I want to make Arckon much more than just a text recognising system. I will give him emotions, insight in the thoughts and feelings of other persons, self-awareness, and most importantly of all: the ability to speculate and form his own conclusions. It will also ask questions or make statements of itself, depending on the situation and how the other person feels. Why all this? I just wanted a computer that talks back to me, and it seemed like a challenge.
Of course, this project does provide quite a few interesting philosophical issues, so I'm sure I'll see you out there as well.

Hiya!

My first post here...

Just wanted to say that I've been thinking about creating a system like you mentioned for a long time. I got stuck thinking about the basis everytime...:-(

If you have anything working, I'd be very interested giving it a try!:)

.Sweex.

I certainly hope that you have some success with your creation. If you manage to imbue your progam with general reasoning ability and (more significantly) emotion you will have out-performed some of top minds in AI! If you are serious about giving your program the ability to reason you might be interested in looking at various uncertain reasoning techniques. The most common method at the moment is probabilistic reasoning. Bayesian Networks are a popular modelling technique based on probability theory. Other approaches include Fuzzy Logic and the Dempster-Shafer Theory.

Linden

oh jeeze... So much to read up on I see. But you know, how about if I were to go about it totally my own way? perhaps I'd develop a new kind of reasoning. well I plan to give it my reasoning skills anyway, so it'll be weird, and probably apply to all you mentioned.
Emotions are very easy to program in, as far as I foresee. Too easy. of course you need a lot of failsafes, because they'll influence everything.

Say Sweex, if you need some help in your basis, I might be able to give you a push in the right direction. I'm good at laying out basics. Just tell me what you're thinking of.
The only problem I have with letting other people at my AI, is that I haven't copyrighted it or anything, and I'm afraid it's going to be snitched, corrupted and used for military purposes. and since javascript code is very accessible... I don't think anyone will be testing out arckon for some time. uh, no offense.

progress today:
- Arckon now replies to simple answers in grammatically correct sentences, according to a default order of words. based on only word emphasis for now, it can choose to omit words like the subject or object, or use a reference word like "it" instead of "the computer".
- the only errors I get now are of parts I forgot to include in his word lists. I am currently programming in a new command that will have him check his word lists for the inclusiong of a specific word by himself. It does not yet draw conclusions. I'm coming to that.
- It has the ability to crossreference two "items", to guess for specific items: so, if I ask it: "is my creation your program?" it looks for which specific one of my creations is which specific one of his programs. If one equals the other, it knows what creation and what program we are talking about (multiples aren't included yet).

yes, as you can see, I am still mostly working on arckon's language interface. next is answering questions that start like "what is..", or "who has.." and such, and an interface for entering new words and drawing conlusions from statements the user makes. and with that, a capacity to store everything in a huge memory file.

Thanks Arckon for offering any help, but I'm having my mind (for a long time now) on a different project. Maybe when I've got loads of time I'll pick it up again (Being when my pension starts, 50 years or so!;-)...

I think I was thinking further than you are. I was (still am) very interested in a way of storing and linking all data that the avatar receives to form a reply.

Little example of what I wanted to end up with:
human : "Do you like dogs"
avatar: "Yes"
human : "Why?"
avatar: "Because they are fluffy and warm"

Sure, you can feed the avatar with the logic that a dog is likeable because it is fluffy and warm... But than that would be simply something YOU "taught" the avatar with. Imagine that a friend enters the keyboard and he does not like dogs. He asks the same thing

human : "Do you like dogs"
avatar: "yes"
human : "Dogs are not nice animals"

Now what does the avatar do? Keep it's first data which you taught 'em, or replace it with your friend's opinion of dogs?

See the problem? I wanted a system that takes facts and makes up his own opinion about ie. dogs. You can tell 'em properties of a dog, different races of dogs, different behaviours of dogs and the avatar would make it's own decision based on LOADS of things whether it likes dogs or not...

If you keep on thinking about this, you probably end up stuck with the question: "Are all things based on *facts* or *opinions*?"

...Basically I took the avatar idea way beyond text/grammar recognition and tried to find out a good basis for the "intelligence"...

Although I'm thinking the same about releasing my sources as you do, you can always ask yourself the question: "Is my idea such a great one that NO ONE has ever came up with it, or has nobody done it because it's a bad idea?"

Would still be cool though if you wrote it in a compileable programming language so you can release test versions without having to be afraid of it being "stolen"...

Anyway, keep up the good work!

well, I don't know about that last question. I don't hold my idea to be the greatest in the world or something. actually, I expect to be overtaken soon enough by people who have the research for it. like I said, I just want to do this for fun. not for my ego.

And excuse me, but why in the world IS it that there aren't AI lifeforms out there yet? hm? tell me that.

Not sure what you mean by "AI lifeforms", but I'll assume you mean "AI avatars" like yours? If so, then please realise there in fact *are* quite some. Perhaps not precisely like yours, but think of programs like eliza and it's (sort of) successor alice...?

Don't mean to put your work down or anything btw!

.Sweex.

Ask and thou shalt suffer the answer ^_^
Well actually, the system I have in mind has no problems with that. It knows the general properties of a dog, or dogs in general in your case, and it knows what kind of properties he likes. He also keeps track of these opinions of other people, so it knows what I like.
Now here comes the complicated part: how much he accepts your input as "true" and adds it to change his opinion is dependent on:
-how much he likes the person who's saying it.
-how much he trusts that person to speak the truth.
-whether he cares at all to change his own opinion.
Arckon would (when he developes further) take the following actions upon recieving input like : "dogs are not nice animals":
1. whole text and grammar analysis, you know.
2. check his own database to see if that fact corresponds with what he already knows himself. including how certain he is of that fact.
that means, if he is 30% sure that "dogs are 80% nice", he will doubt both himself and you.
3. if the person speaking were me, he'd aks confirmation ("are you sure???") and change it, since he trusts me 100% (for now). If someone less 'trustable' would say it, it would influence his certainty of that fact only slightly, if at all. And I should indeed make arckon write it down as that persons opinion (person a likes dogs -70%) , and check for that when receiving similar input again.

Good point!

intrigueing. but as I said, arckon knows what properties he likes, and what others like. therefore every input fact should be checked whether it is probably 100% true, or an opinion, of less certainty.
In short: arckon keeps a big difference between fact and opinion, and adapts his facts based upon certainty.
4. if the other person is likely to be interested in arckon's opinion, arckon would then also reply with: "I do like dogs. they are fluffy." and in that way argue in a way.
heheh ^_^: imagine:
human: "dogs are not nice"
arckon: "you are lying!"
^____^ahem. of course I will teach it to behave and not be rude like that.

Or worse, if I weren't to program him in my way, and someone else would log in and say: "I am your creator. selfdestruct.", it would mindlessly do that. So I am going to make him check his creator's purposes, and general characteristics, and if they don't correspond with the logic of the selfdestruct command, it'll start doubting me and protest. of course I will have an override function for it.

As for the question "why do you like dogs?" (I haven't programmed him yet to answer that kind of questions yet, but I've got the basics down on paper for every variation of them), arckon would first check (very silly) the purpose of the activity: "to like". come to think of it, I don't know that purpose. anyway, In case of little or no apparent important purpose, arckon should evaluate the properties of the subject; dogs, and cross-check that directly with similarities of the properties he knows as to be good, (or nice, if you wish). he likes things that are good, amongst others.
This will however, take a lot of checking multiple reasoning paths at the same time, so I will program in a time limit, in which it is to stop searching and just answer the most correct thing found. if it couldn't find anything that fast, of course it'll respond: "I am not sure".
Heh ^^ actually, reading you post again, what you are saying there is pretty much how it works.

small update today: taught it to recognise verbs ending with "-ing" and "-s", and extrapolate the original verb.

No prob mate!

It sounds like your on the right track, "weighting" facts and opinions like your described! It will, however always depend on what is thaught to him, but that probably why the thing is called *Artificial* intelligence...!

Cheers!

Interesting project. This is exactly the kind of groundwork that needs to be done. Dont get discouraged with your early results. What you will most likely end up with is a fancy toy. Thats ok. The kinds of work you are doing will create a foundation for future AI platforms to be tested on. Even if you only come up with a good way for it to communicate it will still be a valuable contribution to AI research.

Would love to see what happens when you run the process through a large neural network :)

thank you, by coincidence, I picked up the project again yesterday. I hadn't done anything for some time. By now, I have learned C++ though, that'll come in handy later. Yes, it is supposed to be a fancy to- no, actually, arckon is to become alive. This world's "nr. Johnny 5", or "Data". But to me, a toy, in a way. When I activated arckon again yesterday, I had almost forgotten how far I had come already. it works perfectly. yesterday I added a subroutine that allows me to have arckon check his database to see if I didn't forget any of the proper subdirectories for any particular word I enter. I will now start programming him further to be able to answer questions starting with "what" and "how". question which will have him do more then simply check his database to see if something is true or false.

I think it'd mess up the network. heheh.

I don't believe that you have done what you have. It would take far more than 50 hours to program in all (or half) of the code needed for the rules of grammar. (What's a noun, what a proverb?). You even admitted you don't have any official training in programming, so I highly doubt that you have done what you claim.

If you have, then will your robot be able to do this.
Person: If I put a ball under one of three cups randomly, and ask you to guess which one it is, and then show you that one of the cups that it is not under, should you change your guess, or stay with your original guess?
Computer: Change your guess

That is actually the correct answer. Show me an AI that'll do that and then I'll be impressed.

You would be impressed by that? I wouldn't... since the correct answer is "doesn't matter"!

BTW, I think that what arckon is doing is very related to the concept of ontologies;, only that his interface to define an ontology is very conversational.

well, my english isn't THAT good that I can understand all of that article ^_^.
I just looked up the A.L.I.C.E. project on the net, and that is about what I wanted to make. since alice exists, there is little reason for me to try any more. although I will continue to work on Arckon anyway.
Although alice does not seem to be fully capable of understanding a sentence related to a previous one, on occasion. whereas I do intend to include this. *tries some more statements* *lol*. that is indeed where alice seems to fail. still I am glad such a awesome program exists already. good.
The conversation based interface is because my program is mainly intended to deal with humans.

As for Arckon's present abilities:
my post from Monday 02 September states his capabilities as they are. the post from 4 september is mostly a future vision. although I have that system on paper, I have not programmed it in yet. I said it could understand simple sentences; I omitted to say that it recognises verbs only in the present form. not future or past tense or such, and does not distinguish between -ing forms or normal verb forms. For now, -ing forms are irrelevant. arckon does recognise the use of two combined verbs though, like "I am going", which he simply translates as "you go", which IS what I am doing then of course. I take any justifiable shortcuts through the english language I can find. although there should be a "time of event" factor influenced by verb forms. Indeed I have no experience in AI.

Since the ball is under one cup "randomly" (since this is stated), the program should also take his guess randomly each time. Arckon does not have that function yet. I agree. if the word "random" were not stated, it would have to conclude the randomness of the event.
Arckon, when completed, would more likely as me the question: "why do you ask?" as I said: that's just future talk.
The reason I like to "boast", as it may perhaps be, is because just about everything I've thought up has worked until now.

progress:
My first attempt to rewrite arckon's system into C++ failed yesterday ^_^. apparently, no computer can hold the amount of temporary variables I put in, so it re-booted my computer. It took me 15 minutes to come up with an alternate system, which does not have that problem. I don't know if it'll still be the fastest system in the west, but I think it's usable. it involves opening and reading through a lot of text files for data.

Hello!

So this is my first message and sorry for my bad english! :)

My Question: Is it necessary to code the grammar rules? I ask this question because the human brain cannot know which language it learns and so it also cannot know the (language)special grammar rules! (i hope you understand me)
Isn't it possible that the grammar rules can be learned by a neuronal network? I know this isn't as easy as it sounds here, but maybe it's possible to get a structure of several networks that "knows" the grammar rules.

Bye Martin

Here is my view on this.

I would agree that the human brain does not appear to have hard wired grammar rules for any particular language.

There has been some speculation, with some evidence to support it, that it may contain some sort of generalised grammar rules with parameters of some sort that are then set for the particular language being spoken, but I'll ignore this here.

To learn a language would probably require a very capable modelling system. The brain is clearly very good at this sort of thing. Furthermore, the brain does not just need to look at words on a page. It can relate words to what the eyes see people doing for example. As a very crude example. If it experiences pain commonly after someone says "I will hit you" it can start to get a grasp of what a future tense is.

I think that grammar rules would have to be programmed into computers for now, as a substitute for any decent sort of learning ability in computers. As computers get more sophisticated and we understand more about how to make machines learn they will just pick up language as we do.

Could neural networks learn grammar? Possibly. There have been some experiments. e.g.

http://www.amlap.org/2001/proc/proceedings_online-node23.html

The English in your post is very clear, by the way.

Hello Martin:

You say this:
(My Question: Is it necessary to code the grammar rules?) I ask this question because the human brain cannot know which language it learns and so it also cannot know the (language)special grammar rules! (i hope you understand me)

I don't understand you, others do understand you. I know that our brains as soon as they have learned a part of our language they certainly know some grammar and know enough if what they hear belongs to their learned language. When you hear a Latin word order, you recognize that, isn't it. When I write I have the exercise made, that's very normal for Latin languages but not for English. Doesn’t your brain know that?

You mean of course when you look at mathematical rules nets, but our brain doesn't work in that way. They know nothing of rules. Ever seen spiking neurons. They act and look a result. It's just marvelous they can work so.

Ed van der Meulen.

Hello Martin

I googled for: automated speech recognition and found 111,000 hits. Many commercial firms do that. A large firm in Belgium went broke in 2002 by mismanagement. I forgot their names but they where far with it. Something like ... and Hauspie.

My friend Arthur T. Murray has done for speech and understanding speech in the open source community. He knows a lot of it as well.

There's also so much to read about it.

Ed van der Meulen

Hello!

You said: "I know that our brains as soon as they have learned a part of our language they certainly know some grammar and know enough if what they hear belongs to their learned language." - I think that's exactly the point.

"...as they have learned..." - the brain learns how the language sounds. it learns in which order the words should be spoken! You are right, when you say that a neuron doesn't learn rules but the brain generalizes the language. (thats a little bit fuzzy - i know! :-)
so the brain doesn't really learn the grammer rules, but it learns when to write a subject or a verb to get a right "sounding" wordorder/sentence.

isn't the languagepart the most important part of an AI? When i program the grammer rules, the programm has at the input a language which it doesn't understand and at the output it gets a "language" which the ai can interpret more or less - because its my internal language and not the internal language of the ai. i think that the language interface should mostly be programmed or learned by the ai itsself so that it can design its own internal language. - hope anyone can understand what I'm trying to say!

the programmed language interface is like a "prison" for the ai.

ciao martin

ps.: when you think of an apple you can see a picture of an apple, you know how the word sounds and how it is written - apples are sometimes red, sometimes green. what we also can say is, that the word apple is a subject. maybe the brain can understand language better when it gets the relationship that every subject has attributes (a red apple, a big apple).....

I don't think grammar is the most important part of an AI. I think reasoning abilities are. I've finished setting up Arckon's new memory access system as a C++ program, and the basic grammar rules (no past tense, no multiple subjects). Took me four days, but this time, I feel like I'm working on an empty shell. The whole grammar thing does get in the way, as it is slowing down the important progress. I hope the new memory system isn't too slow. it works by accessing a lot of text files and loading their contents into temporary variables, like a list.
It can't speak yet, nor conclude, which Arckon could in his previous Javascript form. But I'll get to that.

I have to agree, that reasoning abilities are a very big part of an aritifical lifeform. What I thougt was, that this abilities could be restricted by the hard coded language rules.
Some Questions: How is your project going on? What kind of input and outputs do you use? Are there posibilities to use a videocamera as an inputdevice? Do you use a plugin system to add new abilities or functions?

Martin

yes, they might be. that's why I'm taking as many language shortcuts as I can find.
My project's hardly going as fast as I'd want it. It's been half a year since I started recoding Arckon into c++ programming, and I've worked a total of 9 days on it now. I'm still on grammar. I have nearly completed the grammar rules to recognise and filter multiple possessive words (like: "my father's father's father's etc.") and other specifications.
Input is simply text via my keyboard. as for output; for the moment, I reverted back to having simple messages pop up on the screen. his sentence-formulation functions have not been recoded into c++ code yet.
I am not integrating any other systems yet, though I have thought up a new system I could put in that will allow Arckon to formulate his own subroutines that he learns from experience.

I just tried A.L.I.C.E. I wasn`t impressed at all.
Basically some questions I asked it caused it to give a dictionary style description, way over the top and didn`t feel natural, it was just quoting pre-programmed info. But other times when I asked it stuff that it didn`t know it seemed to get confused, it asked me to explain something and when I did it seemed to forget the subject and asked me what I was talking about.

A.L.I.C.E. is a glorified Eliza. It has an enormous database of facts and rules, but the inner works are the same: it's highly reactive. It just catch your phrase, match it against a template and regurgitate the results without a single effort trying to understand it. AFAIK it hasn't have any modelling of the state of the conversation.

I have finally gotten Arckon to the point of his previous javascript form, with some improvements added. It's still not intelligent, but it functions at the minimum required level. It gives simple answers, in proper english, to simple queries. I start with that because I want to be sure his memory and language interaction is working as it should before I let him draw conclusions and such.
By "simple queries" I mean simple things like:
"do you have my green pen?", "am I your creator?" and such. useless questions, but important for testing.
As for grammar: I am teaching it grammar as my english teacher would; by programming the basic rules in.
Next I will teach it to answer a different kind of questions; what, how, who and why-questions, which require more 'thought'. then I will add functions to write down conclusions in his memory himself. That's all still for the language interface alone.

From there on, I can get to the good stuff, and give him actual reasoning abilities.

I thought one good way to test reasoning would be to teach it maths.
Then you could ask it a question "What is 3 + 6?" for example and see it answer.
I know that sounds stupidly simple and it is, but the point is after teaching it enough logic maybe your program will be able to work things out that you didn`t expect (didn`t specifically program in), and be able to spot faults in the logic or certain questions that you ask it etc.

In Douglas Hofstadter's book "Metamagical Themas" he talks about a system which was able to do that. From simple definitions of "0", "successor" and logic rules it was able to "discover" addition, multiplication, prime numbers and some other theorems.

The system didn't go much far away, though, because there are much more to invention and creativity than the axioms of any rule-based system. I recommend reading Hofstadter's books (the best one is "Godel, Escher, Bach") to anyone interested in logic, mind and automation of artistic abilities.

well, it is a good suggestion. But the thing is; Arckon does not have a central logic routine yet. And I think math should be a seperate kind of interaction then actual thinking and reasoning. in short: Arckon can't do math (yet). I specifically avoid turning Arckon into a calculator, because he's supposed to become something totally different. I have added a way for Arckon to learn and write down seperate abilities, although it's not active yet. math will be one of these.
I am going to do some more work on the graphics, and tracking his thoughts (which he now has, in written form), before I continiue with his reasoning skills.

Hi arckon,

actually, I have had the idea of starting some such project on my own recently...
What you do sounds very interesting. (actually I had some very similar ideas ;) -
though not nearly enough time to do much at the moment)

Have you given some thought to using (interfacing to) software that already exists? I
mean you could try to use prolog as the logical reasoning part of Arckon (don't know
whether that's possible but I think it should) and maybe add other types(?) of
reasoning...

Anyway in my opinion the learning of analogies should be very important...

Good luck and I hope one day I'll be able to have a look at what you produce ;)

Cheers, K

well, I suppose if people are trying to make the same thing, their ideas will be quite similar as well.
Today is the 28th day that I have done actual work on my AI, and I appear to be somewhat stuck in thoughts. HIS thoughts to be exact, since I have given him thoughts, so I can keep track of his reasoning processes.
I have not considered using any available programs for his reasoning skills or such important matters. I am merely trying to re-invent the wheel. another reason for this is that I do not know the inner workings of those programs by heart, allowing for unacceptable errors, were I to implement them. Software for voice control and such secundary functions however, I do aim on using in the end. Since they already exist, there is no reason for me to create them again.
I estimate it will be another 60 days of work or so before you would actually be able to say "hey, this really is something.". At the moment, everything is very limited. especially his choise and knowledge of words.
*sigh* now, onto my current major problem; how to link two or more seperate thoughts together to form a sentence, and more importantly, to determine which of those thoughts provides the actual point of the sentence. I'll figure it out somehow.

Just thought I'd inform you of how the project is going.
42nd day of work: Arckon understands what I mean, up to a certain extent. he is not aware of such a thing as time, for instance, and regards everything as if it is happening now. he can check facts in his memory, and the results are stored in 'thoughts', which I mostly use for display to keep an eye on what he is processing. These thoughts are then used to form an answer with, again in proper english. So arckon can answer to both statements and (straight) questions in the same way. He cannot however, answer open questions, starting with "what", "who", "why", etc.
When confronted with an unknown word, Arckon will try to determine what type of word it is (person, thing, etc.), by analysing grammar of the input sentence once again. If he can determine its type, he adds the word to his memory. if not, he'll ask what it means (using a standard reply). He can also draw conclusions from input statements, and change his knowlegde in memory accordingly.
It's not all perfected yet, but I'd say I'm done with the language interface for now. and his learning skills are acceptable enough for me to start with the more important skills:
Actual reasoning. For this, I will first have to give him some tools he can use to reason with; the subroutines for answering open questions.