[Freddy]

I always thought chatbots are a lot like children. The attentive programmer is always tweaking and adding new knowledge, or finding a better way to do things.

[Don Patrick]

Many of my friends have also likened my self-learning A.I. to a child. It makes sense as both are young intelligences in development, trying to come to grips with the world. On my part, I have to do a lot of parenting in correcting wrong things that it learned. Human language is a giant pitfall of misunderstandings waiting to happen, no matter how intelligent one is.

[ivan.moony]

and what about statistical analysis of input?

For example, algorithm could do the following:
1. wait for input A
2. when an input happens respond a random function B1 with input A as a parameter
3. wait again for for a response
4. if response is affirmative remember that input A should yield B1 (A). If not, try answering B2 (A) and wait for a response and so on

Here we have try-error-success loop and answers with the most  success points should be used first with certain input.

But, to have totally open algo (no hard coded rules), the algo should learn by itself what responses are affirmative and what are negative. Just to complicate things a little more, answers can have a fine grain on a scale from affirmative to negative.

I'd hate to use red and green button on every input (reinforced learning), I'd like to have affirmation level learned by algo from experience, but how? How to learn that "Yes" means yes and "No" means no? I assume that there should be some behavior pattern analysis to conclude which words mean which affirmation level, but I'm not familiar with those patterns.

[Don Patrick]

Trial and error would allow it to learn even exceptions, yes. Combining that with statistics or pattern recognition would increase the accuracy. Given any question or problem, the AI would then first try the most common answer.

I think that combination has potential, but the description also seems to fit Cleverbot, which gives me reason to think.
If you are looking for a magical formula, I don't know it. Would it be so wrong to give the AI any hardcoded knowledge, even the most basic system for good/bad, yes/no, reward/punishment? There are only a few dozen words for affirmation. What are your objections to hardcoding them?

The impression that I am getting is that you are trying to solve a mathematical equation, like a x b = c, but are not willing to set an initial value for any one of them. Doesn't that make the equation impossible to solve?

[ivan.moony]

I'd hate to hardcode anything (if I ever get to that point) because the most certainly I would forget something important. a few dozens you say? Not much if that is all there exists.

Well, humans have hardcoded God knows what in their instinct. Maybe comparison to themselves make them understand what means yes and what means no.

Which brings us back to emotions. Maybe if the algorithm has internal reinforcement for "yes" and "no"? Like:
on "yes" -> accept it, happens when everyone agree-accept;
on "no" -> avoid it, happens when one disagrees-avoids.
And comparison to that self-behavior would give a basic understanding of what is going on with living being chosen to analyze.

[Don Patrick]

I don't think it's very complex with humans, really. If something causes pain, that means "no". If something causes pleasure, that means "yes". Everything else, even recognising which other words mean"yes" or "no", comes from that simple sensory observation. It's a classic punishment/reward system.

[8pla.net]

Many animals do use language, here is one professor who learned the language of prairy dogs through observation ...

Dolphins also use a language we have yet to decode completely ...

Perhaps we are to dolphins, what prairy dogs are to us?

[ivan.moony]

Here is how people in OpenCog intend to imlpement personality of the machine.

They deal with personality, emotion and other parameters (i.e. in anger resolution is fast, not thorough, valence is negative and arousal is high). They reference to some full blown paper on technique called Psi technique for modelling personality.

Here is the link: http://wiki.hansonrobotics.com/index.php?title=Emotion_modeling

[ivan.moony]

When I think about artificial behavior I can not resist thinking about the world of living beings. Whatever we do there is one constant that follow most of us (except maybe butterflies when they grow up) through whole life: we are forced to cause pain to others. To survive we have to kill and eat someone and that is what makes me at least sad in this world. I also can not resist thinking that this world is some teaching school that is giving us a valuable lesson: we shouldn't force anyone to cause a pain to someone else. That is also (by my opinion) the worst thing we can do to someone.

That being said, maybe can we learn something about future AI from the current situation. Maybe that could be the magic rule No. 1 built into our AI projects: "Do not force someone to cause a pain to someone else". Maybe some secret power lays in that rule, something sufficient to build all other rules and laws about being nice AI personality. Maybe it could be the only hardcoded constant in future AI.

From algorithms published in this thread, I kind of realize how to implement that kind of behavior, but I still don't know how to detect if someone is satisfied or not and how much she/he is satisfied about some action that AI or someone else would take. It seems it is not such a trivial problem. I hope for some simple solution.

[ivan.moony]

Boy, this was fast, it took me a day and a half for this idea...

So, how to recognize wish? I started from empathy, how we compare to external being when interfacing she/he. We put ourselves in her/his shoes and conclude how we would feel if we were at her/his place. Like empathy, to detect wish we just have to compare to ourselves what would we wish if we were behaving the way external being does.

Hence, this algorithm:


If we stitch this function to conflict resolver already published in this thread we get nice complete behavior mechanism that I would like to call "Sky".

I think this could be the base for an AGI. No hardcoded rules, learns by guessing (some well known genetic algorithm would apply) versus checking response and combined with second order math logic I think that Sky could act intelligently and friendly nice to any living being, considering all written in this thread.

I think that my next step should be finishing Synth web programming language (that would allow easy programming of math logic) and implementing Sky in it.

One of first Sky's tasks on my machine would be in chemistry field, explaining how artificial food could be made. So cross your fingers for me, I'll need it.

[ivan.moony]

I composed a short paper in which I summarized some of thoughts from this thread:

(Artificial Intelligence in Theory: "Sky" Behavior Algorithm)

I plan to post this paper in OpenCog mailing list as a Christmas present. Not much, but it's from my heart. If you think I should change anything or add something new, please leave a comment.

Thank you all for giving me an inspiration.

[DemonRaven]

Okay I just did a quick read through but you might this program interesting and if you can find some of the old sites on how they programmed the game it might give you ideas for  a AI. IT is the same game knytetrypper was interested in and i knew him from there but had not really done anything for a while and when i came back i found out the hard way that he had passed. The game was called creatures. The final version i found to be the most interesting as you could teach them to speak but what no one else seemed to mention is that you could also get them to talk back to you in anger if you pissed them off.  They would even hit your hand lol. They acted like a life form. They lived ate breathed air got sick and died just like a real creature. IT was another program way ahead of its time. It was the ultimate alife program. You can still download docking station the final free version and i think the game site creators aka developers section is still up at the current owners website. Check it out it would give you ideas, they spoke, played and had emotions.

[DemonRaven]

I should clarify that no i did not torture them to piss them off i would disagree though.

[Freddy]

I remember Creatures. It was KnyteTrypper who recommended the game to me. It was a lot of fun, I played the Docking Station one and a few other add ons. It was odd too because I found out that one of the people that produced it was someone I went to college with, so it was nice to catch up.

They used to do an annual festival, not sure if they still do. But yes, an interesting example an artificial life. They could learn language too I remember, there was a machine to teach them I think. It was quite sad when they died.

[Korrelan]

The original concept and code was produced by Steve Grand.  He launched a Kickstarter a few years ago to raise funds for a new game (Grandroids) that would use his new brain model.  It makes the original Creatures look like Childs play.  I donated and so have privileged access to his program diary etc.  We have received early versions and it’s all going very, very well.