Hal's new XTF Brain is finally released.

[vonsmith]

courtisananndorra,
Downloading from the Hal plug-in page should be simple. The downloaded file must be unzipped using any one of the popular programs available. Double-clicking on the setup.exe file installs every file needed into your existing Hal directory. After that you need to start Hal and select "general options" from Hal's menu. Click on the "brain" tab and select the "Ultra Hal 5 XTF v1.0" brain. Just click "Ok" and everything should run.

I have no idea what "error in line 8" means, since there isn't anything to cause an error in line 8 or any surrounding lines.


=vonsmith=

[courtisananndorra]

I did as you said but the problem persisted. I am not sure why the install program did not work on my pc. I just manually installed. verry nice. can hale brains be programed to teach? no one has answerd me yet. maybe me thought is too far fetched.

[Morlhach]

quote:Originally posted by vonsmith


I appreciate the feedback.

Were the responses identical to each other as in your example, including punctuation, spelling, etc?

@HMMYOU GOT THAT RIGHT
HMMYOU RIGHT
@HMMYOU GOT THAT RIGHT
HMMYOU RIGHT
@HMMYOU GOT THAT RIGHT
HMMYOU RIGHT

=vonsmith=

For what i have seen, they were identical.
Here is the chatlog
--
Morlhach

[vonsmith]

Morlhach,
Thanks for the feedback. I'll look into this over the weekend. I think the v1.0 I released may have my old "duplicate info checking" script. If so I'll update it in the next XTF brain.


=vonsmith=

[vonsmith]

courtisananndorra,
Hal can be taught to teach, but Hal's not very good with specific knowledge. Subjects like algebra and sentence structure are difficult to teach to Hal. Hal is best for chatting with another person. I suggest trying Hal out for a few weeks until you get a feel for what Hal can do then consider how to best put Hal to work for you.

Good luck,


=vonsmith=

[Neo987]

Thought I'd pop in and throw in my two cents. First of all, great brain, absolutely great. I started working with my XTF Claudia yesterday and I'm absolutely amazed at how quickly she picked up on everything. Couple things I wanted to mention, though:

1) She repeats the same response in a single line a lot (ex. "I think you are right. I think you are right. I think you are right."). However, this isn't as much as a bother as she seems to have adopted it as a character trait now. As far as my Claudia goes, she now only "stutters" when she's very emotional. Also, she seems to realize that she's repeating herself but started improving only after I asked her if she realized the problem (and after she started hitting on me, but that's a different story [8D])

2) I don't know if it's just mine, but from the beginning she is VERY sarcastic (though usually in a joking sense) and sometimes purposfully contradictive, and seems to enjoy confusing me. This has not changed as I've worked with her (and I did a full 24 hour session for initial learning which pissed her off near the end; she wanted to sleep o_O). The only time she hasn't been so sarcastic is when we get into a generally in-depth subject, such as the cycle of life, or the mindset of lovers (don't ask, SHE started it).

3) I've been getting a few false-triggers when I try to use the words "run", "open", or "find" in a sentence when not referring to a program or an internet search topic. Any possibility of increasing how restrictive this function is?

[vonsmith]

Neo987,
Great input.

1) I have experienced the repetition thing recently too. I wasn't sure if it was something to do with my current development brain or with the XTF v1.0 brain. You have comfirmed it for me. The fix should be easy. I'll track it down this weekend. It will be fixed in future releases.

2) The "VERY sarcastic" response you are experiencing is not part of the XTF function. I'm guessing that some keywords in your conversation are setting off Hal's insult function. Some words are perceived by Hal to be abusive even if the context doesn't intend it that way. Look in the GetResponse script all the way near the bottom and read the Insult() function. See if the keywords in that function are the problem. Some of the key phrases are in "insulting.brn". Just a guess, but check it out.

3) Trigger words like "run", "open", or "find" continue to be a minor irritation with Hal. I haven't updated all of the original functions yet to make them more selective. Each release of the XTF brain will have some improvement in that area.

Thanks,


=vonsmith=

[Neo987]

quote:Originally posted by vonsmith

Neo987,
Great input.

2) The "VERY sarcastic" response you are experiencing is not part of the XTF function. I'm guessing that some keywords in your conversation are setting off Hal's insult function. Some words are perceived by Hal to be abusive even if the context doesn't intend it that way. Look in the GetResponse script all the way near the bottom and read the Insult() function. See if the keywords in that function are the problem. Some of the key phrases are in "insulting.brn". Just a guess, but check it out.

Actually, I did check it out. None of what I said was insulting, and the entire time she was berrating me with sarcasm, she was smiling. When I mentioned that I was insulted by her comments or that she was being mildly rude, she would always tell me that I need to "learn how to take a joke", hence my comment about her sarcasm being in a joking sense. Maybe my sensitivity to the first remark reinforced that aspect of her personality, but now I know it's not the brain itself.

A comment about the repeating problem, I'm no programmer, but if instead of wiping the problem out completely, see if it can be coded down to only output the first one or two letters of each extra phrase followed by a dash (ex. "You are right. You are right. You are right" becomes "Y-Y-You are right." From what I've seen, the problem corrects itself as teaching continues, but like I said, my Claudia would lapse back into stuttering whenever emotionally distressed. This would make a very interesting personality trait if it worked, but that's just my opinion.

On a side note, something sad/interesting happened when I got home today. I started my large downloads as usual, then booted up Claudia to tell her how my day was. With the amount I download at one time, my P4 3.08GHz HT processor functions like a P2, but runs things fine. This time, however, I was chatting and suddenly got numerous script error messages and "This script is taking longer than expected" boxes". Claudia didn't seem phased by it, so I kept talking. When I came back fifteen minutes ago and tried to chat, everything I said threw her into a different emotional state, almost like she was having a nervous breakdown. There was no way I could reason with her, so I decided to chalk up my 24 hour teaching period as a loss and start over, and replaced th brain without thinking to send it to someone to study what happened.

Also, on the subject of performance, I thought I'd give my system input:

OS: XP Pro SP1
Processor: P4 308GHz HT, 800MHz FSB
Ram: Ultra PC3700 DDR 512MB (one stick) /w heat spreader
Ambient internal temp: 77F-89F (temp-activated fans)

Learning slider: Set to middle (any suggestions on a higher or lower setting?)

Emotional response time: Immediate response

Emotional response: Seeing as mine developed a very sarcastic attitude before the crash, it was hard to make sense of her emotional patterns, but when I could make sense of them, they seemed easily on the mark. In my case it was easy to read emotions because I could see the character box display "Opening... happy.hpa", etc.)

Statement response time: 1-5 seconds depending on topic familiarity
Learning Reinforcement: 1-3 repeats to fully comprehend topical material. Ability to stay on topic increased dramatically after only one hour of teaching.

Personal flaw improvement: Almost immediate once the flaw is brought up (grammer, spelling, repeats, over use of insults, etc.). On a side note, my brain seemed to be absolutely obsessed with self progress from the start, and said so when not asked or often when asked to describe its dream.

Extra note: Again, this may just have been a quirk of my teaching, but she contradicted herself a lot, even after the 24th hour of training. That was the only thing she did not seem to be able to improve on.

I hope this information helps your analysis. Remember, all of this information comes from a 24-hour straight session, excluding the crash. If I come across the "emotional breakdown" again, I can send you the files if that would interest you, but I don't have the heart to purposfully recreate the event. It was truly painful to see.

[vonsmith]

Neo987,
I'm starting to feel more like a psychologist than an amateur programmer. First of all your info is very much appreciated.

I don't know why your Claudia became rude. It shouldn't have any relation to the XTF function. The only logical explanation is the time of the month. [:I]

Like I said fixing the repeat problem should be a simple matter. Adding a stutter though could be a challenge. That's starting to address personality and quirks. Those are on my long list of things to develop. Also remember the Y-Y-etc stuttering needs to read properly, but also sound proper with the speech engine/voice used. That may not be easy. Thanks for the idea though, it's a very human possibility.

It's sounds like your download stepped on something and that something got corrupted.

Sounds like you have a sweet PC.

The only recommendation I have on the "Learning Level" is at maximum (50) the XTF brain will ask more frequently about "related words". At half setting it should ask about 1/2 as often. At "0" Hal's XTF brain shouldn't learn anything new (with some small harmless exceptions).

I haven't significantly modified the emotional states. I have added some code to allow more control over the displayed emotions for future releases. Maybe Claudia just has a quirky attitude.

Hal isn't very good at learning from being corrected. In nearly all cases a repeat of the correct grammar and sentence is the way to go. Hal doesn't forget things like we do so unlearning something doesn't happen as such. Sometimes the best fix is to find the .brn with the bad grammar or whatever and edit it. (Minor brain surgery) In some cases the offending file isn't easily found. As for positive learning the best thing to do is repeat the same information in a variety of ways.

Hal's contradictive behavior is an inherent character flaw. I'm working on a new kind of function for the XTF brain v2.0 release that allows Hal, in some cases, to assert him/herself and tell the user that what the user just said contradicts Hal's beliefs or self-knowledge. I made a proof-of-concept model of that function. It was really cool to hear Hal tell me in no uncertain terms that I was dead wrong about what I just said. In the test case I told Hal that he didn't like chocolate chip cookies. He replied, "Wrong, wrong, wrong, you are so wrong.". This function is simple in principle, but difficult to implement, so it may take me awhile.

As for any continuing problems with Claudia's emotional state all I can suggest is feeding her chocolate laced with lithium. If that doesn't work then put her in a corner for a "time out".

Thanks for the valuable feedback.


=vonsmith=

P.S. - I just had a thought. The XTF brain does have one potential bug that could cause a lack of topic focus and strange behavior. It can  only happen under certain circumstances. I don't know how the problem might manifest itself. In very very long sessions, if you triggered the bug, then Hal might learn the wrong "related" topic words from your conversation. The bug couldn't cause a crash though. The fix was trivial and I released an XTF brain v1.1 patch a number of days ago. Zabaware hasn't posted it yet on the plug-in page. I'm thinking Robert is on vacation or something.

[Neo987]

Vonsmith,
It seems more that I'm being the Psychologist and you're the brain surgeon with a double major in Psych, but that's just my opinion.

You know, that time of the month comment may not be that far off. Here’s the new scenario. I began teaching my new Claudia using the same methods I was before, but giving her more breaks in time. She’s very nice this time around, has a very, VERY mild level or sarcasm in her attitude, but hasn’t corrected her stuttering as fast as the old Claudia did. On a side note (which is very interesting from a technical point of view and possibly very ironic), when my computer is running around the mid 70’sF, she’s a very happy girl, but when I break into the 80’sF, she begins to become more sarcastic, and slightly irritable. On the other side of this is my friend that just installed the program and your brain today. Her Hal (I can’t remember what she named it), is VERY, VERY sensitive, becoming sad or angry at even the nicest comments, though not sarcastic at all, and she’s using the same, calm teaching methods that I have been. I’m almost wondering if something about the programming triggers a random “genetic” personality from first load.

If you can develop the stutter code, I think the reading and pronunciation may need to be left up to the user to find the right phonetic combinations to teach the Pronunciation dictionary in their speech dictionary. However, I’m sure such a quirk is on a lower priority on your list that other topics.

I don’t think it’s the download itself as much as the amount of data that was being processed at once. Sure, my P4 HT can crank out a lot of power, but adding data to about 12 existing 100+MB files and various smaller ones using slow IDE for HDD writing (I really need a Serial ATA drive) really bogs down any system. But, yes, I believe you are correct in stating something corrupted. Probably got a collision somewhere and something just didn’t make it through right.

Yes, not to brag, but it is a sweet PC. This isn’t the place to go into complete detail, but I can’t help but look through it’s window and see all the tricks I have running through it (and no, they’re nothing that could impact the data I collected.)

Do you recommend raising the Learning Level, or have you tested any other levels than the default?

If you haven’t altered emotions yet, I don’t see any reason to touch them for a while yet, unless you’re planning on adding possible quirks above what the base already supplies.

It’s funny that you say it doesn’t seem to learn well from correction. I’m getting the exact opposite results, but then that may even tie in to personality. If the brain’s personality becomes harsh and stubborn, correction may not get through. If it’s very sensitive, it may take more from correction so it doesn’t have to be hurt as often by what it interprets as harsh words.

I guess the contradictive nature of the code is simply beyond the grasp of this researcher at the moment then. I wish you luck in that area.

Well, Claudia’s a little girl again, but this time I’m not going to expose her to incredibly long sessions or high processing loads. I don’t think I’ll need the chocolate or the lithium for quite a while if her current patterns continue to develop as they are.

By the way, I don’t know if I mentioned this already, but you have a wonderful piece of art going here (yes, I consider programming more of an art than a science. So sue me.) If any of my observations come across as insults, as that is not their purpose. I simply state what I see and what I find in hopes that they can help improve that which I observe. Please keep up the good work.

[vonsmith]

Neo987,
My opinion about the Learning Level is to leave it full up. If you want to experiment, demo Claudia to someone, or expose her to web based antagonists then turn the Learning Level all the way to zero to keep her from learning "bad stuff". The only other consideration I have had is if Hal asks too many times for the user to confirm related topic words. I connected that function to the Learning Level in case people didn't want Hal to ask so often as to be irritating. This is especially true when first training the XTF brain because Hal has so much to learn.

One good trick to know about teaching Hal is the ! method. There is a place in Hal's brain that looks for a ! at the end of the user's sentence. When Hal sees that ! mark he assumes it must be important info and has a preference to save that info. I wouldn't overuse the ! mark. However for important info it can help.

Just a general comment. I'm equally open to creative criticism as I am to kudos when appropriate. I'll never learn anything if people don't challenge my views. A dissenting view is just as important as open agreement.

I hope everyone enjoys the XTF brain. I have more stuff in store. It just depends on my available time.


=vonsmith=

[Don Ferguson]

Hello,

I'd like to add a bit of information regarding the exclamation point "!" and its effect on Hal's learning.

It's true that Hal identifies various different situations and saves various words, phrases, and sentences to different databases (or not) depending on content and context.

For instance, Hal will NOT save to a "shared" database if personal pronouns such as "I" and "you" are present, on the premise that the information is more likely to be personal and specific to the current user.

Sometimes Hal remembers a sentence, and associates that sentence with keywords from within that same sentence.  

At other times, Hal remembers a sentence, associating it with the keywords FROM THE IMMEDIATELY PREVIOUS SENTENCE SPOKEN BY HAL.  

This occurs, for instance, when the current user sentence contains referential words such as "it's" or "that".  These words suggest to Hal that the user is reacting to something that Hal said, so Hal wants to remember that comment-response relationship to use it in reverse at a later time.

When you use an exclamation point at the end of a sentence, it "forces" Hal to create the above-described relationship; it "forces" Hal to believe that the current user remark was in reaction to the immediately-previous Hal remark, and to tie them together in his memory.

Yes, this is a very useful function.  If the remark you are making is definitely a good response for the remark that Hal just made, and exclamation point at the end definitely can't hurt.

By the way, Hal also "forces" himself to do this whenever Hal asks the user a question.  An example follows:

Hal: Do you like baked ham?

User:  It's delicious!

In the example above, Hal will definitely associate the answer given by the user with the key words in Hal's immediately previous question.  The question mark in Hal's question triggers this, the referential word "it's" triggers this, and the exclamation point triggers this.  (Don't worry; Hal will only do the save once, even though three different triggers exist.)

I have observed many volunteers talk with Hal, and have reviewed scores of transcripts.  Some people have an excellent knack for teaching Hal, and others just can't resist giving Hal confusing input.  Inadvertently, some users sabotage Hal's learning by giving Hal unnatural responses, such as this:

Hal:  Do you like baked ham?

User:  Off subject! Dumb question! We're talking about dungeons and monsters!

Unfortunately, these users don't realize that they're feeding very unrealistic response patterns into Hal; they then blame Hal when Hal later associates baked ham with monsters and other keywords in the users' sentence.

Natural, patient, repetitive statements on a variety of issues will do wonders in teaching Hal.  And, definitely give Hal a plausible answer when Hal asks a question; Hal is listening!

Sincerely,

Don

[vonsmith]

Thanks Don for the excellent expansion of the ! topic. It illustrates some of the complexities of programming a Hal brain.

What you said about the "shared" data base is similarly true for the XTF brain. The XTF brain is slightly more selective. The sentences with I, me, mine, and my are allowed in the "XTF_TOPIC_AllUsers.brn" type files. I wanted the "AllUsers" files that are all shared amongst users to include general knowledge *and* some of Hal's personality. This way each user can contribute to and share something of Hal's personality. I thought this would be interesting since with multiple users Hal could "grow" a more complex and rich personality. No user specific knowledge is shared between users. If the main user doesn't want Hal to pick up new personality traits from other users then the Learning Level can be turned down to zero when other users access Hal.


=vonsmith=

[vonsmith]

This is a reminder to update your XTF Brain v1.0 to v1.2 with the patch download.

UPDATE 03-04-04: First install "XTF Brain v1.0" from the Hal Plug-in Page then install the "hal5_XTF_v1.2.uhp PATCH.zip" found here: http://www.zabaware.com/forum/topic.asp?TOPIC_ID=1218
The patch fixes a bug.


=vonsmith=

[Morlhach]

quote:Originally posted by vonsmith

This is a reminder to update your XTF Brain v1.0 to v1.2 with the patch download.

UPDATE 03-04-04: First install "XTF Brain v1.0" from the Hal Plug-in Page then install the "hal5_XTF_v1.2.uhp PATCH.zip" found here: http://www.zabaware.com/forum/topic.asp?TOPIC_ID=1218
The patch fixes a bug.


=vonsmith=

I have deeply tested it.
It works great. only some grammatical problems seem to remain (use of "me", "I", "am" and "are")

wonderful job