Independent Learning Algorithm for Verbots

[dcorey]

Have you tried incorporating a learning algorithm program that Verbot can use along with the knowledgbases? I understand that the Learned Knowledge vkb allows you to actively teach it new knowledge, but I am referring to a heuristic where the Verbot can make its own extrapulations. This would make the chatbot have the best of both worlds. The Verbot editor is easiest to use for people without a programming background and it is much easier to have some preexisting knowledgebases rather than having to program from scratch. Yet the disadvantage of Verbot programming alone is that to my knowledge, please correct me if I'm wrong, there is no self learning algorithm or learning heuristic that it can extrapolate from to create new knowledge from existing knowledge without active teaching from the user. Basically, it just parrots the outputs it is taught from the knowledgebase or from the Learned Knowledge. If a chatbot could be programmed with a combination of prewritten knowledgebases and a learning algorithm which will activate if the chatbot cannot find an appropriate response in its knowledge base, this would be ideal. In other words, when a user gives an imput, first the verbot would check for a response from all of its knowledge bases like it does now. But if there is nothing in the knowledgebases, rather than immediately going to default rule it attempts to use its learning algorithm to synthesize existing knowledge and come up with its own response. Then if the response is not appropriate, the user could correct it. It would be ideal if there would be a way to correct it that is as simple as the following example I got from an AIML discussion board,
User: Where do I find the webpage with Australopithecus?
>Alice: Where do you want to find it?
>User: Bad Answer Alice, time to Learn a new Line
>Alice: How would you answer "Where do I find the webpage with Australopithecus?"
>User: I would say: "http://www.ancientskulls.net/australopithecus.html"
>Alice: Thanks, human. Try asking me again.
>User: Where do I find the webpage with Australopithecus?
>Alice: http://www.ancientskulls.net/australopithecus.html
What are your thoughts on this and any plans for creating such a program? Thanks, Debbie

[Somniator]

I am not quite sure but I think what you want to do is generally possible with some additional code module (VCM) programming, because you can create a knowledge base that fires a predefined number of questions to an input term and save the answer matrix in a rule.
Later the answer can be activated by a <send> command, can be evaluated (with some csharp scripting) and so on.

The aleady existing number of Verbot commands can be organized to a pretty mighty set of tools.

However the question is which kind of "intelligence" you want to have.
We have already discussed some quite advanced kinds of AI.
An example for this hard-to-develop autonomous 'intelligence' can be found here. However I don't believe that a verbot can proof such a kind of knowledge frequently in a normal conversation. Matt called it an expert system.

Basically, it just parrots the outputs it is taught from the knowledgebase or from the Learned Knowledge.

I think you should examine the "teacher.vkb" very carefully. I really wonder that this jewel did not become a prey of experienced programmers because only a trickle of it's potential is already used.
Imagine teacher rules that do not only create one but a number of rules, including others that command to unlearn the previous ones when being activated.

If a chatbot could be programmed with a combination of prewritten knowledgebases and a learning algorithm which will activate if the chatbot cannot find an appropriate response in its knowledge base, this would be ideal.

Now this is quite easy. Write into the * rule a number of verbot questions that can trigger the teacher.vkb rules.
The main problem of unexpected inputs collected by the * rule is that they are unexpected. So the Verbot should have tools for a semantical and syntactical analysis of the input. ("What is the subject? Where are the verbs and dependent terms?" And so on) This feature does not exist yet.

In other words, when a user gives an imput, first the verbot would check for a response from all of its knowledge bases like it does now. But if there is nothing in the knowledgebases, rather than immediately going to default rule it attempts to use its learning algorithm to synthesize existing knowledge and come up with its own response. Then if the response is not appropriate, the user could correct it. It would be ideal if there would be a way to correct it that is as simple as the following example I got from an AIML discussion board,
User: Where do I find the webpage with Australopithecus?
>Alice: Where do you want to find it?
>User: Bad Answer Alice, time to Learn a new Line
>Alice: How would you answer "Where do I find the webpage with Australopithecus?"
>User: I would say: "http://www.ancientskulls.net/australopithecus.html"
>Alice: Thanks, human. Try asking me again.
>User: Where do I find the webpage with Australopithecus?
>Alice: http://www.ancientskulls.net/australopithecus.html
What are your thoughts on this and any plans for creating such a program? Thanks, Debbie

The example above can easily be adopted using the Verbot Editor.
Alice got her signal to learn something by "time to Learn a new Line".
The keyword in the teacher.vkb is "remember" or "rem".
At this time the learning rule is well hidden, so that the Verbot will not learn things unintentionally.
But you can divide the rule in a childrule set, eg:

1. User says something unexpected.
2. Verbot rulename: *
3. Verbot output: Do you want to tell me something important?<mem.set unknown [_input]>
4. User input: Yes ==> goes to the childrule (yes) of *
5. Verbot output: How would you answer [unknown]?
6. User input: (gives the answer)
7. Verbot output: <send remember when I say [unknown] say [_input]>
==> activates the learning rule in the teacher.vkb
... or write alternatively a new complete learn command.
That's all.

Guess I should implement something like this in my KB...

[Matt]

User: Where do I find the webpage with Australopithecus?
>Alice: Where do you want to find it?
>User: Bad Answer Alice, time to Learn a new Line
>Alice: How would you answer "Where do I find the webpage with Australopithecus?"
>User: I would say: "http://www.ancientskulls.net/australopithecus.html"
>Alice: Thanks, human. Try asking me again.
>User: Where do I find the webpage with Australopithecus?
>Alice: http://www.ancientskulls.net/australopithecus.html
What are your thoughts on this and any plans for creating such a program? Thanks, Debbie

This is definitely possible today using the Teacher_learn_default.vkb knowledgebase that comes with verbots, put this kb on your list (nothing with a default rule can be above it in the list or that will override the teacher's default), when the default rule hits, it says "I didn't understand that, in the future, what would you like me to respond to [input]", then the user says what he/she wants and the verbot adds the new rule. Try it out!

[dcorey]

Hi, thank you both so much for your help. I want to make sure I understand how to use Teacher Default correctly. I have a few questions:
1. Do I use the Teacher Default vkb in conjunction with the regular teacher ckb or vkb or in place of it? If I am using both do I use Learned Knowledge on top, then Teacher (ckb or vkb does it matter) then other knowledge bases and then teacher default)

2. It sounds like if I am going to use teacher default vkb I need to delete all default rules from other knowledge bases. If I put my knowledge base with the default rules on the bottom, the teacher default will always have something so the default rules in any other knowledge base would be useless anyway and might as well be deleted? Am I understanding this correctly? I would want the teacher default on the bottom so that the Verbot would have the opportunity to search all other knowledge bases I want to use before going to teacher default, correct?

3. Finally, is there any way to train the verbot on line. If I compile the teachers and learned knowledges into ckbs and put them on Verbots online, then will the online Verbot work the same way?

Thank you again and thanks for updating the Julia 2 file. Have a good weekend.

[Somniator]

1. Do I use the Teacher Default vkb in conjunction with the regular teacher ckb or vkb or in place of it? If I am using both do I use Learned Knowledge on top, then Teacher (ckb or vkb does it matter) then other knowledge bases and then teacher default)

The CKB is a precompiled knowledge base. If you're using only VKB's in your Verbot Player knowledge base list you see that your Verbot always compiles all of his VKB's when loading or reloading the KB's.
If you teach him new rules frequently (so I do) the verbot wastes pretty much time compiling KB's - even those that have not been altered.
This waste can be spared if you klick on the Compile-button after saving your KB in the Verbot Editor.
Then the Verbot only loads the CKB stuff without compiling it starts much quicker.

So my advice: Use the teacher.ckb instead of the teacher.vkb, but first compile it in the Verbot Editor.

2. It sounds like if I am going to use teacher default vkb I need to delete all default rules from other knowledge bases. If I put my knowledge base with the default rules on the bottom, the teacher default will always have something so the default rules in any other knowledge base would be useless anyway and might as well be deleted?

This happens if you set up a "*"-Default rule in the teacher.vkb.
That's exactly why I don't do it.
Putting the teacher.vkb down to the end of the list is not useful though because there are a lot of superfluous, old-time "dunno"-statements in the Default.vkb that would catch the user's questions and prevent an "intelligent" learning process.
I tried out the "Alice"-concept from your posting above this morning and decided to integrate the learning command into a childrule of * instead of sending the question to a rule of the teacher.vkb.

I got the following childrule hierarchy:

Code: Select all

#input: *
#output: My brain must be fried. You wanted me to remember something?<mem.set unkn_input [_input]>
   (and other statements that require a "yes" or "no" reply)
---CHILDRULE YES
    #input: yes
    #output: You were saying: [unkn_input]... What's your opinion about the appropriate answer to this question?
------CHILDRULE CONFIRMATION
        #input: *
        #output: <mem.set desiredanswer [_input]>Do we agree that the answer to "[unkn_input]" is: [desiredanswer] ?
---------CHILDRULE YES
            #input: yes
            #output: I'll make a note of it. If you want to test me just ask: [unkn_input]
                <learn
               #input:[unkn_input]
               #output:[desiredanswer]
                >
---------CHILDRULE NO (the desired answer does not fit the first input for some reason)
            #input: no
            #output: Okay - do you like to give me another try to learn it correctly?
------------CHILDRULE YES (user wants to retry the learning process)
                #input: yes
                #output: Then please retype your statement.

This is a short survey about a possilbe learning structure. Another way can be to send an input to a rule in the teacher.vkb with a <send> command. But then the Verbot leaves this childrule structure and it is hard to get back and execute the latter childrules.

[Somniator]

Sometimes I am in awe of the powerful simplicity of the verbot command structure.

In dcoreys opening post Alice memorizes an internet link. So I tried to add on the level of the CONFIRMATION rule another one that is specialized on links.
The trigger is "http://":

Code: Select all

------CHILDRULE LINK
        #input: (open) url http://[link]
        #output: So you want me to open the url http://[link] after your input: "[unkn_input]"?
---------CHILDRULE YES
            #input: yes
            #output: I'll memorize that.
                 <learn
                 #input:[unkn_input]
                 #output:I'll try to open the following link: <linkonly http://[link]><run http://[link]>
                 >
---------CHILDRULE NO (the desired answer does not fit the first input, or the link information is somehow wrong)
            #input: no
            #output: Okay - do you like to give me another try to learn it correctly?

In the same way you can set up handles for opening applications or media or other special activities.

[Matt]

dcorey,
I thank you for your questions, I've tried to answer as best as I can below.

1. Do I use the Teacher Default vkb in conjunction with the regular teacher ckb or vkb or in place of it? If I am using both do I use Learned Knowledge on top, then Teacher (ckb or vkb does it matter) then other knowledge bases and then teacher default)

Use the Teacher Default in place of (instead of) the regular Teacher kb. As Somniator said, the ckb is "compiled" so it loads a bit faster, but either vkb or ckb will work.

2. It sounds like if I am going to use teacher default vkb I need to delete all default rules from other knowledge bases. If I put my knowledge base with the default rules on the bottom, the teacher default will always have something so the default rules in any other knowledge base would be useless anyway and might as well be deleted? Am I understanding this correctly? I would want the teacher default on the bottom so that the Verbot would have the opportunity to search all other knowledge bases I want to use before going to teacher default, correct?

Yes, this is a bit of a pain, in the perfect situation you would want the teacher learn default kb at the bottom of the list and you need to make sure any knowledgebase listed above it do not have a default (* - catch all) rule.

3. Finally, is there any way to train the verbot on line. If I compile the teachers and learned knowledges into ckbs and put them on Verbots online, then will the online Verbot work the same way?

We don't have a learning capability for the online verbot, the best thing you can do today is view your online chat logs and see what is getting asked of your Verbot and adjust the script accordingly (the export to CSV function can help with this). Also be sure to check out our Chat product:
Conversive Chat
You can add chat to your website and your visitors can browse your FAQ's when you aren't online.

We have wanted to increase the automation of this product so that you could hook your online verbot up to your chat and have it auto-respond, similar to how our enterprise Conversive Automated Chat (ARA) works.

[dcorey]

Hi Matt and Sominator,
Thank you both for all of your help. My programming knowledge isn't advanced enough to create a command like Sominator suggested, so for now, I'm just going to go with the teacher default. I tried to set it up but I'm having problems getting it to work. I will tell you what I did and I would appreciate it if you would tell me what I'm doing wrong. I am using 3 knowledge bases, my own, which is the julia knowledge base with my edits and new rules, the alice knowledge base used with the AIML converter, and Eliza. I deleted all default rules in all 3 knowledge bases by deleting all the rules entitled just *. Then I loaded the knowledge bases in the following order:
1. Learned knowledge vkb
2. Juliaedit (my edited knowledgebase)
3. Eliza
4. Alice
5. Teacher default knowledge
When I try saying really obscure things to get it to come up with the default rule on Teacher so I can teach it something, the Verbot always comes up with something that doesn't make sense and never gives me the opportunity to teach it. I also have attempted to correct it using the model in the teacher you should have said, etc, and the Verbot just responds as if I wasn't trying to teach it anything.
Would the resource files get in the way? For example, would Eliza's reflect file act as a default rule even when I deleted the default rules? Would you please answer the following:
1. What might I be doing wrong?

2. How do I cue the Verbot to learn something? Do I use brackets like in the knowledge base or just regular words? I was able to use the remember that command but in most contexts I want to use the you should have said command and then type the line the Verbot should have said.

3. Is there any way to edit existing vsn and synonym files without creating a new one from scratch? They don't appear to open in the editor. I would like to see the synonyms so I know what I would have to add.

4. Are there any other files, including resource files that I should delete to get the teacher default KB to work?

5. I'm really struggling with the WIKI instructions on conditionals and child rules. I want to get the Verbot to respond a certain way to a question depending on how it responds to a previous question from the Verbot. For example, with one of my rules, a user asks "What is an independent variable?" I program the Verbot to answer this and then ask the user "What is the dependent variable?" I can't figure out how to get the Verbot to respond one way if the user answers correctly and another way if the user answers incorrectly. I tried creating a child and it didn't work right. I'm not sure if I know what a child really is; the WIKI is confusing. Could you please write for me how to do this using different wording than the Wiki instructions?

6. Are there are other instructions somewhere on the site that are not WIKI or written like the WIKI that could teach me about child rules, replacement files and conditionals? I have read the whole Wiki manual and it is hard to understand. The wording isn't clear and a lot of the pictures don't correspond to the directions.

Thank you very much for your patience and help. Your instructions that you post in reply to questions are so helpful. I'm trying to understand the already written instructions and it seems like another language to me. I have no programming background at all. Debbie

[Matt]

1. What might I be doing wrong?

Even when you try something vague, some of your other knowledgebases may try to handle that, especially the Eliza KB, she is designed to work with just about anything you might say. I would first start with the learned knowledge, one other like Julia, and the teacher learn default. You may want to try your vague input within the verbot script editor, the test knowledgebase tool, then you can see exactly what input in the script is firing. One thing that could be firing a rule is a "capture" input with the syntax: [variable_name] in the input you can match to and capture any part of the users input, so this can behave like a default rule and it's not just scripted like "*" like the default. For example you might make a rule (Eliza has lots of rules like this)
input:
Tell me about [thing]
output:
I'm sorry I don't know much about [thing].

2. How do I cue the Verbot to learn something? Do I use brackets like in the knowledge base or just regular words? I was able to use the remember that command but in most contexts I want to use the you should have said command and then type the line the Verbot should have said.

If you get this setup correctly this should just work.

3. Is there any way to edit existing vsn and synonym files without creating a new one from scratch? They don't appear to open in the editor. I would like to see the synonyms so I know what I would have to add.

Yes, you may not be able to see the .vsn files by double clicking in windows on the file, open the script editor, then choose tools -> synonyms, then open your file.

4. Are there any other files, including resource files that I should delete to get the teacher default KB to work?

I don't think the resource files are causing the problem, so that shouldn't matter.

5. I'm really struggling with the WIKI instructions on conditionals and child rules. I want to get the Verbot to respond a certain way to a question depending on how it responds to a previous question from the Verbot. For example, with one of my rules, a user asks "What is an independent variable?" I program the Verbot to answer this and then ask the user "What is the dependent variable?" I can't figure out how to get the Verbot to respond one way if the user answers correctly and another way if the user answers incorrectly. I tried creating a child and it didn't work right. I'm not sure if I know what a child really is; the WIKI is confusing. Could you please write for me how to do this using different wording than the Wiki instructions?

A great way to learn this is too look at an example, check out some of the quiz knowledge bases in the download area, they do this kind of thing and work pretty well.
State Capitals
Country Capitals

6. Are there are other instructions somewhere on the site that are not WIKI or written like the WIKI that could teach me about child rules, replacement files and conditionals? I have read the whole Wiki manual and it is hard to understand. The wording isn't clear and a lot of the pictures don't correspond to the directions.

I'm sorry it is not understandable, where the wiki fails you should be able to search through the forums for good information or just ask people directly, as you have seen we are very helpful around here.

One tip: since our forum search isn't that good you can have google restrict your search to just the verbot site, so for example you want to search for replacement profiles, try this in google:

site:www.verbots.com replacement profiles

[dcorey]

Hi, I've been reading Sominator's excellent ideas and links provided about independent learning algorithms. I've also looked at Jabberwacky, Ultra Hal, they all have a heuristic type algorithm, plus like Sominator's link stated, these algorithms were available in the 80's. The problem with the other AI programs is that the preexisting knowledge bases are limited or difficult to edit without programming knowledge. Verbot is ideal because even someone like me is beginning to get the hang of editing knowledge bases and creating new rules. The only crucial element Verbot is missing is that learning algorithm where the Verbot can make connections and extrapulations with existing knowledge without having to be directly trained by the user. I have been reading on the forums that this idea has been discussed extensively, but no one has discussed concrete plans for creating a learning heuristic in the Verbot. I don't even want to begin to understand how to write such an algorithm.. I'm challenged enough as it is learning the basic skills. I do know, however, that such an algorithm is possible since Hal and Jabberwacky have it. My question is, is there someone in this community with a programming background that could and would be willing to create such an algorithm, similar to a Hal, Jabberwacky, etcfor the Verbot and put it in the community download as a KB. This way, those of us without the programming background don't have to learn how to do it and don't have to understand how it is done, but everyone can put the algorithm into their Verbots and use it in conjunction with the other knowledgebases. In summary, if other programs have them, we know its possible and this would be a crucial next step for the Verbot. If there was someone that was able to do this, it would be a wonderful gift to the Verbot community. I wish I could do it myself, but I know you have to have advanced programming skills to do it. What do you think?

[Matt]

Hi, I've been reading Sominator's excellent ideas and links provided about independent learning algorithms. I've also looked at Jabberwacky, Ultra Hal, they all have a heuristic type algorithm, plus like Sominator's link stated, these algorithms were available in the 80's. The problem with the other AI programs is that the preexisting knowledge bases are limited or difficult to edit without programming knowledge. Verbot is ideal because even someone like me is beginning to get the hang of editing knowledge bases and creating new rules. The only crucial element Verbot is missing is that learning algorithm where the Verbot can make connections and extrapulations with existing knowledge without having to be directly trained by the user. I have been reading on the forums that this idea has been discussed extensively, but no one has discussed concrete plans for creating a learning heuristic in the Verbot...What do you think?

Agreed, the Verbot 4 engine today is more of a simple pattern matching engine, it was not designed to do some of the advanced heuristic features that those other bots use extensively, AIML/A.L.I.C.E. is more closely comparable to the verbot engine in terms of the inner-workings. The beauty of our technology, as you mention, is the simplicity of designing the brain, where in the other bots it requires much more research and time to develop a "good" bot.

Since Verbots wasn't designed from the ground up to use this type of algorithm it is very difficult, if not impossible to simulate this functionality in the current verbot engine. Maybe Leseur could chime in here, he has made some development efforts towards expanding the capabilities of the verbot through the use of variables.

I don't fully understand how jabberwacky and ultralhal work, because I haven't played with them very much, I have been thinking lately about how an engine based completely on word associations (similar to an ontology) would be neat, it could even come up with an entirely fluid/dynamic response based on the users input, this is maybe how those other bots work today, I'm not sure.

[Somniator]

Guess for an "intelligent" bot you need 3 things:

1. A bot who can fathom what things are (like 20Q) and compare them to others.

2. A bot that can analyze a written text and pose reasonable questions. (like SAM)

3. A feature that comines these two abilities to a network of meaning.

Well, number one is possible with the present Verbot toolset, so I am convinced.
I've begun to check this out - however it is damn complicated and a heck of work (for a freelance hobby programmer).
But at least I know the way to do it: Every new term can be saved in a rule containing a numeric information about it and be loaded at _startup with a volley of <send> commands that store all of this Verbot knowledge in a huge array list / string list (that has to be converted to an array whenever the rule get's an input ).

Guess a new Verbot version should not only enable an overall string variable list (like "vars") but also an overall list of array lists. (or - I don't dare to imagine this - database access).


2. More important than no.1 is the second feature. Verbots would be so much more entertaining if they'd ask appropriate, never-before-heard questions about an input or a written text.
But even identifying word classes (like "verb" or "substantive") requires an immense amount of grammar knowledge, i.e. HUGE KB's and a lot of coding. - The "pluralizer" VKB/VCM that uses a quite simple algorithm is nothing compared with a real vocabular- and grammar-parsing analysis.

The best thing would be to dig out somewhere the source script of SAM ...

3. No.3 is too far away from me to waste words about it...

[Matt]

I think you are correct Somniator, a frame based system where the engine knows (or at least approximates) the underlying meaning of a user's input is the ultimate in chatbot technology.

You should take a look at FrameNet it is an open framework for this technology, I haven't played with it at all though.

-Matt

[Somniator]

You should take a look at FrameNet it is an open framework for this technology, I haven't played with it at all though.

Thank you for this link.
Oh my gosh, apparently there exist a lot of experts doin a bunch of work in this matter!

I am a bit resignative now about the amount of work to translate all this to the scripting system of the Verbots!
Maybe it is possible to build some rules that can use the commandline version of the tree tagger, a tool that analyzes the words within an input sentence and parses it. --> Hereyou can try it out on line.

However, there's a long way from this to the intriguing abilities of semantic parsers and word frame analyzers or all-in-one editors - and after all analysis isn't worthless without storing it in rules and an additional programming of flexible, apporpriate questions about the read Verbot input!

Guess it would be clever to contact one of these expert system developers directly and recommend the Verbots as an amazing output system for their work. Up to now their tools seems pretty dry to me!