sign in sign up
<<
Home , Affective computing, Cognitive science, Hope, Rosalind Picard, Trust (social science)

AI Magazine Volume 20 Number 1 (1999) (© AAAI) Book Review

machines. Although Picard warns Review of Affective about ethical issues, she apparently welcomes the use of detec- tors in a wide range of contexts and Computing relationships (for example, teacher and pupil). The final chapter, in par- ticular, suggests that computing devices will help us choose our mood ; decide which scenes to record for our photo albums; and find out about exhibitions, plays, and other events likely to suit our . Reac- tions to this prospect will differ wide- riting a multidisciplinary implants; so, it will be possible to have ly. Many will dislike the idea of using book is a risky business. a wide range of sensors, processors, remote devices to tell them which WSome are likely to and transmitters constantly monitor- mood to expect in their spouse: It will be fiercely critical because of omis- ing, analyzing, recording, and trans- strike them as an improper intrusion. sions or errors. Others with tunnel mitting about one’s However, there probably are some vision are likely to miss the point. blood pressure, temperature, blood- couples who, having found bedroom , with considerable sugar level, muscular tension, and ceiling mirrors tame, will relish mutu- , addresses a broad collection many other physiological states. Some ally informative intimate sensors. of themes, including the nature of of these devices, suitably hidden, If I were a pilot or a bus driver I , , and ; could also monitor various aspects of might accept the right of my passen- the detection of emotional and other the environment, including other gers to insist on my being wired up to affective states and processes; the people. Thus, even your friends and minimize the of disasters if I fall nature of and the rela- colleagues will easily be able to record asleep at the controls or have a heart tionships between intelligence and your conversation; your facial expres- attack while in control. However, I emotions; the of the sions; and, perhaps with remote sen- would not want a linked to and other aspects of human physiolo- sors, your muscular tension, tempera- such sensors to select music for me, gy relevant to affective states; require- ture, sweating, and so on. Picard and I have no wish to use such ments for effective human-computer that such machines can learn devices, especially hidden devices, to interfaces in a wide range of situa- to predict our reactions and use such tell me what my wife, my colleagues, tions; wearable devices with a range of predictions to inform us of and or my students are feeling or to sensing and communication func- inform them of what I am feeling. I’d tions; philosophical and ethical issues prefer us all to become more sensitive. relating to of the future; Others might have a different view. and a brief encounter with theology. , All this discussion begs the deeper This is a book with a bold vision. Rosalind Picard, question, Can computational devices Some readers will find it inspiring and really be used to detect emotions, as stretching. Some will find it irri- The MIT Press, Cam- claimed? Yes, a subset, emotions that tating. Some will have both reactions. bridge, Massachusetts, produce characteristic patterns in It gives many pointers to the vast lit- measurable physiological states, can erature on emotions, including useful 1997, 292 pp., detect emotion. How such sentic mod- recent material, for example, books by ISBN 0-262-16170-2 ulation (defined on p. 25) might be LeDoux, Goleman, and Damasio. produced and detected is discussed at The book ranges over themes of length in chapters 5 and 6. I argue varying depth. The main theme con- below that this will not work for most cerns the nature of intelligence and opportunities (“You’d really like that socially important emotions. the role of emotions in intelligence, film ...”). She argues that such devices which I discuss later after comment- will increasingly be able to measure Communication ing on some of the simpler themes. and categorize emotional and other affective states to help us comprehend and what is going on in ourselves and oth- The book makes an important claim and Sensing ers. (The New Scientist [1998] reports that is relevant to AI research on com- on a Japanese device that purports to munication and teaching, namely, It will increasingly be feasible to tell you what a pet or infant is trying that the ability to detect and influence install sensors and computing devices to communicate!) affective states in others is important in furniture, walls, car seats, driving Some people might be alarmed by in human communication and will be controls, clothing, jewelry, and even the prospect of being “spied on” by necessary for machines to interact

Copyright © 1999, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1999 / $2.00 SPRING 1999 127 Book Review effectively with humans. I agree with technically defined terms and use matic evaluations to be performed. this statement and made similar these, which is best done using archi- Possible effects include (1) reactions in claims in Sloman (1992). However, I tecture-based : Starting from the primary emotion , includ- am not sure the methods proposed by an architecture, we can derive the ing physiological changes, for exam- Picard will be very effective. types of state and process that the ple, muscular tension, weeping, flush- Can emotions be recognized on the architecture can support (Sloman ing, and smiling, which can produce a basis of physiological changes? Chap- 1998, 1992; Wright, Sloman, and Beau- characteristic “feel,” for example, “a ter 6 describes affect recognition “as a doin 1996; Simon 1979). Animals or flush of ” and “grow- pattern recognition problem and affect machines with different architectures ing tension” (try imagining a surgical expression as pattern synthesis” (p. will not be describable in these terms. operation on your eyeball), and (2) 165). Such patterns can be measured by On this basis, we can distinguish pri- rapid involuntary redirection of devices such as physiological sensors mary and secondary emotions. processes (compare Sloman and cameras recording facial expres- and Croucher [1981] and Simon sion or posture. “Despite its immense Primary Emotions [1979]). It is not always appreciated difficulty, is easier Human have many compo- that effects of type 2 can occur with- than thought recognition.” It is easier nents that are evolutionarily old. out effects of type 1. “largely because there are not as many Some are responsible for “animal” emotions as ” (p. 167). emotions, for example, being startled, Two Types of Picard’s discussion may give some frozen with terror, sexually aroused, Secondary Emotion readers the impression that all emo- or nauseated. Information from per- Damasio conjectures that triggering tions essentially involve measurable ceptual fed to a fast pattern- by thought contents depends on physiological responses, that is, sentic recognition mechanism can rapidly somatic markers that link patterns of modulation. However, that impres- trigger massive global changes. Such thought contents with previously sion would be accurate for only a sub- mechanisms apparently include the experienced or or oth- of emotions, as explained later. brain stem and the limbic system er strong . Such triggering Moreover, many emotions involve (Goleman 1996; LeDoux 1996). Engi- enables secondary emotions to play an thought contents, and they are no less neers will appreciate the need for fast- important role by directing and redi- diverse than thoughts. You can, in acting pattern-based global “alarm” recting in dealing with com- principle, be afraid of, for, be mechanisms to ensure that an agent plex decisions (Damasio 1994). Picard pleased about almost anything that reacts appropriately to important also believes that secondary emotions you can believe or expect. Most of the risks and opportunities (Sloman always trigger primary mechanisms, emotions people are interested in are 1998). producing sentic modulation. Howev- constituted largely by their contents, Damasio (1994) calls these primary er, I think we need a distinction and the contents of emotions are also emotions (pp. 131–134), as does Picard. between two subclasses: (1) central the contents of thoughts. Even if there These products of our evolutionary secondary emotions and (2) are a few physiologically detectable are still often useful. Because secondary emotions. universal categories covering all emo- they involve physiological reactions Central secondary emotions involve tions (happy, sad, angry, and so on), relevant to attacking, fleeing, freezing, involuntary redirection of ongoing which I , it won’t be much use and so on, sensors measuring physio- cognitive processes such as planning, knowing that your spouse is happy logical changes (including posture and reasoning, reminiscing, and -mon- without knowing whether it is happi- ) can detect such pri- itoring. Such shifts of attention can ness about your promotion or about a mary emotions. occur entirely at the cognitive level message from a secret lover. Measuring without involving sentic modulation. sentic modulation will not, in general, Secondary Emotions An example might be , which answer such crucial questions about Primary emotions can be less impor- involves negative assessment of one’s human emotions. tant for civilized social animals than own motives, decisions, or thoughts certain semantically rich affective and can produce thoughts about Ambiguity in the states generated by cognitive processes whether detection will occur, whether involving appraisal of perceived, or to confess, what the likely punish- of Emotion imagined, situations. These are re- ment is, how to atone, how to avoid It has been clear for a long time ferred to by Damasio as secondary emo- detection, and so on. Other emotions (Arnold 1968) that the word emotion tions and described by Picard on pp. (, , and so on) will has no unique and clear meaning. A 35–36 and 63–64. They can arise only have different effects on attention. proliferation of definitions can be in an architecture with mechanisms Peripheral secondary emotions occur found in philosophical and psycholog- for processes such as envisaging, when cognitive processes trigger states ical literature. It might, therefore, be recalling, planning, and reasoning. such as primary emotions without any wise to avoid discussing whether com- Patterns in such processes can trigger to redirect thought pro- puters can have or detect emotions. learned or innate associations in the cesses (for example, the shudder pro- Alternatively, we can introduce new “alarm” system that cause rapid auto- duced by imagining scraping one’s fin-

128 AI MAGAZINE Book Review gernails on a blackboard). Subtle patterns expressing , , A hybrid secondary emotion could involve a mixture of both types, for , or infatuation can be clearly visible to example, guilt or embarrassment others long before the subject notices the accompanied by sensed bodily changes. emotional state. In Sloman (1992), I Central secondary emotions are conjectured that some involuntary often important to novelists, play- wrights, poets, and garden fence gos- emotion expressions are necessary in a sips. There need not be any overt society of intelligent agents because expression, but when there is, it will a society where all expression was typically be some sort of verbal utter- ance or intentional action. I don’t voluntary would be too unstable: mean that people label their emo- Nobody would know whom to . tions: Like other animals and young children, even human adults can lack the sophistication to recognize and classify their own mental states. Rather, a central secondary emotion can be expressed involuntarily in Can Secondary Emotions Be tions and sentic modulation, depend- choice of words or as an extended Detected by Physical Sensors? ing on the person and the type of thought or behavior pattern such as emotion. The can also make a If Damasio’s and Picard’s theories of frequently returning to a theme or difference (for example, the British secondary emotion were correct, then always expressing disapproval of a cer- “stiff upper lip”). Subject to the need secondary emotions would always tain person. to calibrate patterns for individuals, is generate sentic modulation, which Subtle patterns expressing anger, the first hypothesis true? could be measured externally and jealousy, pride, or infatuation can be I suspect the answer depends on the used to detect emotions using pattern- clearly visible to others long before person. Even if secondary emotions in recognition techniques. Two hypothe- the subject notices the emotional most people sometimes trigger the pri- ses need to be considered: state. In Sloman (1992), I conjectured mary emotion mechanisms, it does that some involuntary emotion Hypothesis 1: The general not follow that they always do, for all expressions are necessary in a society nature of a secondary emotion people, or that they need to in artifi- of intelligent agents because a society can be detected (for example, cial emotional agents. Thus, as an where all expression was voluntary anger, jealousy, , apprehen- unqualified generalization, I suspect would be too unstable: Nobody sion, consternation) but not the even hypothesis 1 is false. would know whom to trust. In semantic content (for example, It is specially problematic for mixed humans, emotional reactions become who you are angry with and what emotions, such as those reported by less physical, more central, and more you are angry about). the captain of a woman’s yachting controlled as part of the process of Hypothesis 2: Both the general team at the end of an around-the-world growing up and becoming emotional- class of emotion (for example, race: She was looking forward intensely ly mature (Goleman 1996). Damasio’s anger) and the semantic content to seeing friends and relatives whom architecture could easily be modified (for example, anger at Fred for she had not seen for many weeks and to explain these changes; for exam- giving away a secret) can be to eating again (because food had run ple, the global alarm system might be detected from such patterns. out two or three days earlier); she was able to control cognitive mechanisms The first hypothesis is no more delighted at completing the race suc- directly without causing sentic modu- plausible than the hypothesis that our cessfully but disappointed at not win- lation, and there might be a second thought contents can be inferred from ning it, sad at the thought that the global alarm system whose output are externally observable physiological adventure was over and the team directed only at the deliberative (cog- patterns. For example, primary emo- would have to separate but looking for- nitive) mechanisms. tion mechanisms that evolved long ward to the next such event. Could any With either alternative, some sec- before political systems cannot be collection of grimaces, shudders, pos- ondary emotions could cause cogni- expected to produce a distinctive ture changes, tears, sweating, and so tive interrupts, redirection of atten- physiological pattern for consterna- on, have expressed this mixture? tion, and new evaluations, without tion over an election defeat or pride at Words are more than apt to the task. necessarily triggering primary emo- being elected president. tion symptoms. Of course, the older Is the first hypothesis, the weaker How Are Mixed primary emotion mechanism might hypothesis, true? Picard acknowledges Emotions Possible? sometimes be triggered concurrently, that there will be considerable varia- Picard offers blending and rapid alter- producing hybrid emotions. tion in the mapping between emo- nation as possible models of such

SPRING 1999 129 Book Review mixed emotions (p. 171). Neither is or ) in internet news organs by nerve cell terminals, plausible because coexisting emotions groups concerned largely with mutual under the control of a dedicated (jealousy and guilt at feeling jealous) consolation and support. brain system, which is respond- endure and preserve their identity. A The writer of the following message ing to the content of thoughts more accurate model would be a collec- clearly feels guilty, apologetic, and a relative to a particular entity or tion of coexisting , possibly little apprehensive without saying so: event. (Damasio 1994, p. 139, implemented as concurrent (mostly I know I promised to give you a my emphasis) unconscious) cognitive processes striv- chance to check our paper, but Emphasizing the contents of ing for attention and control. Of my portion was not finished until thoughts is correct. Emphasizing body course, a combination of infatuation just before the deadline and I had state ignores central secondary emo- with X and jealousy of X’s lover could to send it in. I hope it won’t cause tions involving only involuntary be regarded as a blend but only insofar you any embarrassment. attention control mechanisms. Dama- as one is the cause of the other. The sio shows later that he is fully aware Compare Marc Antony’s soliloquy infatuation could also coexist with that there can be such chains of causa- when first left alone with the corpse of unrelated jealousy over a colleague’s tion at the cognitive level: “…con- Caesar: promotion. There’s no more need for sumed as we are by using the past to these to blend or alternate than for dif- O, pardon me, thou bleeding plan what-comes-next, a moment ferent coexisting beliefs or expectations piece of earth, away or in the distant future. That all- to blend or alternate or all the many That I am meek and gentle with consuming, ceaseless process of cre- attitudes that we have to different these butchers! ation is what reasoning and deciding things, people, policies, places, and so Thou art the ruins of the noblest are about…” (p. 165). on. A highly parallel information-pro- man These are processes in a mental “vir- cessing architecture makes blending tual machine.” We can modify his the- and alternating unnecessary. That ever lived in the tide of ory by allowing that some secondary times. emotions involve only such virtual Linguistic Expression Woe to the hand that shed this machine processes, without any sentic Picard mentions the power of written costly blood! modulation (although, obviously, communication regarding emotions … there are physiological processes in the (pp. 13 and 97), but she underesti- brain). He discusses a general require- Cry ‘Havoc!’, and let slip the dogs mates its importance: ment for such mental processes later: of war, If computers are to utilize the nat- The personal and immediate That this foul deed shall smell ural channels of emotional com- social domain is the one closest above the earth, munication used by people, then to our destiny and the one that when computers learn to recog- With carrion men, groaning for involves the greatest nize human emotions, they will burial. and complexity. Broadly speak- have to rely primarily on sentic These words are far more powerful ing, within this domain, deciding modulation, as opposed to hav- than explicit telling. When human well is selecting a response that ing people explicitly tell them emotions are expressed, we don’t usu- will be ultimately advantageous the names of their emotional ally “tell them the names.” In fact, to the organism.… Deciding well feelings. (p. 26, my emphasis). very often, the emotions have no also means deciding expeditious- Contrary to Picard, I suspect that for names because they are complex and ly, especially when time is of the the emotions that matter most to sometimes even unique combinations essence, and, in the very least, humans, the primary and most natural involving interactions between deciding in a time frame deemed mode of expression is linguistic. As semantically rich coexisting motiva- appropriate for the problem at indicated previously, I don’t mean that tional states, attitudes, evaluations, hand (Damasio, p. 169). people recognize and label their own expectations, beliefs, and so on. In Damasio’s remarks draw attention emotions. Rather, as novelists and play- addition, even when they have names, to what might be called metalevel con- wrights know well, our choice of lan- having the emotion does not require trol, that is, combining the ability to guage can convey rich information knowing the name or being able to solve problems and make plans with about mental states by making our recognize the occurrence in oneself. the ability to notice that such a pro- thought processes “readable” external- cess needs to be interrupted or redi- ly. Such choice of allows rich- rected, for example, because time is er and more diverse affective states to Toward a Modified Version too short to continue planning: be expressed than either a set of physi- of Damasio’s Theory Action must start. Damasio assumed ological patterns or a set of verbal that only emotions can perform this labels. Part of the evidence is the long Damasio wrote: kind of high-level control. However, history of intensely emotional letter I see the essence of emotion as the recent work in AI has produced alter- writing and poetry. There are also the collection of changes in body native metalevel control mechanisms, profound outpourings (for example, of state that are induced in myriad including planners that use anytime

130 AI MAGAZINE Book Review (Boddy and Dean 1989), which allows an interrupted planning process to produce useful initial partial If emotions are not required for plans in intelligent machines. Some plan-based theorem provers metalevel control, is affect necessarily use high-level strategic to relevant to intelligence? guide the search for a proof. Similarly, Luc Beaudoin’s (1994) Ph.D. thesis discussed the notion of a metaman- agement layer in the architecture of an agent with reactive and delibera- tive mechanisms. This layer can mon- itor, evaluate, and, to some extent, control and redirect attention in understanding of the structure of motivational and emotional states. deliberative mechanisms. There is human . Chapter 7 of the book However, must intelligent systems much work still to be done exploring discusses various recent attempts to also have emotions? I think they such mechanisms. model cognitive affective processes, almost certainly will but not for the We therefore do not need to follow which are totally different from the given by Picard and Damasio. Damasio and Picard in assuming this sentic modulation capabilities that are I have argued (Sloman 1987; Sloman rapid redirection of attention by a the main focus of chapters 5 and 6. and Croucher 1981) that certain kinds global alarm system necessarily works This topic is important for future AI of emotion will be side-effects of by invoking the primary emotion research on self-aware and socially mechanisms designed to overcome mechanism. Cognitive control mecha- aware agents. My feeling is that look- resource limits in intelligent systems. nisms can directly redirect attention ing back in years to come, we’ll find They argue that emotional mecha- to new goals, new items of informa- that all this work is very shallow and nisms are required for intelligence. tion, and so on (Sloman 1997; Simon inadequate, especially models involv- ing explicitly labeled emotional states Must a Skilled Communicator 1979). In some cases, this redirection Have Emotions? goes on even though such redirection and special emotion-generating rules. is of little or no , for example, in It might be possible for a robot with- obsessive jealousy or anger. Damasio Emotions and Intelligence out any emotions of its own to learn a comes close to accepting my sort of great deal about human emotions, theory (pp. 197–198). As Picard points out, most AI including how to recognize Building on Beaudoin’s work, my researchers ignore motivational and them on the basis of both physical colleagues and I have argued that emotional mechanisms. There are a states and, more generally, on the humans have at least three important few exceptions, notably Simon’s pio- basis of understanding what the pupil coexisting interacting control systems: neering 1967 (Simon 1979) paper or client is saying. (1) an old reactive layer; (2) a newer responding to criticisms of AI by the It is often suggested that we deliberative layer; and (3) a still more Neisser. Randall Davis about mental processes of others by recent metamanagement layer con- (1996), in his American Association simulating them in our own brain. If cerned with monitoring, evaluating, for Artificial Intelligence 1996 confer- this were the only way to predict emo- and redirecting internal processes. ence presidential address, views AI as tional reactions, it would be impossible Control by the third layer is generally exploration of the space of designs for for a completely unemotional agent. only partial because there are auto- intelligent systems, including both However, we can also use general matic mechanisms that can redirect natural and artificial systems. He dis- knowledge to reason about reactions attention, including one or more reac- cussed at length the diversity of capa- of other people. Nevertheless, it is like- tive global alarm mechanisms, as dis- bilities and mechanisms that have ly that one source of information cussed earlier. We have suggested that evolved naturally, but there was no about emotions is one’s own emotion- different sorts of emotion correspond mention of motivation and emotion. al experiences; so, perhaps a Spock to processes in these three layers. If emotions are not required for without emotions might find it Picard summarizes these ideas about metalevel control, is affect necessarily difficult to deal adequately with metamanagement and perturbance on relevant to intelligence? A partial humans, although not impossible. pages 211–213. However, this theory is answer is one of the Picard’s themes in its early stages of development. mentioned earlier, with which I agree: Are Emotions Required for There is still no adequate account of If intelligent machines are to commu- General Intelligence? and . nicate effectively with humans in a Picard, like Damasio, makes a much If computers are to be able to detect, wide range of situations for a wide stronger claim, namely, that the abili- respond to, or model these types of range of purposes (teaching, advising, ty to have emotions is required not affective state, they will need rich and and so on), then they will often need only for communicating about emo- subtle linguistic abilities and a deep to take account of actual and likely tions but more generally for control-

SPRING 1999 131 Book Review ling reasoning, as indicated in these said about expertise in more abstract mechanisms may be capable of gener- quotations: domains, such as , algebra, pro- ating central secondary emotions. We all know that too much gramming, and games such as chess. In a culture, the experience of many How to Interpret Damasio’s emotion can wreak havoc on rea- Evidence soning, but now there is evidence generations can be transmitted in a that too little emotion also can compressed time scale to new mem- Damasio, Picard, and others have mis- wreak havoc (p. 10). bers. Likewise, it is now commonplace interpreted the evidence about brain to allow chess programs to make use of damage in Damasio’s book as imply- Apparently, a balance is a great deal of “book learning” to ing that emotions are essential to needed—not too much emotion, avoid massive combinatorial searches. intelligence, which is a simple non and not too little emotion (p. 11). Although the full workings of sequitur. Damasio’s findings point to an human associative learning are nei- Certain sorts of frontal lobe damage essential role of emotion in ratio- ther well understood nor replicated in produce two effects: (1) patients lose nal thinking (p. 12). current AI systems, it is clear that they the ability to have certain kinds of Nevertheless when a system need not involve emotions (although (secondary) emotional reaction or to problems where the possi- in some cases, they do, sometimes care about things that previously mat- bilities cannot be enumerated with bad results). For example, a tered to them (including physical pain and evaluated in the available young child picks up a huge vocabu- in some cases) and (2) the patients time, I suggest that affective deci- lary and many subtle grammatical become less creative and decisive and sion making provides a good rules simply by being exposed to less able to take strategic decisions. solution. Humans use feelings to speakers of the language. This can render them totally unable to help them navigate the oceans of Another thing we learn is informa- manage their own lives, even though inquiry, to make decisions in the tion about control. Experts learn to they retain normal functions of per- of combinatorial complexity detect patterns in a situation that sug- ception, , language, motor (p. 222). gest that a different approach is need- control; perform well on all standard I this view of the role of emo- ed, the problem is insoluble, or some intelligence tests; and even have tions is an overgeneralization from more important and urgent problem explicit knowledge about how they how a subset of humans work. There has arisen. Thus, besides hierarchically ought to behave in various circum- are two problems with these claims: structured goal-directed processes, an stances. (1) misconstrual of the expertise animal or a robot with multiple inde- It is fallacious to infer from this evi- involved in avoiding search and (2) pendent sources of motivation inhab- dence that effect 1 is the cause of effect misinterpretation of the evidence iting a dynamically changing and 2 because there can be mechanisms from brain damage. partly unpredictable environment for controlling and redirecting atten- There are many areas of expertise needs mechanisms that can redirect tion in the cognitive virtual machine, that potentially involve massive attention away from the current goal which are essential for intelligence searches but where humans somehow and its subgoals. (Compare chapter 6 and also produce secondary emotions. manage to avoid the searches. Anyone of Sloman [1978].) When they are damaged, there is a loss who has worked on natural language These mechanisms need not be of secondary emotions as well as bal- processing or image understanding emotional, although sometimes they anced judgment and control of will know that both utterances and will be, namely, when the intrusions thought processes. It does not follow retinal images have huge amounts of involve highly positively or negatively that emotions are necessary for intelli- local ambiguity regarding segmenta- charged evaluations and cannot be gence. Rather, mechanisms required tion; grouping; selection among possi- prevented by metamanagement deci- for intelligence sometimes produce ble meanings; and, in the case of sions of the agent. I have called these emotions. Such emotions are emergent. images, possible occlusions. Typically, perturbant states. (A perturbant state Compare disconnecting the car bat- the problem of resolving the local can become dormant when attention tery will prevent the radio working ambiguity subject to global con- is grabbed by something more power- with the car starting. It doesn’t follow straints can involve enormous search ful. Dormant emotions don’t go away; that the radio is required for the car to spaces. However, humans seem to they simply await their chance to start. home in rapidly on a unique interpre- regain control, for example, when Of course, Damasio’s evidence does tation without searching, except in grief is temporarily forgotten because support the hypothesis that in addi- the case of garden path sentences (for an urgent and important task grabs tion to the mechanisms studied in par- example, “the horse raced past the one’s attention.) ticular subfields of AI and cognitive barn fell”) and puzzle pictures. In summary, the -control , an The obvious explanation is that powers that Picard and Damasio requires more global control mecha- expertise is based on a very large col- attribute to emotions can occur with- nisms that attempt to ensure that lection of slightly generalized special out emotional mechanisms, although these mechanisms are deployed ap- cases stored in some kind of content- as argued in Simon (1979) and Sloman propriately. This is what metamanage- addressable memory. The same can be and Croucher (1981), such control ment is about. This need is obvious to

132 AI MAGAZINE Book Review any software engineer accustomed to liography pointers to much relevant Sloman, A. 1998. Damasio, Descartes, designing systems with multiple capa- literature, although I suspect the best Alarms, and Metamanagement. In Proceed- bilities, for example, operating sys- literature on this topic has yet to be ings of the International Conference on tems or plant control systems. written, perhaps by readers stimulated Systems, Man, and , 2652–2657. Washington, D.C.: Institute of Damasio’s patients reveal that the and challenged by this book. Such Electrical and Electronics Engineers. high-level control mechanism may be work requires a broad multidisci- Sloman, A. 1997. What Sort of Control Sys- incapacitated, while many more spe- plinary background. Unfortunately, tem Is Able to Have a Personality. In Creat- cific aspects of intelligence remain there are still too few researchers like ing Personalities for Synthetic Actors: Towards intact, which should not be very sur- Picard who are willing to combine Autonomous Personality Agents, eds. R. Trap- prising to software engineers. psychology, ethology, , pl and P. Petta, 166–208. Lecture Notes in , computer , software AI. Berlin: Springer. engineering, AI, and philosophical Sloman, A. 1994. Explorations in Design Conclusion insight in the context of creative engi- Space. In Proceedings of the Eleventh Euro- This wide-ranging, ambitious book neering design. pean Conference on AI, ed. A. G. Cohn, presents work from several disciplines, Maybe one day, their numbers will 578–582. Chichester, U.K.: Wiley. Reprint- ed in Boden, M. A. 1990. The of including empirical results, theoretical reach a critical mass, they will discover Artificial Intelligence, 231–247. Oxford Read- a common analysis, and practical applications ings in Philosophy Series. Oxford, U.K.: along with some of the ethical issues within which to communicate, and Oxford University Press. they raise. the subject will really take off. A regu- Sloman, A. 1992. Prolegomena to a Theory The majority of the discussion of lar section in AI Magazine on affective of Communication and Affect. In Commu- emotions is based on the widely held computing might help to accelerate nication from an Artificial Intelligence Perspec- assumption that they always involve this process. tive: Theoretical and Applied Issues, eds. J. externally detectable sentic modula- Ortony, J. Slack, and O. Stock, 229–260. tion, as primary emotions do. This Acknowledgments Heidelberg, Germany: Springer. assumption, however, ignores the pos- I gratefully acknowledge ideas gained Sloman, A. 1987. Motives, Mechanisms, sibility of central secondary emotions, from, or sharpened by, colleagues in and Emotions. and Emotion 1(3): which I have claimed are the most the Cognition and Affect Group at the 217–234. important emotions in our (adult) University of Birmingham and also Sloman, A. 1978. The Computer Revolution lives and certainly of most in useful comments and suggestions on in Philosophy: Philosophy, Science, and Models of Mind. Hassocks, U.K.: Harvester. much of our thinking about one an earlier draft by Adele Howe, Fanya Sloman, A., and Croucher, M. 1981. Why another. From this viewpoint, the Montalvo, and Rosalind Picard. Robots Will Have Emotions. In Proceedings emphasis on externally detectable pat- References of the Seventh International Joint Confer- terns of physiological processes is ence on , 197–202. unfortunate. However, as an account Arnold, M. B., ed. 1968. The Nature of Emo- Menlo Park, Calif.: International Joint Con- of how primary emotions and some tion. Harmondsworth, U.K.: Penguin. ferences on Artificial Intelligence. peripheral secondary emotions are Beaudoin, L. P. 1994. Goal Processing in Wright, I. P.; Sloman, A.; and Beaudoin, L. expressed and how they might be Autonomous Agents. Ph.D. thesis, School P. 1996. Towards a Design-Based Analysis of detected, it may be a good beginning. of , The University of Emotional Episodes. Philosophy, Psychiatry, The book includes many topics I Birmingham. and Psychology 3(2): 101–126. have not had space to discuss, includ- Boddy, M., and Dean, T. 1989. Solving ing several challenging and potential- Time-Dependent Planning Problems. In Proceedings of the Eleventh International ly extremely interesting and probably Joint Conference on Artificial Intelligence. very difficult applications of affective Menlo Park, Calif.: International Joint Con- Aaron Sloman (B.Sc., mathematics and computing, for example, automating ferences on Artificial Intelligence. physics, Cape Town, 1956, D.Phil., philos- ophy, Oxford, 1962) is professor of artificial the process of searching a for a Damasio, A. R. 1994. Descartes’ Error, Emo- intelligence and at The picture or a piece of music with a spe- tion Reason, and the Human Brain. New York: University of Birmingham, where he has cific type of mood for use in an adver- Grosset/Putnam. tisement or as background for a film. been since 1991, having previously helped Davis, R. 1996. What Are Intelligence? and to found the School of Cognitive and Com- Although I have been critical of Why? AI Magazine 19(1): 91–110. puting at Sussex University. His some major themes, there is much of Goleman, D. 1996. : research interests include philosophical interest and value, and I believe the Why It Can Matter More Than IQ. London: foundations of computing and AI, book has no competitors. As a wide- Bloomsbury. architectures for intelligent agents, the evo- ranging and provocative groundbreak- LeDoux, J. E. 1996. The Emotional Brain. lution of , architectural sup- er, it can be recommended to students New York: Simon and Schuster. port for motivation and emotions, varieties of representation, mathematical thinking, of AI who need to have their minds Simon, H. A. 1979. Motivational and Emo- vision, and the design of tools for exploring stretched. However, they should be tional Controls of Cognition. In Models of architectures. His web page is available at warned not to believe everything they Thought, 29–38. New Haven, Conn.: Yale www.cs.bham.ac.uk/~axs/. read! University Press. Readers will find in the text and bib- New Scientist. 1998. New Scientist 2135:11.

SPRING 1999 133