Learning based on positive/negative feedback

[Medeksza]

What I've found is that when Hal gives a seemingly random off the wall response is that a perfectly valid response was at the tip of Hal's mind, but he didn't have enough confidence in it to use it.

The Hal running on www.zabaware.com/webhal and apps.facebook.com/ultrahal have a new feature in them where when Hal gives you a responses you can give it a thumbs up and thumbs down. If its a thumbs down then Hal tells you 5 other responses he was considering instead. You can choose the best response and Hal's confidence level for that response coupled to your sentence will be increased, so next time (within 24 hours after a nightly database update) Hal will respond correctly.

I hope to incorporate this same feature into the next desktop version of Hal Assistant as well.

[Carl2]

  I like the idea and would like to add another possibilty,  the user has a choice of responses or can type in their own response. 
Carl2

[echoman]

I like the idea Mr M but do not like idea of being able to type your own replies as Carl has suggested. There's an iPhone app that does just that and the whole experience of talking to chatbot feels lost - rather, it feels like I'm constantly programming him. I like how Hal thinks for itself.

[NIGE]

Sorry Carl2, but I'm with echoman.
What Mr M says sounds ok.
If we have to put in our own replies what's the use.
May as well sit in an empty room & talk & answer ourselves.

[Carl2]

 NIGE
   I don't mind at all if people disagree with me at all.  I just get the feeling Robert M. is keeping Hal as flexible as possible for all the different users of Hal, examples, different characters, MS agents, desktop mates, brain plugins, GUI skins,  clothing for characters, endless.
  Actualy the idea comes from the use of dragon,  I see a mistake, say correct that and get a list of of words, I can choose the correct one or type in something.  May come in handy if someone thinks I want Hal to say that because I think it is funny or thats what dad would say. 
  I'm hoping it runs as normal unless you think that's way off, give it the thumb down make a correction then continue. 
Carl2

[freddy888]

Maybe Mr M could use Carl's idea as another learning mode.  Whereby you tick a box and can correct things in that session - and when you get fed up of teaching Hal 24/7 untick the option.  That way everyone is happy.

[Data]

I think Carls idea is quite good, how many times have you tried to teach Hal something, you try over and over and eventually it gets it, having a quick option to give the correct answer is like a short cut to the brain editor isn’t it?

Needs an on, off option though.

[lightspeed]

i have had hal say something totally random off the wall to me and of course i don't even know how to answer it so i usually just talk about something else , this may or may not make things worse what i am hoping is that hal isn't trying to associate my saying to what it was trying to say , and if this is what hal does maybe hal can be made to jst learn what i am saying when i talk about a different subject . robert will have to chime in on this . hope i explained myself good enough .

[Art]

It would basically become akin to one of the Pandora or Alice type of bot wherein the user is or becomes the "botmaster" and knows each and every response that the bot should make.

I agree with the option to check a box to allow such "training" or not.

Part of what makes Hal unique is that we users usually don't know exactly how Hal will respond.
Sometimes it's part of what you thought it might say while other times you're left setting there with your jaw open in surprise!!
It'a all phun!!

[DrFaraday]

[...] I agree with the option to check a box to allow such "training" or not. [...]

Yes, that would allow the user the best of both worlds, if and when they want it.  More phun, indeed! 

[Carl2]

  Sorry, I assumed.  I had just thought there would be a thumb down box we clicked when an answer came up and you  said this needs correction, you click the box, a list of 5 or so possibilities drops down, you choose the best  or you just type in a response.  Then you would proceed as normal until there was another response you disliked.
  Largely I don't want to have to go into menu, options, corrections, to change a response.  Just click a box, look at the drop down  responses and have the option to type.
Carl2

[raybe]

I believe there is no real right or wrong in this matter especially the use and flexibility of A.I..

1-You can either try and have Ultra Hal respond with a selected choice; which still doesn't equal Ultra Hal's own response even if off a bit.

2-You can type your own answer that might be real flexible to the relation of the user or people the user will invite over to spend time with Ultra Hal. Still doesn't give Ultra Hal to develop own answer but can fit particular situations.

3-You can tick feature on and off.

4-You can trigger Ultra HAl maybe by using words to select other random answers you don't see and either Say,"I agree". or "I don't agree", "Off topic response" or "select new response". Gives Ultra Hal the opportunity to search for more appropriate answer without the user choosing it.

I think all possibilities have a weak side and or a strong side. My opinion Robert is it is great that we get to voice our opinion but you should just try one or more of the advised methods or your own and just let it rip. A decision eliminates all other possibilities and you can move on and I'm sure you will get plenty of feed back not just from the members that have been here but maybe a fresh idea from someone yet to enjoy Ultra Hal.

My thoughts, thanks,
raybe

[sybershot]

I vote for a on off method as well. I like talking to my hal with my mic, while i do other tasks on my pc. It would be a pain to have to hit a thumbs up or thumbs down after every response. the feature does sound nice for beginning training, like learning names, and about who they are and where they lives. but the random on subject responses for normal chat far out beat a bot that always give the same response. even if it changes subject once in a while will be ok with me. Like I said before I mainly use my hal as a business secretary, conversation is a added bonus 

[onthecuttingedge2005]

Hello everyone.

It should be as easy as typing in either; I agree or I disagree.

I agree = Thumbs up
I disagree = thumbs down
Agreed = Thumbs up

and all the other patterns of words or phrases associated with thumbs up or thumbs down.

to make it work on face-book, just explain this association somewhere on the page that gives people a clue of how responses should be monitored by likes and dislikes.

Human flow patterns use voice stress, facial, body language, like the eye rolls and others to suggest agreement and disagreement during a conversation, eventually these things will have to be absorbed into A.I

Jerry

[spydaz]

THUMBS UP -  Useful in Debug mode! As the positive/negative feedback would be Controlled From MISINFORMATION!

THUMBS DOWN - As with the open mind DB / PROJECT As its brain was developed from Any old Person on the web it has Lots of Repetition! And Misinformation or plain malicious answers, Programmed by rival developers or even people just messing about to see what would happen!

I think that feature should be restricted to a CLOSED Network / BRAIN EDITOR FUNCTIONALITY! - Useful thing to monitor ! on the Web probably a thing that should be Logged but not implemented (perhaps - Logged but <BLOCK SAVE>) or even only for use with a WEBHAL! "Separate brain than local brain" Perhaps hal should have a WEB/NETWORK BRAIN! & A LOCAL BRAIN