# How could emotional intelligence be implemented?

I've seen emotional intelligence defined as the capacity to be aware of, control, and express one's emotions, and to handle interpersonal relationships judiciously and empathetically.

1. What are some strategies for artificial intelligence to begin to tackle this problem and develop emotional intelligence for computers?

2. Are there examples where this is already happening to a degree today?

3. Wouldn't a computer that passes a Turing test necessarily express emotional intelligence or it would be seen as an obvious computer?

Perhaps that is why early programs that pass the test represented young people, who presumably have lower emotional intelligence.

Architectures for recognizing and generating emotion are typically somewhat complex and don't generally have short descriptions, so it's probably better to reference the literature rather than give a misleading soundbite:

Some of the early work in Affective Computing' was done by Rosamund Picard. There is a research group at MIT specializing in this area.

Some of the more developed architectural ideas are due to Marvin Minsky. A pre-publication draft of his book, The Emotion Machine' is available via Wikipedia.

Emotional intelligence would certainly seem to be a necessary component of passing the Turing test - indeed, in the original Turing test essay in Computing Machinery and Intelligence implied some degree of "Theory of Mind" about Mr. Pickwick's preferences:

"Yet Christmas is a Winter’s day, and I do not think Mr. Pickwick would mind the comparison."

As for your comment about a computer program showing lower emotional intelligence, you may find Eliza (which you can try here) interesting. It is classical in the history of AI and pretends to mimic an analyst (psychology).

However, I think your question fits nowadays more in the field of Human-Robot Interaction, which relies largely on vision for recognition of gestures and follow movements, as well as soft, natural movements as a response. Note that the movements of the face and hands belong to the most complex tasks, involving many muscles at a time.

I strongly recommend the film Plug&Pray to have an idea of what people are researching in this area.

On the purely human end of the scale, I sometimes wonder about our (my) emotional intelligence myself. Would I want to implement such an intelligence in an artificial agent at all?

I remember why I thought of Eliza: not because of its emotional intelligence, but because it was apparently taken seriously by a couple of humans. Could this be taken as a sort of (approved) Turing test? What does it say about the humans it met?

• Eliza is not really a good model of human emotion: it was chosen precisely because the `professional detachment' of the analyst means that it is somewhat credible to act dispassionately regardless of the user's statements. – NietzscheanAI Aug 3 '16 at 7:03
• @user217281728 You got a point there, definitely. – Luis Aug 3 '16 at 9:22

Emotions aren't something that you can implement - they're very complex. However, you can attempt to mimic them. Human emotions are closely related to conscious experience characterized by intense mental activity, which is based on interpretation of events.

Recent brain studies (including research in cognitive psychology and neurophysiology) suggests that human emotional assessment of every action or event plays an important role in human mental processes.

The recent 2016 Annual Meeting of the BICA Society brought together scientists from around the world to approach principles and mechanisms of human thought to create biologically inspired AI.

For example, in Samsonovich's (a professor in the Cybernetics Department at the MEPhI) proposal, the idea is to test AI in computer games which involves actions with emotional content, where AI may engage with players in different types of social relationships (such as trust, subordination or leadership).

Jonathan Gratch of the ICT, invented virtual characters capable of identifying and expressing emotions by communicating with humans in their natural language based on the situations where for example AI can deceive a human to achieve the desired result. The effect is obviously not achieved by re-creating human consciousness, but by achieving statistically adjusting parameters.

Researchers from the Institute of Cyber Intelligence Systems in MEPhI are hoping to be able to create in the near future virtual beings which are capable of planning, setting goals and establishing social relationships with humans, also by possessing both emotional and narrative intelligence which can interpret the context of events.

Source: Researcher proposes social emotions test for artificial intelligence

• Emotions aren't something that you can implement [Citation needed]. This suggests that emotions are somehow on a different level than what Turing machines can run, and there is no element that shows that it is true (excluding religious arguments, which are irrelevant). – Fatalize Aug 19 '16 at 12:31

So you may be familiar with Word2Vec, (W2V) which as Wikipedia describes1 "captures the linguistic contexts of words" using vector arithmetic. For example, subtract 'Paris' from 'France' and add 'Italy' and you get 'Rome'.

What you need is something like a Sentiment2Vec (S2V) that captures the similarities between emotional transitions. Something like: subtract 'fear' from 'sadness', add 'joy' and you get 'hope'. Or: subtract 'sting' from 'papercut', add 'smashed' and you get 'throbbing'.

The catch is that you don't have an easily accessible corpus of emotional contexts to train with, like you have with words. If you had a million hours of fMRI - mapping the transitions between emotions in hundreds of subjects - then you could use that data to build an S2V. You probably don't have that data though.

In the mean time, you could just build a W2V that specializes in sentiment. You could even try to use a current sentiment analysis engine to bootstrap it. Perhaps if you read enough text that says "I got a papercut and it stings" and "I smashed my finger and it's throbbing" then you could eventually produce an S2V. Children's books often use explicit language regarding emotional context ("this made the boy feel sad").

But words are still a far cry from the experiential context that a connectome map would provide. To test whether you have something useful or not, you might want to implement your S2V in a mouse foraging simulation - see whether it produces typical behavior and if any cooperative or competitive dynamics can organically grow out of your S2V.

Some further info on the subject:

In 2014, Glasgow University claimed2 that there are four primary emotions: happiness, sadness, fear and anger.

This website3 provides nice (if somewhat short) hierarchical breakdown of secondary and tertiary emotions under primary emotions.

References

1: en.wikipedia.org/wiki/Word2vec

2: www.bbc.com/news/uk-scotland-glasgow-west-26019586

3: changingminds.org/explanations/emotions/basic%20emotions.htm

• You can always add links... – user58 Aug 25 '16 at 18:55

The way i do emotion in a AGI system are by a bunch of little parts, agents, voting in system statis state registers. If the union of agents are working together correctly. This is the subconscious part.

The conscious part that plan out movement in the environment include these system statis state registers in all planned movements.

All emotions can be derived from these registers: