ZDNet Australia recently reported on a radical Artificial Intelligence system called EMIR (the Emotional Model for Intelligent Response). This new system will allow software and machines to mimic human emotions, creating devices such as "friendly fridges" or "angry robots". ZDNet speaks to the creators behind the project to find out more about the science and practicality of emotional machines.

The original story prompted a torrent of questions from our readers. "Why would you build such a machine?" "What would it be like?" "How can you make human emotions with mathematics?"

ZDNet delves deeper into the subject with two key minds behind the project: Charles Guerin of I & A Research and Dr Albert Mehrabian, of the University of California. We asked the questions from our readers and more, unravelling the mysterious 'Frankenstein' world of a machine that has feelings.

Our first Q&A is with Dr Albert Mehrabian, about his research into psychology which was used in the emotional simulator.

Our first question is regarding your research into emotions. How did you collect data? Did it involve zapping university students with electrodes?

No, it is not possible to obtain the kind of detailed information that we needed for understanding emotions without resorting to verbal reports. So our challenge was to devise highly sophisticated sets of protested verbal questions that would yield reliable and valid answers about emotions.

How did you measure the feelings of test subjects?

I began developing the test items back in 1972 and developed and refined these up until 1998. Basically, subjects were presented with contrasting pairs of emotion terms such that (and here is the key) each pair only differed with respect to (a) pleasure, (b) arousal, or (c) dominance. In this way, subjects were forced to report their emotions on a single dimension at one time. This helped greatly in obtaining detailed and precise answers.

People talk about the emotional components of intelligence. Do you consider emotion to be a critical component of intelligence?

Emotion is a critical component of all human and animal function. In fact, emotions are what adults share with infants and animals. Communication about one's emotions and interpretation of emotional signals of others are the very key to human existence. The literature on emotional intelligence...suggests that individuals who are better aware of their emotions and those who are better able to decipher the emotions of others are behaving more intelligently. One way to use my emotion measures to apply them in this arena is to compare a person's interpretation of another's emotional expression to the consensus reports (i.e., interpretations of the same emotion by an average group of peers, as reported using the PAD).

Why is your emotional model in three dimensions? What does three dimensional mean in this context?

There are literally hundreds--and if you choose to be specific enough, in excess of a couple of thousand--different emotion terms (or labels for human emotions). The challenge to a researcher was to reduce this enormous range of terms to a few that could be studied systematically and experimentally.

I drew my inspiration from the work of Charles Osgood who had developed the "semantic differential" technique to describe human reactions to a great variety of stimuli. Osgood has used the dimensions of "evaluation", "activity" and "potency" based on factor analytic findings. Our work with emotions showed that what Osgood had been finding was in fact the substrate (or lowest common denominator) of human reaction to any situation. And, this lowest common denominator is emotion, the very basic way in which humans deal with any situation and before they even begin to verbalise what they are experiencing.

Anyway, our studies showed that we needed at least three dimensions for a reasonably complete description of emotions--we could have used more, of course--however, adding more dimensions did not significantly improve our ability to describe specific emotions.

Three-dimensional means that we regard the three basic dimensions of emotions as constituting a three-dimensional space. Each emotion term, in turn, becomes a single point in this emotion space. Thus, once emotions are mapped into this space, we can quickly see which emotion terms are positioned close to one another in the space (e.g., fear, anxiety, distress) and which emotions are spaced considerably apart (e.g., elation vs. boredom).

What is an 'emotion term'?

An emotion term is a verbal label that we use to describe a specific emotion (e.g, tired, sleepy, empowered, vigorous). There are many ways in language to express the subtle nuances of emotion and each word used can be analysed in terms of its basic PAD components.

Why are there 259 of them?

There is no specific reason for this except that these seemed to be the most commonly used emotion terms.

Take the word "abandoned". This emotion term rated -59 (out of a range of -100 to +100) on pleasure-displeasure, it rated 41 on arousal, and rated -65 on dominance-submissiveness. Thus, "abandoned" is a substantially unpleasant state, it is a moderately high arousal state, and it is a very submissive state. Take the word "belligerent" with ratings of 45, 66, and 35. You can see that the major difference between "abandoned" and "belligerent" is that the former involved submissiveness, where the latter involves dominant. Both, however, are unpleasant and aroused emotional states."

Does the makeup of an individual's personality skew their emotional responses in decision making?

Absolutely yes. This is the other area of major research effort for me. I have used the three basic dimensions of emotion to also characterise people and their characteristic emotional reactions (or temperament). Thus, for instance, you can describe people as having "anxious", "bored", "exuberant", "relaxed", or "dependent" temperaments. The reactions of a relaxed person vs. a bored one to the same situation are apt to differ.

A relaxed person will typically respond with greater pleasantness and dominance, whereas a bored one with respond with more unpleasantness and more submissiveness. Thus, a relaxed person will appear to be more outgoing and decisive and, generally, will make a better leader.

As I have already noted, emotions constitute the lowest substrate of human experience that we share with animals. In my opinion, emotions come first developmentally and ontogenetically. Conceptual functioning is built on emotions and tends to be far more unstable and transient. When you place a person under a great deal of pressure (e.g., as in a marital conflict situation, in a high-tension stock market investment situation...you very quickly wipe away the "façade " of high cognitive development of sophistication and the observant psychologist will readily note the reflex-like emotion-triggered reactions. In the investment situation, for instance, fear may propel many a decision, irrespective of the high level of technical knowledge the person may possess.

Are tolerance and self-control components of emotional intelligence?

I think self-control, meaning ability to control one's behaviour that is automatically and unthinkingly triggered by emotions, is a key part of emotional intelligence. Tolerance is somewhat more removed from the subject matter. Yes, I believe you can teach self-control and modelling (that is having subjects observe a "model" exhibit self-control in various situations) can be one way to teach it.

Considering your understanding of emotions, does it give you a new insight about your own emotions? Has it affected the way you emote?

My studies of psychology and emotions have definitely helped me understand myself and my reactions and emotions in various life situations. When I take the time to carefully consider my options (for reacting) in a situation, I can of course benefit from this knowledge.

However, in most life situations, it is altogether too easy to simply go ahead and respond intuitively and in an unplanned manner. Otherwise, interactions with others tend to become overly stilted, as in a therapist-client situation.

What I am saying is that I often take the easy way out and live with the consequences. This makes life more interesting for me.

Can people benefit from education on the causes influencing their emotions?

Absolutely yes. I have devoted a tremendous amount of research effort to trying to pinpoint the various ways in which our emotions determine (in positive and negative ways) our behaviours.

Examples are my studies of effects of emotions on (a) hunger and food consumption (overeating or loss of appetite), (b) sexual desire and sexual dysfunction.

EMIR can have innumerable applications and this will depend on the ingenuity of the people who find a physical situation that they want to infuse with human-like characteristics. The key in any single situation will be to identify the Pleasure-displeasure, Arousal-nonarousal, and Dominance-submissiveness counterparts of emotions in the physical variables.

Thus, for instance, an airplane control system that registers very strong winds would be programmed to respond with high arousal and low dominance (submissiveness); a computer program that is extremely complicated and difficult to teach or use would be translated into unpleasant, arousing, and submissiveness-inducing emotional components.

You could, for example, apply EMIR in political campaign situations. Each candidate (or even each political slogan or advertisement) can be quantified in terms of its PAD values; my findings that relate human preferences of different combinations of PAD experiences would in turn allow campaign managers to predict the effectiveness or success of various strategies.

How complex is the emotional makeup of a given individual?

As I have noted above, each person can be described in terms of his/her characteristic level of pleasure-displeasure, Arousability (or general level of emotionality), and dominance-submissiveness. This would be in sharp contrast to how a poet or a novelist might describe a person; however, in contrast to the latter, it would give us a very powerful set of tools for comparing and studying differences in individual reactions to the same situation.

There is a tremendous amount left to learn about our brains and emotions. I think ultimately we will get to the point whereby we will be able to tie the PAD dimensions to variations in brain chemistry. And, we will probably be able to relate temperament differences in PAD to genetic markers.

Once the relations of brain chemistry to emotions and the relations of genes to temperament are known, you can imagine the consequences, both good and bad.

Dr Albert Mehrabian is the Consulting Editor of the Journal of Psychology, published in North America.

The Technology of Emotions

How and why did you decide to develop the EMIR?

EMIR was originally developed out of an odd circumstance. I was part of a development team that had just delivered a control system that monitored temperature in large stadium-sized rooms for millions of dollars worth of fruit. One day, we got a call from the owner who started yelling that our system had frozen thousands of dollars of grapes overnight. He was having a fit, and was extremely angry. When I got there, I noticed that the system had reported bizarre temperatures that were obviously screwy, but the midnight staff upon viewing the bad readings hadn't responded to them. The numbers and graphs were all obviously wrong...but they hadn't elicited any reaction from the human operators.

The bug turned out to be simple--a bad ground on a sensor--but I kept thinking about the operators who ignored the readouts, thinking, 'If they had only felt the full depth of the owner's emotion, they would have certainly responded quicker'. It dawned on me that perhaps the problem with automated systems was that they didn't really seem to "care" about whatever it was they were controlling. We simply reduced all readouts to numbers or cryptic messages, instead of the simple emotions that humans are trained from birth to respond to.

I looked around to try to find an existing system for simulating human emotion, and found that none existed. After around six months of researching many different models of emotion, I developed a theory of analysis that when it met up with the work of Dr. Mehrabian, formed into a very elegant solution.

People talk about the emotional components of intelligence. Do you consider emotion to be a critical component of intelligence?

Intelligence relies on emotion. Without emotion, you can't have intelligence...all you can have is 'data', data that is difficult to communicate and difficult to understand. I think emotion helps us derive intelligence from information in three important aspects.

1. Emotion adds a lot of subtle information content to any situation.
   
2. Emotion helps us to remember information.
   
3. Emotion-laden information is understood faster than data.

If you look at a bird and think 'That is a bird, and it's blue, and it's making a sound'...you are being a robot, analysing and developing 'data'.

If you look at the same bird and think 'what a beautiful bird, of such rich blue colour and such a fabulous voice' and feel some tingling pleasure inside...that is being an intelligent human and you are developing 'information' which is of much deeper meaning to another human than the 'bird was blue'.

Another important aspect of emotion in intelligence is memory. Memories tend to be strongly linked with emotion states, and the stronger the emotion state, the stronger and longer lasting the memory. Think September 11th or the day that Kennedy was shot, and everyone can tell you where they were and what they did. Ask them what they ate for breakfast yesterday, and there was likely little emotional impact, thereby no memory.

In terms of speed, consider your mother mentioning 'Your room requires cleaning' versus 'Clean your room, now, or else'... and you'll grasp how the latter example will elicit a more rapid response, though the data was there in the first instance.

Can you tell us about your favourite application of EMIR, or a use for it that you would like to see fulfilled?

There are many applications I'd like to see. In particular, I'd like to see applications that help people with emotional disorders understand and deal with their emotions. Software could be built to help children focus and expand their breadth of emotion and feelings.

The obvious commercial uses of EMIR lie in computer games and environmental simulators, but I believe that we may see the day when emotion is just another API, another tool for communication that is part of every operating system.

Did the EMIR system produce any unexpected results during development?

EMIR is always surprising, often hilarious, and sometimes uncanny. Robby, the emotional thermostat, was the first 'real-situation' test of our theory, the first few hours running live were very special. We initially had messed up the accuracy of some of the internal mathematics and were getting sort of random results that were fun, but not accurate. Our programmers were rolling on the floor with some of the light scripts we'd given Robby to say to enunciate the various emotions. 'I'm relaxed, thinking of my vacation in Lake Tahoe'.

However, as soon as we fine-tuned it, the feeling was 'wow', Robby started to respond with the same descriptive terms a human would use, in the right situation, he was getting nervous at the right times, scared, and going through an emotional regime that was perfectly natural... and that's when the hairs on the back of my neck went up and I realised we were onto something.

The Robby demo seemed to 'care'.

An interesting aspect of this convincing demo was that all emotions were deduced from only two input pieces of data--the temperature and whether a cooling motor was on or off--and some fancy math.

Could you give us an example of a practical difference between the decision making of an AI system with and without emotions?

Sure, let's say that you have a verbot (a talking robot) that is programmed to learn from talking to you about your interests and knows that you enjoy flying airplanes. You start to tell the verbot that you intend to fly today. The verbot checks out an online weather database and notices light rain followed by thundershowers in the area where you want to fly. In a classical decision making system, the verbot might decide that IF 'Thundershowers' THEN Warn-User.

Programmed with EMIR the system would feel "concerned and worried" about the light rain, and ask you if you would change your flight plan? If you said no, the system might get more aggressive, even angry about not being able to help you. In short, the difference is that emotion systems make it a lot easier to simulate 'caring' and manage subtle information situations, very well.

Is is possible to design a certain personality for the EMIR system?

Yes, definitely...at least that aspect of 'personality' that is affected by emotional response and tendencies.

Could the unique personality of a person be copied using the system?

Well, not really someone's 'personality'. The emotion response to situations though can be mimicked very effectively though using the mathematics of EMIR.

What can you tell us about emotional voice inflection, where a computer generated voice can be altered to communicate particular emotional feelings. Have you seen this in action?

Yes, actually I've seen three different voice generators that do a good job of synthesising emotion into voice. Our technology is an excellent input device for taking 'real world' situations and using them to control that type of technology.

Do you think emotional voice inflections could be universally applied to computer-generated speech without pre-recording every word and sentence?

Definitely. At first I expect that you'll see simple systems that are "cute and interesting", and then advanced systems that add a lot of emotion inflection to speech systems.

Can you imagine a future where this exists?

I have this one nightmare that someday my refrigerator and my washing machine are going to get into a heated debate over whether it's 12:01 am or 12:02 am and scream at each other all night.

Charles Guerin and I & A Research developed the EMIR system, based, in part, on the theoretical work of Dr Albert Mehrabian.

Copyright © 2009 CBS Interactive, a CBS Company. All Rights Reserved.
ZDNET is a registered service mark of CBS Interactive. ZDNET Logo is a service mark of CBS Interactive.