception and emotional interpretation, and are learned in the context of societal values, maps more closely to modern psychological theories than the other architectures reviewed herein.

Jiang et al. [12] have implemented a model of emotional agents in a simulated “Tileworld,” where virtual agents attempt to cover holes with tiles. Multiple agents in the world can communicate information about the locations of tiles and holes. In the model presented by the authors, some agents can learn to trust and cooperate with other agents, or “lie” in their communications, and so become untrustworthy. They are able to show that agents that track how they “feel” about specific other agents are more successful in their tasks than agents that do not incorporate an emotional model.

Kim et al. [14] describe an autonomous robot, called Silbot, intended as a “mascot,” or companion for the elderly. Silbot is capable of expressing emotion in multiple modes, including word choice, facial expression, and posture, the latter category including arm position, neck position, and wheel movement. The basic library of gestures and expressions are developed by motion capture of professional actors, rather than generated from somatic principles. Terms that the robot intends to communicate are filtered through an emotional layer, which provides specific word-choice selections and settings for motor components that are intended to communicate the robot’s current emotional state.

4 Grouping the Architectures

We considered various groupings of the presented papers. Ordering by date would make it possible to demonstrate the tendency toward complexity of implementations that we saw when investigating the various architectures. Grouping by the type of implementation would make it easier for architects building different types of artificial emotion systems to see which papers were most central to their area of interest. Ordering by the most closely associated emotional theory would show which theories have been implemented by the various teams and which have been neglected by authors to date. We determined that it would be best to provide all these groupings to allow researchers to select the grouping or ordering that best fits their current needs.

5 Conclusion

Christian [5] describes his preparation for being a confederate at the Loebner Prize competition for the Most Human Computer. The confederates are the human beings that act as the decoys for the computers in this international competition held each year. The computer system that is able to gain the most “human” votes when compared with these confederates is awarded the “Most Human Computer” award for that year. The descriptions Christian provides of the historical test constraints as well as the current constraints of the test, and of the systems that have won make it clear that, at least in the case of the Loebner Prize, we no longer consider rationality to be the only thing that defines “intelligence” in our AI systems.

In fact, the humans gaining the most “computer” votes are generally those with the highest degree of rationality and crystallized knowledge. The computer systems that have won the prize in most recent years, however, either used recorded bits of human text conversations or used systems that behaved as if they were abusive humans or emotionally obsessive ones. This would indicate that the “humanness” sought by the Turing test has shifted in recent years from meaning “rational” to meaning “emotionally rational,” with an emphasis on the emotional.

Despite some authors’ interpretations of the Turing test as a test of rational behavior (e.g., Schroder and McKeown [22]), the most famous implementation of that test on computer systems is no longer a test of rational thought alone. While historically, there were constraints in place that gave the advantage to “rational” systems, current human judges for the Loebner Prize, focusing on the capabilities and limitations of current AI systems, are pushing the boundary of the definition “intelligence” more toward an expectation of fully human behaviors.

Users and designers are increasingly coming to see AI systems as social systems (e.g., Schröder and McKeown [22]). Understanding how previous emotion-based AI systems have been designed, and connecting these with the psychology literature, may provide a bridge that future developers can use to pass these evolving definitions of the Turing test.

The systems we have examined in this paper use various techniques, including supervised and unsupervised machine learning, data mining, fuzzy logic, Bayesian inference networks, neural networks, statistical analysis, natural language processing, and behavioral tagging. These are used to recognize, construct, or use human-like emotions to improve system performance in both interacting with humans and other agents and choosing effective plans or actions. Each of these conceptualizes emotions in their own way. We have attempted to focus our classification of these systems on the way past designers have viewed emotions and the purpose for which emotions are used. This may help future developers of robots and AI systems to decide how to conceptualize emotions and their roles for new systems.