Yes, it should be feasible. I suggest a steady activation

function like the sigmoid function. Start with all neurons

connected (n * n * 2 connection). Later you can prune

the connections with a modified fitness function.

Emotional states in such a type of neural net

are not different from any other state: it's

a combination of activations of the neurons.

In real life there would be hormones.

You may want to read out and display a certain

component of emotion (e.g. hunger or pain). For this

purpose, this component should be represented as

the output of a single neuron, rather than a spread

representation. So you should train the network

to use a chosen neuron for representing this

certain component of emotional state.

This could be done using an evolutional approach.

One training pair would consist of an activity

of inputs that leads to the desired emotion (or

the opposite), and an activation state for the

neuron(s) considered. One training step would

mean activating the input neuron(s), propagating

the causation forward and then let the emotional

output neuron(s) influence the fitness function,

comparing the output with the desired one.

Spread of activation could be: choose one neuron

randomly and calculate its activation and output.

Do this a couple of times, e.g. 5*n where n is the

number of neurons.

Regards,

Joachim

Nice, but do you have any such program to sustain whatever you said??

You can use

http://www.augos.com/temp/Neuron.cs

as a template. I made experiments with

pattern recognition, but I did not call the states

of the network emotions. Further research is

up to you.

Regards,

Joachim

Recent research in psychology and cognitive neuroscience are increasingly showing how emotion plays a crucial role in cognitive processes.Gradually, this knowledge is being used in Artificial Intelligence and Artificial Life areas in simulation and cognitive processes modeling. However, it lacks a theoretical framework that allows them to deal with emotion. In addiction, regarding emotion-based computational projects, controversial questions concerning the nature, function, and mechanisms of emotions, that must be considered, are mostly neglected on researches.

I had done research and experiments but there is less research till now. We have to do further more deep research to develop emotions

My paper "On modeling the affective effect on learning”, Fifth Multidisciplinary international workshop on Artificial Intelligence, December 07-09, 2011. Proceedings, Springer, Lecture Notes in Artificial Intelligence Volume 7080; studies the effect of moods and emotions during learning. This paper uses symbolic machine learning techniques.

Now, classifying the emotions by weight values in an ANN, becomes a representational issue! (A research topic)

Yes it is possible, In our research we created artificial evoked potentials for simulation purposes, same techniques with necessary domain modifications, will give you the desired AE.

Will it be useful add to include emotions to an artificial intelligence? Do we want computers that can get angry or sad?

Of course it will be useful since it will help to provide better understanding of emotion in a psychological point of view. Furthermore, it would be a big leap in AI as the main concept is to develop an Artificial being.

First we have to develop a complete theoretical model for emotion than we can move forward, the method suggested in first comment is debatable, because neuronal network is later stage first we need to understand the phenomenon of Emotions they are more related to heart than mind in early stage... that's what I think not necessarily correct ... but would love to understand more

Dear Dr. Bilal, I agree with you. As of today, we are missing a comprehensive theory that would be able to explain human emotions in their entirety, articulation and interaction. Furthermore, we do not know enough about the interplay of emotions with perception, volition, cognition and motivational states yet. We can mimic certain affective states or arousal mechanisms of human beings, even transfer these to artificial systems, but it sounds too ambitious for me to think of or talk about the implementation of emotional complexity – in the lack of sufficient insights and proper descriptive/predictive theories. Research still has a lot to do in this direction … We also have to consider that emotions, as complex psychophysiological experiences, differ person by person and that they also have time aspects (episodes or dispositions).

An ocean is a blending of water droplets. But you will never be able to understand the ocean by studying water droplets.

I am total in agreement with Dr Imre.

In fact this may be true for most of the research undertaken..

Could it be that the emotions are there to solve a problem rather than be a symptom? Suppose that the test for a convergent search process was a positve emotion and for a divergent process a negative emotion? This would mean that the underlying goal resolution mechanism (thought) used in planning included its own test conditions. Or would that be cheapening them (emotions) too much? And is that last question somewhat emotive?

John,

what a good idea. Only one sentence should be formulated a little different. I think, there exists two independent data processing mechanism. The older one is quick and unconscious: emotions. The other one is new and slow but conscious: thinking.

This would comply with the test condition which has to be independent and should use different ways to solve the problem. So we are able to test any situation without danger but reliable results.

Hi Wifried,

from you...

" I think, there exists two independent data processing etc"

There are two independent mechanisms. The look-up mechanism that simply responds rapidly to inputs in an appropriate way. This is an autonomic type of systems that is using what might be called "knowledge" . The other systems is goal-based mechanism capable of internalised feedback. This refines (to yield "knowledge") what might be called belief or even: solves open-loop control problems by modelling. We call this thought . It is that process that needs to converge and the emotions system could be seen as part of the system's test for convergence. The emotions system must be a component of evolutionary development so it somehow keeps us viable. Thus when we are not able to respond directly using knowledge we are still able to plan/model and derive viable responses - belief refinement using internal feelings to plan responses that are viable... perhaps emotions are generic tests for what is best for us..?

The brain has no obvious data storage and I would suggest that memory is more akin to state-based recall - like a delay line. the states can be triggered and that creates apparent memory. Thus I do not see the brain as a data processing system but instead as a vast state-machine and the feelings (emotions) can be interpreted as states - highly generic expressions of viability that can be triggered whenever we model/plan/think.

(The emotions system and the goal system are strongly related and both can be interpreted as generic actions - that also fits in with the concept of a state machine.)

John,

I am fascinated. Do you have some more information about this subject?

Heman,

I do not know so well in artificial neuronal networks but we know emotions are very old in biological organisms. The sense of emotions is to set a fast signal to react in also fast manner. Monst primary emotions are warnings: fear, disgust, surprise, may be also sadness. The message is clear and the nesessary reaction is also simple: turn round and hurry, avoid, attention, gather and orientate yourselves anew. If an organism is not able to react in adequate manner, it will not survive. This should stay stable in the evolutionary process so we find similar reactions in totaly different species over a very long time in biological evolution and there is no learning process in ontology. I am sure emotions are hard wired.

Very well said WIlfried....I feel i can develop an AI system apt to sense emotion. However, lack of research equipment and funding have refrained my from going full steam on my research......:)

I am just curious if a software approach is the best one to simulate the effect of emotions on neural networks (e.g. the EMANN model). Do you know about hardware implementations of emotions on artificial neural networks?

Conference Paper EMANN - a model of emotions in an artificial neural network

Up to now I have never heard of practical implementation but I have an idea of the approach that needs to be undertaken. however, due to the lack of appropriate tools in my lab, I have been restricted solely to simulations.

Actually I don´ t know any hardware implementations, except the experiment on robots done in our lab (until now unpublished). Our aim is to generate an swarm of autonomous robots, that is able to:

To solve this, we implement  "emotion" (=neuron-chemical modulation of specialized areas) in to the  to give evolution the ability to "reuse" a given area for different tasks. For example: the neural area evolved for "wall-avoidance" shold be active in any task of the robot, but the neural areas responsible for "high level behaviour" (e.g., where to go, what to do) should be influenced by emotions like "hunger" or "fear".

Due to the fact that in biology the neural and the neuron-modulatory hormonal system are extremely interwoven in very interesting feedback loops (and not as it might be expected on the first look in a hierarchical system) we decided not to clearly separate the two systems from each other, but to develop an architecture, where the concept of the continuous change from neurons to neuron-modualtory cells to neuronal controlled hormoneglands to "stand-alone" hormoneglands" is reflected.

Emotions are part of our "viability mechanisms". They serve a purpose or should that be define purpose? We do not implement emotions for some arbitrary reason, they should become obvious as a mechanism as our real-time control-based system seeks to improve its ability to survive. For example as a convergence or avoidance mechanism in arbitrary encounters.

As in Ronald's model wall avoidance (see previous post) is a fundamental of basic (closed loop) control  whereas obtaining resources or staying alive requires modelling - higher level of behaviour (solving open loop problems). A process of modelling in a neural system must be a feedback (recursive) mechanism perhaps even a resonating condition ...emotions?... drive?