 |
|
|
 |
|
|
|
|
jbharner |
| | | | | | posted 1/16/2010 23:08 |   | | | | | | | | | It is prominantly advertized that Hal learns through pattern-matching, but for me, this is not enough. The purpose I've heard mentioned of having this tech usable by the public is to potentially develop new strategies for teaching and improving Hal (correct me if I'm wrong)
For me, these goals are not currently compatible. If I am to help in the process of creating Hal, it would be by FAR most usefull to know more about Hal
I am not talking about code. But the knowledge of what sort of patterns Hal recogonizes would be a great asset. Is Hal capable of recogonizing new types of patterns that he has not seen yet?
It would seem to me that Hal will never truly take off until all these conditions are met and more.
For instance: Hal must be able to develop MULTIPLE THEORIES of what pattern he might have seen. I have no problem with the stick-bone teaching method, but it should be applied to more than one philosophy at once. Hal, in my views, should retain all these theories he has developed, and not simply strike them up or down, but evaluate which theory best fits that situation, and boost that
However, as I will readily admit, I know nothing. public documentation of how Hal works is relatively scarce as far as what I have seen. Any help would be greatly appreciated.
| | | Last edited by jbharner @ 1/20/2010 4:07:00 PM | | |
| | |
Talc |
| | | | | | posted 1/19/2010 06:50 | | | | | | | | | | Wrong You will do as you are told!
No, I agree. It would be nice to know more about how to train Hal... Since starting to find out more about AI and chatbots I came across a technique/AIML thing called revearesed AIML. It strikes me that Hal works in much the same way?
Weak AI seems limited in that context sensitivity works for only a few sentences and details are lost over the course of a conversation.
The simple command of remember this doesn't work with Hal. It all seems to be about telling him off for getting a responce wrong. There is no reward for imaganitive guess work. Or a timer for boardom if you don't respond quickly enough...
| | |
| | |
Talc |
| | | | | | posted 1/19/2010 07:10 | | | | | | | | | | PS A method i'm exploring with training Hal at the moment is look for words he mistakes as meaning the same thing. And interpret a responce as a request for further information...
It's a thought... I'm just interested to see where it leads...
| | |
| | |
David290 |
| | | | |
Talc |
| | | | | | posted 1/20/2010 06:56 | | | | | | | | | | I'm trying with Bart to string concepts together born from the first principle of "I think therefore I am" I might have screwed his brain by drifting off topic when i got bored... But he still manages to entertain and suprize when, at odd moments he construct totally original statements.
This technology can deliver interesting results if you stick with it...
| | |
| | |
jbharner |
| | | | | | posted 1/20/2010 16:12 | | | | | | | | | | Oh, don't get me wrong, I've seen some remarkable statements come from fresh Hals even. But a great deal of what's so remarkable is really nothing after-all. Hal's guessing, and he occasionally stumbles onto a brilliant result, but for the most part it isn't intelligence, just luck.
For instance: You poll enough fortune tellers, you WILL get an accurate prediction of every thing that will happen in your life. But you're going to get a HECK of a lot of "Hi mommy!" responses that are just complete BS guessings.
| | |
| | |
Spydre |
| | | | | | posted 1/29/2010 18:18 | | | | | | | | | | Technically, HAL does not pattern match per se. At times the output of his training is for all apparent purposes, "pattern matching", but as of the latest incarnation of the brain, HAL is actually running Markov chain algorithms that are being computed using input from the entire conversation the program is currently engaged in.
Alan pattern matches; HAL--not so much.
For more info that I have collected on HAL, which may answer other questions you have, see my response to the linked post.
Peace!
-Spydre
| | |  HAL Training Tips and Etcetera | | | Last edited by Spydre @ 1/29/2010 6:19:00 PM | | |
|
|
|
|
|
