I was asked the following question by Deb Roy. > Finally, you said that you "make sharp distinctions between desires, > attitudes, moods and emotions". Can you send me a pointer to your most > relevant papers on this point? My answer follows. This is a very hasty first draft. To be revised and extended. 8 Apr 2002 ======================================================================= For several years I have been talking about how our ontology of affective states and processes is design-based or, more recently, architecture-based, and I have been giving fragmentary examples (e.g. emotions, goals, and in some recent slides evaluations), but I don't think I've ever written it down in a systematic way. Matthias and I have a long incomplete disorganised draft joint paper begun when he was working here for a year, that addresses some of it. We keep meaning to get back to it. But it is a huge topic, which keeps growing as we look at it. I can't remember whether Marvin Minsky has addressed this sort of thing in a systematic way, either in Society of Mind or the Emotion Machine (I read lots of things that I absorb without remembering that or where I have read them). I have some not very good diagrams related to varieties of affective states, scattered around my slides e.g. this old one http://www.cs.bham.ac.uk/~axs/fig/control.eps http://www.cs.bham.ac.uk/~axs/fig/control.jpg (I've just created the jpeg file in case the other causes problems.) It (somewhat obscurely) makes the point that affective states are all control states within a complex system, but that they vary in a number of "dimensions", e.g. - how long term/short term they are (many intermediate cases) - whether they have semantic content or are just control states (we'd also need to distinguish different kinds of semantic contents, e.g. propositional and other kinds - not indicated in the diagram, and different kinds of control) - whether they hard or easy to change - to what extent they are innate or produced by learning or development (including being influenced by a culture) - whether they are global or specialised (different notions of "global": affecting the whole organism vs affecting only part (like an eye-blink); or affecting all decisions vs affecting only a subset of decisions. etc. etc.) - whether they are concerned with controlling actions or with controlling control states (e.g. preferences are concerned with selections between (arbitration between) goals or desires). Those "dimensions" are not necessarily orthogonal. NB: Within the framework of an architecture one can talk about varieties of control states and processes and different collections will be supported in different architectures. E.g. not all architectures will support all the above dimensions. With that background, here's an incomplete, sketchy set of notes forming a first draft answer to your your question. (I must write all this up properly some time -- though perhaps you will instead!) 1. A desire (or want, not in the sense of "lack") is a state involving a goal, which has some semantic content which may be true or false (or more or less nearly true, etc.) The goal is to make it true, or make it false, or keep it true, or keep it false, or move towards making it true, or to keep it true as long as some condition holds, etc. etc. (there are many variations in goal contents, presupposing different varieties of representational sophistication, not all of them propositional, e.g. in most animals they can't be). Besides the goal component a desire has other components referring to its role in the architecture and its control functions. E.g. having a desire may or may not involve having a commitment to acting on it (e.g. it may clash with some other desire as a result of which it is rejected, though it remains as a desire). I.e. the desire need not be the content of any *intention*. (Fortunately). Also desires vary in a number of important ways, e.g. in the time periods for which they exist, whether they are dormant or not, how intense they are, whether they are recognized to exist, whether fulfilling them gives pleasure or not, how they grow and subside, how semantically focused they are (wanting something to eat, wanting an orange), whether they are intrinsic or derivative (sub-goals) etc. There are many things that can make a desire go away besides satiation. I don't think we have anything like a complete theory of the variety of life-histories of desires, though good novelists have implicit theories which are good though partial. (H.A.Simon had things to say about this in his 1967 paper.) You can have a desire, even a fairly strong desire (e.g. to get a PhD) without being at all emotional about it. However desires often form parts of dispositions to become emotional, e.g. if the desire is about to be thwarted or an unexpected opportunity for fulfilling it turns up, etc. But no emotion need follow from thwarting of a mild unimportant desire (e.g. a desire to scratch a mild itch in a part of your body you cannot reach). Some desires are idle wishes: they generate no action at all, though events that fulfil them will be welcomed and evaluated positively. (Evaluations are different from desires, though connected. There's a lot more to be said about positive and negative evaluations, pleasures, pains, etc. and how they function within the whole architecture.) 2. An attitude may be either long term (mostly dispositional, and dormant) or short term (episodic). The long term ones are what we most often refer to using the word "attitude", but the word has other uses. A long term attitude towards something is (usually, but not always) a cluster of beliefs, desires, preferences, evaluations, and intentions concerning some object or class of objects. Mostly these are dormant dispositions, doing nothing but capable of being triggered in some way. What they trigger need not be external behaviours: they more directly trigger mental processes. Examples include attitudes towards a country, a composer, a type of music, a person, oneself, one's job, a programming language, a way of life, a religion, a field of study, etc. Human adults in our sort of culture have thousands of attitudes, almost all of them dormant at any time. (There may be very wide cultural variations in the ontology of attitudes. E.g. think of "family honour" in some cultures.) A full analysis would have to include the variety of ways in which attitudes can become active, e.g. by perceiving the object of the attitude or something related to it (e.g. a threat to it), or hearing news about it, or being asked a question about it, or discovering that something you are about to do will impact on it, etc. Most people give "love" and "hate" as examples of emotions, because they think only about the intense passionate varieties. However, most varieties of love and hate are attitudes (or parts of attitudes). Someone who really hates football need not be in an emotional state, except occasionally when forced to watch football, or when learning about payments to football players, etc. You can love something or hate it even though most of the time you don't even think about it. Mild attitudes may be capable of generating decisions and actions, but not emotions. Other attitudes can trigger emotions when activated. A more complete theory would break this up into quite a lot of specialised sub-cases. The excellent Ortony, Core and Collins book which claims to be about emotions is mostly about attitudes as analysed here. I don't know of any comprehensive theory on the aetiology of attitudes. There is another use of the word "attitude" which is much closer to "mood", though it usually implies something longer lasting and possibly more controllable (e.g. "he has a very positive attitude today"). Sometimes "attitude" is used to refer to an emotional state, though rarely. I think this is connected with its links to "posture". Sometimes it refers to a personality trait, though what that means is another long story: personalities operate differently in the different layers of the architecture.) 3. Moods are (usually) temporary, relatively short term -- e.g. hours or days) global control states that affect a variety of mental processes, including possibly perception, reasoning, generation of desires, decisions, plans, and actions. Examples include optimistic moods, pessimistic moods, happy or sad moods, being excited, sleepy, lazy, apathetic, depressed, irritable, placid, contented, and many more. They may be more or less intense, more or less easy to change, more or less dominant. (If you understand the evolutionary benefits of such global control states, it all makes good sense.) Unlike desires and attitudes, moods need not have any specific objects nor any semantic content: someone who is in an irritable mood need not be irritated by anything specific, but has a disposition to be irritated by almost anything that happens. I recall that this paper by Laura Sizer (Hampshire College) has some useful things to say about differences between moods and emotions: Laura Sizer (2000) Towards a computational theory of mood, in, British Journal for the Philosophy of Science, 51, 4, December, 2000, pp. 743--769, (We have exchanged some email, but I don't know her.) Insofar as moods are global control states without semantic content and without rapid changes, they are the sorts of things you'd expect chemical mechanisms to be able to turn on and off or modulate. There may be other kinds of faster-acting global signalling mechanisms for speeding things up, slowing things down, reducing or increasing the degree of risk-taking, etc. So perhaps we need to distinguish different sorts of moods on that basis. They probably shade into emotions as we currently understand them. Sometimes "mood" in English refers to a desire or inclination: "I am in the mood for a drink, ... dancing, reminiscing, etc." I think that's a straightforward ambiguity, or to be more precise "mood for" labels a different concept. (Can you be in a "mood against" dancing?) 4. Emotions are also control states, though they are usually (though not always) transient and shorter-lasting than attitudes. (Think of the etymology: Being "moved" by something.) I think the notion of an emotion is linked to a particular type of distinction between *normal* and *abnormal* (e.g. emergency) processing which I find hard to define precisely with complete generality. It's easier to explain "abnormal" in the context of an architecture where "abnormal" processing is the responsibility of special sub-mechanisms for dealing with emergencies and urgent or over-riding needs (the "alarm" sub-systems in my diagrams), but hard to give it a general definition. Maybe there isn't one: just a lot of special cases (it's a cluster concept that needs to be refined by specifying different varieties of abnormal processing, which may lead to useful refinements of our notion of "emotion" among other things). So: Emotions involve this contrast between "normal" processing (perceiving, deciding, learning, approving, disapproving, thinking, planning, etc) and something relatively "abnormal" that disrupts, interrupts, or modulates, normal processing. (Oatley and Johnson-Laird had an over-simplified version of this when they went through a phase of defining emotions as related to things that disrupt plan-execution, if I remember correctly.) This disruption usually involves a positive or negative evaluation of something, the object of the emotion. But it makes a difference where and how that evaluation is done (e.g. in a reactive sub-system, in a deliberative mechanism, in a self-monitoring system, etc.) -- and also what sort of thing is being evaluated: a thing an event, a possibility, something happening now and here, something happening far away, the probability of something having happened, etc. (Ira Roseman, once produced a partial taxonomy, though it was not architecture-based. The Ortony Clore and Collins book adds many useful additional points.) Different sorts of emotions are linked to different processing levels in a multi-layer architecture, e.g. primary, secondary and tertiary emotions (a first draft crude, inadequate, subdivision). So agents/organisms with simpler architectures cannot have the full range of emotions. An enduring emotional state is sometimes called a mood though it may be more accurate to say that there may be both an emotional state (with semantic content, etc.) and a mood (without, though caused by the emotion). Long-lasting emotions can be temporarily dormant while something else holds attention, e.g. grief, jealousy, infatuation. Individuals may be, but need not be, aware of their emotions. Discussed at length in connection with grief and attachment in I.P. Wright, A. Sloman and L.P. Beaudoin, (1996) Towards a Design-Based Analysis of Emotional Episodes, Philosophy Psychiatry and Psychology, 3, 2, pp. 101--126, Repr. in R.L.Chrisley (Ed.), Artificial Intelligence: Critical Concepts in Cognitive Science, Vol IV, Routledge, London, 2000, And online here http://www.cs.bham.ac.uk/research/cogaff/0-INDEX96-99.html Emotions may have a rich internal semantic structure including beliefs (about the object of the emotion - e.g. being angry with X typically includes believing that X did or wanted to do, or intended to do something you evaluate negatively). They may have very different time scales, different degrees and kinds of intensity, different side-effects (e.g. generating second order emotions, such as guilt at feeling jealous) very different types of semantic and ontological sophistication (e.g. being frightened of something big moving rapidly towards you vs being frightened that world war 3 is approaching fast). They can have a very complex aetiology, which various psychologists (and psychiatrists?) have tried to study, but without an adequate architectural framework for doing so. ======================================================================= Well that's a really *tiny* start at answering your question. But our vocabulary of mental states, even just our affective states, goes far beyond these few categories and I doubt that anyone has an adequate overview as yet. (Maybe it's in Cyc??? Or something else I have not read.) Till we have an adequate overiew we won't even know *what* we need to put into machines that can understand stories and conversations..... let alone *how* to do it. I've just remembered that in 1971 I was visiting Denmark and talking about emotions to philosophers. One of them said that Danish had relatively few words referring to different sorts of mental states, compared with English. So we opened a Danish dictionary and started searching through it. He was very surprised to find out how wrong he was: there was something relevant to mental phenomena on almost every page, and even more comes out in phrases. I.e. dictionaries contain a huge collection of useful clues about the nature of mind. This is elaborated in Chapter 4 (on how to do conceptual analysis) of The Computer Revolution in Philosophy: http://www.cs.bham.ac.uk/research/cogaff/crp/chap4.html Comments, criticisms, suggestions welcome. I look forward to talking to your robot about its feelings. Cheers. Aaron