From: Chris Paget (email@example.com)
Date: Mon Aug 22 2005 - 19:00:47 MDT
I mentioned in my JOIN that I'm working on a model of intelligence using
emotion as the driving force. When I posted that message to SIAI, I got
some questions back about it, so I'll post my reply here as well. I was
writing up another email clarifying some things and explaining some more
of the complexity involved, but for the sake of netiquette I'll just
send this out for the moment and save the extra stuff for when I
(hopefully) get some questions back.
The fundamental driving force behind emotional intelligence (as I see
it) is a desire to be happy. "Happiness" in the case of AI is just a
number, as are all emotions. In a biological intelligence there is some
correlation for emotions as simple vectors - it might, for example, be
reasonable to measure the fear level of a creature based upon the amount
of adrenaline in its system.
The catch is, the system cannot make itself happy. All of the input it
is given affects its emotional state - if it observes people who are
happy, it becomes happy itself. In essence, it is driven by a desire to
make other people happy - or, to put that another way, it is driven by
Alongside basic happiness sit a number of other basic emotions - fear,
anger, etc etc. Again, each of these cannot be influenced directly by
the being itself (at least not by conscious thought), they are primarily
controlled by external influences.
Above these primal emotions sit more complex emotions, such as
confidence, stubbornness, and optimism. (There is a threshold at which
these should be regarded as personality traits, although that's not
strictly relevant). Each of these more complex emotions has a smaller
influence on overall happiness, but at the same time are influenced less
by external factors. As a general rule, every emotion is controlled by
two things - external influences, and other emotions. Any emotion which
has a large effect on overall happiness is controlled more readily by
Example: Optimism is strongly affected by external influences. Optimism
controls how likely you are to take a chance - if the chance pays off,
your optimism goes up, and you feel happier because of it. At the same
time, stubbornness affects optimism, but does not directly affect
happiness much. If you are stubborn in a given situation, you are less
likely to take a chance, and the happiness increase when that chance
succeeds is far less.
Memory is a combination of two things. Firstly, a word - the concept in
question. Secondly, attached to each word is a set of emotional vectors
which comprise the beings total experience about that object or concept.
If the being encounters an object that it recognises, it consults its
memory to see how that object has made it feel in the past, determines
whether the emotions it presents are appropriate for its current mental
state, and either promotes or avoids the encounter accordingly.
Example: Let's say that in the past, you have been bitten by a dog.
Pain is a universal way to reduce happiness (much like pleasure is a
universal way to increase it), so the emotional vectors associated with
your memory of a dog include negative happiness. However, if you see a
person holding a dog and you remember that the person has, in the past,
made you very happy, you may still decide that the dog is not worth
avoiding; the negative happiness based upon your experiences with dogs
are offset by the positive happiness of your experiences with the
person, and you can make an intelligent decision on how to respond to
Automated learning, in any given situation, is simply the product of
combining a number of different emotional memories together in order to
achieve the required goal. You program the system with a number of
basic operations that could be applied to the task, and let the system
experiment randomly, learning based upon emotion along the way.
If, for example, correctly recognising a face stimulates "pleasure"
(either by seeing that the face is smiling, or by the programmer
pressing the "pleasure" button), then whatever operations were used to
perform that recognition (adjust brightness, decrease color depth,
adjust contrast, etc etc) are then given higher happiness ratings, and
are more likely to be used again. The act of randomly combining
operations together is based upon emotion, and the success or failure of
each attempt is similarly stored as emotion. If, for example, the
computer takes a chance on a new graphics operation when attempting the
recognition, it will remember whether it tried it before and failed
based upon the stored value for confidence; if its confidence is high at
that time then it may still take a chance on it.
There's a lot more complexity than what I've presented here, but I think
this should convey the gist of what I'm thinking. Hopefully that's
enough to let people start exploring the idea themselves, and hopefully
I can answer some of your questions.
-- Chris Paget firstname.lastname@example.org
This archive was generated by hypermail 2.1.5 : Wed May 22 2013 - 04:01:06 MDT