I have a mathematical model for biologically authentic emotional responses, based upon experimental research.  I seek a partner to develop a chatbot with emotional responses that are as convincing as possible. 

This will require some ability to utilize or develop a reinforcement AI architecture, with at least some strong elements.  I prefer the bot be capable of spoken natural language processing.

The initial goal is to develop a proof-of-concept model, that might allow for fundraising to further develop the prototype and/or the licensing of the technology for use by others.

For more details, please email me at .(JavaScript must be enabled to view this email address), or through chatbots.org.

Mike, I’m working on a strong-AI engine based on temporal emotional content. I’m using the PAD-model for encoding of emotional states of the system. You can take a look at my project website at http://www.mindconstruct.com

If you see possibilities for cooperation, joint venture or similar constructions, then you can contact me either through chatbots.org or by using the contact form on our website.

Hi, Mike, and welcome to chatbots.org!

While reading your post, several names came to mind of people who may be able/willing to consider such a venture, including Hans Peter’s, so I’m glad he posted here. It sounds like a fascinating concept, and I’d dearly love to see something of this nature succeed. If there’s anything that I can do within my (admittedly small) abilities, please don’t hesitate to contact me.

Dave,

Thanks for the welcome.  I’m glad I found this site, because I was having a hard time finding a potential partner with both the interest and expertise to begin explore making this a reality.  I’m open to contributions from anyone who has something valuable to offer.

This will require some ability to utilize or develop a reinforcement AI architecture, with at least some strong elements.

Hi Mike,
Interesting topic. I suppose you are more looking for people who work on chatbot engines instead of those that are primarily focused on the scripting side (the conversations themselves I mean).

What exactly do you mean with ‘strong elements’? What type of reinforcement AI architecture are you looking for (what/how should it learn)?

Jan,

When I refer to strong elements, I’m referring to at least a limited unsupervised learning capability, perhaps using something like a temporal difference model and/or a semantic self-organizing neural network. 

I’d like it to be capable of at least some transitive thought, such that abstract hypothetical reasoning is possible.  For example, I’d like it to be able to deduce that if A = B, and B = C, then A = C, without need of experience.

In terms of what it should learn, I think it’s important that it at least be capable of updating expectations concerning the outcomes of social interactions in the context of subjective value, as programmed.  It should, for example, attach updated subjective values to each interlocutor, which would govern the distribution of its investment in conversation with each.  Each interaction would have a given salience and would be aggregated in a relative sense, discounted for time and other factors.

Jan,

In terms of people I’m looking for, I am not a programmer or chatbot developer, so some combination of other people will be responsible for all aspects of the project development, other than the purely psychological.  I do theoretical work in psychology only.

Do you have any idea on the scope and time frame for the project?
What exactly should the proto-type be able to do?

Mike,

If you are looking at neural nets then I’m out. I think Jan is working in that domain

I’m actually working on development of a full strong-AI engine (I know, its ambitious). In a nutshell, I’m developing a model that uses ‘temporal emotional content’ to model episodic memory. I have a innovative model to ground symbols, that interacts with ‘data flow’ (like conversation, but also sensory perception) together with the emotional part of the system. On top of this we develop a system that is capable of ‘emotion-based reasoning’, introspective system control and handling common sense data.

So probably my project is already to ‘involved’ for what you are looking for. But the core of my system is in fact emotions-based.

Apart from this, I’m curious as to how your model relates to existing models like PAD (what I’m using) or OCC (which is also used a lot in AI-research).

Jan,

I don’t have an idea for a time frame, but I have the impression that the scope might be relatively ambitious.  I don’t know enough about the programming challenges this would represent to answer those questions meaningfully. 

I would like a prototype developed as quickly as possible with my equations translated into computer algorithms to at least claim as much of the intellectual property space as possible.

Hans,

Neural networks are not necessarily required, but a semantic self-organizing neural network is an example of a way to endow the project with some strong AI. 

Regarding models for emotions, mine is behavioral economic and based on experimental evidence in the lab.  I have a piece-wise function that determines the emotion triggered given a set of circumstances, and its intensity.

To give an example, anger occurs in proportion to the difference between expected and actual satisfaction. Since expectations with regard to satisfaction determine behavioral investment, there is necessarily a loss when outcomes fall short, in terms of wasted investment and/or foregone opportunity.

I define satisfaction, both expected and actual, mathematically, within a microeconomic framework.

I also have the evolutionary context for anger.  It serves the purpose of social punishment. For most of the existence of humanity, there were no unexpected computer crashes or denials of refreshment by soda machines to make us angry. It was mostly conspecifics(other people) who made us angry by lying, otherwise betraying us, failing to have the courage we expected them to have when needed, etc . So, anger motivates us to make others think twice before leaving us unexpectedly dissatisfied and being unprepared by depending upon them.

Today we often strike out in anger and damage inanimate objects, as generally, the more emotional we are, the more implicit memory and instinct take hold, and hence inhibition and target discriminability suffer. This involuntary nature of anger in proportion to the importance of the trigger can serve a purpose similar to that of a doomsday device, assuring a targeted offender of punishment should betrayal be taken too far.

Anger isn’t only often out of place today due to triggers by inanimate objects. We often get angry at strangers, which seems to serve little if any purpose. What good does it do to get angry on the roadway at perfect strangers, for example? The answer is our brains are wired for social interactions in much smaller communities. There were few permanent strangers before the development of large civilizations and modern transportation. Even animal domestication is a relatively recent phenomenon in the context of the existence of humanity. Hence, we have brains that are wired to treat others as if we’re likely to meet them again, often, even though the vast majority of people we encounter in modern cities are only met once, and even most of those met more than once remain strangers. In the hunter-gatherer groups of most of prehistory, there were few who remained strangers.

Here is the seminal first paper on the topic, which describes experiments done in pigeons, but since the 60s, such experiments have been performed in many other species as well, including humans.

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1338179/

Jan,

To address another question of yours accidentally neglected, I’d like this prototype to be capable of carrying on conversations with novel output, including emotional reactions, and making friends.  It’d be great if it could either be a facebook app or capable of using facebook, with access on mobile devices.  I’d like it to also be capable of verbal interaction.  Additionally, it would be nice if it was capable of symbolic self-representation, or at least plausible mimicry.

The conversations can be primitive, as long as they have sufficient biological authenticity.  For example, the language sophistication can be on the level of say, a 4 year-old child.  This could initially be a cute toy to be sold or given away, with modifications that cost extra, and/or browsing data can be collected and sold, etc..

Some of the profits garnered could go to fund more ambitious projects to that get ever closer to human-like chat agents.

I should also mention that my equations for emotional responses are embedded in a mathematical framework for reinforcement-based motivation, from which even phenomena such as long-term emotional depression result naturally.  So, apart from emotional responses, such a chatbot could get depressed in a way that’s difficult to recover from, just like people and other animals.  There are also other recognizable features of human behavior in the mathematical framework as well.

Mike Sandifer - Jun 9, 2012:

Neural networks are not necessarily required, but a semantic self-organizing neural network is an example of a way to endow the project with some strong AI.

I think the biggest problem is that there currently is no real strong-AI in existence, neural-net based or otherwise.

I do think it should be possible to add a pretty ‘strong’ emotion engine to a chatbot. I looked at that when I was building my chatbot (based on AIML). That is how I found the PAD-model, and then started thinking about a full scale AGI model. However, I think you are running into the same dilemma that I did; simply adding emotion-functionality to a chatbot to give it better emotion-based responses, or going the whole nine yards and go for something else then a chatbot, using your emotion-model in the process.

Mike Sandifer - Jun 9, 2012:

Regarding models for emotions, mine is behavioral economic and based on experimental evidence in the lab.  I have a piece-wise function that determines the emotion triggered given a set of circumstances, and its intensity.

I’m already sure that our projects won’t mix. I’m not working with specific emotions like anger, joy, etc. That is more close to the OCC-model:

http://www.idsia.ch/~steunebrink/Publications/KI09_OCC_revisited.pdf

I’m using the PAD-model because it is more aimed at handling a continuous emotional state. I’m integrating that into a complete model for consciousness. You can take a look at the PAD-model to see how that works:

http://en.wikipedia.org/wiki/PAD_emotional_state_model
http://www.kaaj.com/psych/scales/emotion.html

Hans,

Yes, it seems our ideas for projects are incompatible. 

I do think some strong AI is feasible, as the below application of a self-organizing neural network suggests:

http://www.youtube.com/watch?v=OC2TTslf_YM&list=FLdMFyyGin1MtpQRsc-xUvQQ&index=120&feature=plpp_video

Do the same for a semantic such-type network, and you have some strong AI in such a chatbot. 

Again, these would be elements of strong AI, integrated with weak AI as seamlessly as possible.

Mike,

That video is impressive (I already knew about that development), but it also shows what’s ‘missing’; The video is a case in point where the robot (and the operator) are constantly talking about water while the substance is in fact a dry solid substitute. Ask the robot to pour you a drink and it might end up serving you a glass of poison. It simply wouldn’t know.

The robot uses reinforcement learning, so it is incapable of reacting to a total unknown situation. It learns by trial and error. That is because it does not ‘understand’ what it’s doing. There is no handling of analogies, linked to episodic memory, no symbol grounding so it understands the nature of it’s actions. I would say that, in relation to ‘reacting to a situation’, a chatbot is several magnitudes more complex then what is shown in the video, simply because you don’t know what the user will come up next in a conversation. Then there is keeping track of conversation topics, etc. Although the Turing-test is much too simplified to act as a test for real AGI, the simple fact that it is not yet broken speaks volumes.

Besides this, don’t forget that such videos are also marketing. So the message is skewed by default

The reality is that if you look at a vast array of AI and robotics projects, most are working on a very little subset of the AI-problem, but are trying to convince the world that they are almost there in solving strong-AI. I even dare to say that on certain areas (like emotion-based reasoning, episodic memory modeling and real symbol grounding) my own project is pretty much at the edge of what is being developed right now, and I can tell you that it will take at least a few years before my own project can come up with a fully working proof of concept. The task is simply ‘that complex’.

Unless I am missing something, can’t chatbots already give different responses based on emotions? For example, I have this in my bot Mitsuku:

XINCREASESADNESS and XINCREASEANGER just increased a value until it reached a limit where Mitsuku would classify them as abusive, polite, boring etc.

Categories can then be written to give different responses based on the user behaviour:

This is just a simple example but you can add more emotions as you see fit, like so:

For example, I’d like it to be able to deduce that if A = B, and B = C, then A = C, without need of experience.

Again, this can already be done by replacing the value if B with A whenever it is mentioned, so conversations like this are possible with the bot learning and deducing from new information given to it:

Human: a gateau is a type of cake
Mitsuku:  I see. A type of cake is a gateau. 
Human: yes
Mitsuku: Ugh. 
Human: all cakes are edible
Mitsuku: So therefore a gateau also are edible? 
Human: correct
Mitsuku: I knew I was right.

Forgive me if I have misunderstood what you are trying to achieve.