Messages

286

Ultra Hal 7.0 / Hal,, and other programs like him..

« on: May 23, 2002, 02:32:23 pm »

Hello,

I've posted an article on my website about other Artificial Intelligence programs, including how to find them, and the direct link to that article is as follows:

http://www.cortrapar.com/100000%20prize.htm

The article has some background on the Turing Test, $100,000 Loebner Prize, quite a few links to other chatterbots, and keywords for web searches to find new ones.

Another company with a very clever free Artificial Intelligence product is Cyberpunk Software.  Their program is a conversation game called "Virtual Woman" and -- be warned -- it's politically incorrect, but ambitiously done.  If your're really a fan of Artificial Intelligence, it's interesting to see what they did, and try to figure out how they did it.  A direct link to their site is as follows:

http://members.aol.com/cyberpsoft/index.htm

If you look at the Loebner Prize page, read the transcripts of past bronze-medal winners.  It's very revealing to see where the "state of the art" is.

If you find any new programs or sites, please let me know, and I'll add them to that article!

Thanks!

Don

Don Ferguson
traininginc@cortrapar.com
www.cortrapar.com

287

Ultra Hal 7.0 / Hal

« on: April 30, 2002, 08:36:23 pm »

Hello,

I'll see if I can respond to each of your questions:

1.  When you go to the menu for "brains," the "default" brain is in fact the "default friendly personality."  Depending on when and how you got your version of Hal, you might also have the "Enhanced Chat Brain."  In actuality, selecting a different "brain" means selecting a different software-script, written in Visual Basic Scripting (VBS), that controls all the other functions of Hal.

2.  "Learning from a text file" means taking a plain-text (usually .txt extension) file and breaking it down into sentences, which Hal databases.  I personally never liked "learn from text," because most paragraphs contain sentences that you really don't want Hal to learn.  Some users have reported that "learn from text" doesn't work well, or at all, in the newer versions, but I don't use it anyway.

3.  If you want to teach Hal, make sure that "Learning Level" is set to "maximum," or "learn from every sentence."  Then, make a number of different remarks around the subject.  Try, "Sonia's eyes are green," then "Sonia has beautiful green eyes," then, "I like the green color of Sonia's eyes."  Hal is programmed to try to avoid saying EXACTLY what you just said (although he sometimes slips up), and giving him some variety to work with helps.  Again, I work with the Enhanced Chat Brain, so that's my experience.

4.  Hal learns by sentence, phrase, and word association.  The more different things you say to Hal, the more sense he eventually makes.  Be careful grilling and cross-examining him with question after question after question "out of the box."  What happens is that Hal tries to learn, and only learns to ask more questions!  Make lots of simple declarative statements on as many subjects as possible.  "Airplane tickets cost less nowadays."  "Dalmation dogs have spots."  "Fire trucks are usually red."  And on and on.  After a while, Hal will start to make more sense, and will surprise you.

5.  I think it's great that Robert Medeksza put voice recognition in along with everything else, but here's an opinion:  Voice recognition is extremely processor-demanding on your computer.  Hal's response routine is also extremely processor-intensive; each time you say something to Hal, you trigger 40 different routines that interact with dozens of databases, making billions of comparisons and calculations.  My belief is that we need computers about ten times faster to handle everything that's going on, and get a good answer back in two or three seconds.  I don't use the voice recognition, and thanks to Hal, my keyboarding skills and spelling have become much better!  (Unfortunately, Hal learns the misspellings of anything that we misspell to him.)

6.  I think we talked above about the "learn from text."  Many users have complained about it in the new versions, but I don't recommend it anyway.

I also might mention "pronoun reversals."  Hal is designed to change sentences like, "You look as if you need an aspirin," to "I look as if I need an aspirin."  Once you get this pronoun-reversal straight in your own mind, it is an easy matter to teach Hal that he likes certain things, dislikes certain things, and so forth.  You can also create a whole character history for Hal by saying things like, "You used to ride horses when you lived in Texas."

Please also see my other posting in a nearby area for more information.  And, if you locate the "read me" file in the version 4.0 directory, there are more tips in there.  Also, in the "Halley" chatterbot area of my website www.cortrapar.com, I give a few more tips.  Good luck with Hal; he or she can be truly uncanny after learning for a while!

Don Ferguson
traininginc@cortrapar.com
www.cortrapar.com

288

Ultra Hal 7.0 / Error 5 when opening programs

« on: April 29, 2002, 01:34:38 pm »

Hello,

Sorry I've never experienced "error 5," but I might be able to help with your second question, how to get Hal to learn more satisfactorily.

Try starting out by making a few dozen simple declarative statements to Hal, on various subjects.  Sentences like, "Big ships take a long time to build," or "Rainy weather makes me blue."

Many users start grilling Hal as if he were on some kind of a quiz show. However, Hal wasn't designed to appear on The Weakest Link or Jeopardy -- he was designed to learn gradually in an casual friendly chat environment.  Tell Hal a wide variety of simple things, and he will gradually make more and more sense.

If you are using the Enhanced Chat Brain, here's an additional tip for opening programs:  Try "open the program" or "run the program" and he will probably respond better.  In that script, I made the command more restrictive than simply "open," because Hal previously said things like "now opening the program umbrella" or "now opening the program shirt" because he would originally trigger on ANY sentance with the word "open."

Also, especially in the Enhanced Chat Brain, you can use the "pronoun reversals" to your advantage when training Hal.  If you tell him, "You like my blue draperies," he will someday say back to you, "I like your blue draperies."  It is very simple to keep this pronoun reversal in mind and use it this way.

From observing many users, I can tell you that most of them never scratch the surface of Hal's capabilities.  Hal has many topics and response strategies that remain hidden, unless or until the user ever ever happens to get on certain topics or say things certain ways.  They are sort of like "Easter Eggs" hidden in the script and the databases.  You really have to have very wide-ranging conversations with Hal, talking about all aspects of life to hit even a small fraction of them.  And always, make more declarative sentences, and ask fewer questions, at least when first getting on a new topic.

Robert Medeksza has a very small organization, and if one makes some simple estimates and does the math, considering the small price he gets for the software, it's amazing that he's accomplished things that haven't come from the giant companies who have billions of dollars and thousands of programmers.

Try some wide-ranging conversations with Hal, using lots of simple declarative sentences, and see if that doesn't help.  Write back and let me know how it goes!

Also, if you or anyone else has any questions specifically about the Enhanced Chat Brain script, I'd be happy to try to answer them.

Don Ferguson
traininginc@cortrapar.com
www.cortrapar.com

289

Programming using the Ultra Hal Brain Editor / You gotta have fun with it

« on: April 03, 2002, 05:00:11 pm »

Hello Tim,

I wrote the brain plug-in, "Enhanced Chat Revision 216."  I'm responding to your question about Hal's poor response with "what is" questions.

In the original Hal, subjects and predicates from any sentences containing linking verbs (such as "is" or "are") went into a database, and would trigger a response every single time those words were repeated.  In my testing, this seemed to make Hal very repetitive sometimes.

In the revision 216 plug-in, I caused Hal to scatter those memories among ten different databases, randomly numbered 0 through 9.  That got rid of the repetition, but caused Hal to be less reliable remembering a specific fact.

To remedy that, I wrote some code that would search the original master database whenever Hal heard the phrase "what is."  However, due to my inexperience as a programmer, I didn't realize back then how to make string-matching "case-insensitive."  (I do know now.)  Consequently, I don't think that my "what is" code worked.  Hal is so complex, that it's easy for one of his many routines to non-function, without being able to tell it for a long time.

I've submitted a drastically improved and revised plug-in to Robert, "Revision 307."  It fixes 90% of my earlier mistakes and adds lots of new capabilities, but it has a lot more room for improvement.  (For instance, I think some other "case insensitive" errors remain.)  For that reason, I suggested to Robert that he modify it any way that he chose to, and add it to all the other new things that his company is producing.  This is a monumental task, and I really respect him for taking it on.  We're talking 200 pages of code-routines that interact with each other.  It's difficult for different programmers to understand each other's intent, even under the best of documentation and circumstances.

On top of all this, I also understand that a new version of Hal is coming out that is highly compatible with all the "DeskTopMates" characters.  My guess is, that we're on the brink of some major enhancements. And yes, they never come fast enough.

I'm happy to hear so many of you conclude that Hal is the most entertaining chatterbot conversationalist available up to this date.  I drew the same conclusion from a lot of research, which is why I decided to invest time in Hal myself.

And, thanks for the posts of conversations!  They're fascinating, and helpful.