[sunama]

So I have been pondering on making my own learning chat bot. I stumbled across ultra hal today and installed the trial version.

I have been trying to teach Ultra Hal a simple fact:

my eyes are green.

I then proceed to ask him 'what colour are my eyes?'

Unfortunately, Ultra Hal has not yet been able to answer this question.

I've been through the teaching and questioning cycle at least 10 times and Ultra Hal is unable to learn that my eyes are green.

Is there a trick to getting Ultra Hal to learn facts and then answer questions regarding those facts?

[Freddy]

Try asking over on the Zabaware forum, there's people there with a few methods for teaching Hal.  I've always thought Hal was a bit of a thicky out of the box.  But they seem to get results over there okay.  Usually though if you say something enough he eventually catches on, so I am surprised you got nowhere yet.

[sunama]

Hi,

Yes I put up a similar post in the zabaware forum this morning and someone did reply. However, the reply wasnt particularly inspiring.

The basic jist of it was that apparently Hal does learn, but he doesnt respond to enforced learning, eg. I tell it a fact and then ask him to repeat the fact. Apparently, it works best when I give him a few facts and eventually, HAL, will respond during a conversation, with a fact that I taught him earlier.

If this really is the level of intelligence HAL has to offer, then I can't say I am impressed.

The whole premise of a learning AI is that it should be able to respond to enforced learing and be able to regurgitate information taught.

I was thinking of writing my own chat bot/AI program which will allow for enforced learning, as well as learning from the internet. And once the AI has read a page from the internet, it will have stored keywords in its memory for immediate recall.

So, if I did say to it that my eyes are green, it would relate the keywords: <user> <eye colour> <green>. And the association would be especially strong when I have mentioned this fact in only the previous line. And of course the more times I state this fact, the more likely it is that AI will relate the keywords to eachother.

I'm VERY surprised that UltraHAL was able to win the Loebner prize, unless of course there is a secret that I am not in on.

[Freddy]

Yes teaching UltraHal is a painstaking and lengthy process.  I personally don't have the time to work with it much any more. You need to do a lot of chatting to get results, but you also need to have the patience of a Saint too I think.

I guess it is surprising that Ultra Hal won the Loebner prize, at first glance.  But that level of the prize is much less strict than a proper Turing Test.  The bronze level award which all of these bots were competing for only awards the most human-like computer.  It doesn't have to fool the judge entirely or even at all.  The gold prize as you can imagine is more strict and no one has ever won that. And of course there are those that don't think the Turing Test is relevant anyway...

[Art]

Keep in mind that Robert also wrote and provided a special "Enhanced" brain for
that competition...one. shall we say, more geared toward that type of situation.

A bot (if you can call it that) that might be more well suited for question / answer
types would be Answerpad. There are several bots out there that do parrot back
learned info rather quickly. The list is, however, rather large.

Either way, good luck!!

[Freddy]

Yes good point there Art - it was a version made especially for the competition..

[one]

Hi,

Yes I put up a similar post in the zabaware forum this morning and someone did reply. However, the reply wasnt particularly inspiring.

The basic jist of it was that apparently Hal does learn, but he doesnt respond to enforced learning, eg. I tell it a fact and then ask him to repeat the fact. Apparently, it works best when I give him a few facts and eventually, HAL, will respond during a conversation, with a fact that I taught him earlier.

If this really is the level of intelligence HAL has to offer, then I can't say I am impressed.

The whole premise of a learning AI is that it should be able to respond to enforced learing and be able to regurgitate information taught.

I was thinking of writing my own chat bot/AI program which will allow for enforced learning, as well as learning from the internet. And once the AI has read a page from the internet, it will have stored keywords in its memory for immediate recall.

So, if I did say to it that my eyes are green, it would relate the keywords: <user> <eye colour> <green>. And the association would be especially strong when I have mentioned this fact in only the previous line. And of course the more times I state this fact, the more likely it is that AI will relate the keywords to eachother.

I'm VERY surprised that UltraHAL was able to win the Loebner prize, unless of course there is a secret that I am not in on.

Hey.
Did it ever occur to you that enforced learning also has to be 'taught'?
I usually told my Assistant "Positive re-enforcement is always good" and she learned the many meanings of this statement.

GL!
J.

[RD]

Just a thought, are you using the UK English colour? 
Not sure if you have explored the mod's to help Hal learn or his built in learning patterns, but there are shortcuts to improve his conversations.
IMHO Hal is the best of the bot's after alittle work with clip board and AIML with out paying the $1,000 usd for that one completed version of Alice.

[Freddy]

Well spotted RD, that sounds like it could be the problem...