Ai Dreams Forum

Member's Experiments & Projects => General Project Discussion => Topic started by: ranch vermin on May 11, 2015, 09:13:57 am

Title: program developing off a memory (simple and well known knowledge)
Post by: ranch vermin on May 11, 2015, 09:13:57 am
Its a program, developing new fields off the old fields. a ga. and its the craziest thing.
Watson doesnt do it,  it just searches with whats there with a program.
Title: Re: program developing off a memory (simple and well known knowledge)
Post by: Art on May 13, 2015, 02:40:59 pm
SO you just threw that statement out there to keep us guessing or what? Come on...details....
Title: Re: program developing off a memory (simple and well known knowledge)
Post by: ranch vermin on May 13, 2015, 05:54:43 pm
I was just stating a fact.

That watson doesnt generate anything when it searches.

A genetic algorythm searches for old fields then through a method transforming them makes new motor fields, then rinse and repeat over again.

But its limited, and its tricky to make something like a huntsman spider, because I dont think it would ever be as agile, at all. 

If you wanted to treat it as the winning solution,   then im afraid you cant temporally search far enough into the future for a positive reinforcement, because it will take 1000000 parallel instances of your markov chain over and over again,  and negative reinforcement is tricky because its confused in statistics if it happens or not.    And if you make it too careful, then you make it more unco-ordinated, in it being too careful.

And then your stuck with something quite primitive, poor generalization of the moments that cause these happenings of reinforcement, they are taken as brittle states to be compared to, as a whole, and it doesnt know what of that happening is to be ignored, that didnt matter to the moment.  (like if a cube was there, then it suddenly starts thinking its lucky, if the trigger was placed when the cube wasnt there.)

It finds it hard to know what a smile means,  even if you categorize it perfectly,  what does it have to do with success and failure.  and thats the secret of intelligence, if this was the way.

Everything in the eye sensor shows a strict raw data resemblance to itself, (in the end after a lot of work)  But language and this "smile" thing, have no raw data resemblance to become synonymous, to be treated the same in a way.
Title: Re: program developing off a memory (simple and well known knowledge)
Post by: Art on May 13, 2015, 10:27:08 pm
Perhaps it's because Watson is searching for the Truth...a Fact...the basis behind that with which it has been tasked at searching.
If a human were given a similar task, then said human should not add to nor subtract anything that might slant or deviate from the hard facts of the task. (hey...media moguls, newscasters and reporters...take note here!! You should try it sometime!..the way news used to be...non-biased!).

But I digress. What else would one (or Watson) be expected to make or add to the search in question?

If Watson were tasked (assuming it was actually capable), of creating / writing a short story or novel, then yes, I'd certainly expect it's programming to be able to draw upon a vast array of knowledge upon with to set ink to paper and create something...well, in this case...different, to be sure.

And no...I never bought into that premise of the monkey and a typewriter producing the next "Best Seller".  ;)

Maybe William Shakespeare said it best, "To search, perchance to find..."
Or something like that...maybe not. :)
Title: Re: program developing off a memory (simple and well known knowledge)
Post by: Don Patrick on May 14, 2015, 08:33:33 am
Art is right. Watson was simply not tasked to create anything. With some modification however the same underlying techniques could be used to mix and merge 1000 similar sentences from its database to create a new average sentence. It would need some grammatical enforcement and it wouldn't know what it had written, but it could produce the result with similar techniques as it now uses to search.

The downside of both Watson and genetic algorithms is mainly that they only get enforcement from analysing a huge amount of data. Watson can't operate with a small knowledge database, and genetic algorithms can but take their sweet time getting haphazzard results. And one of the bigger problems of AI is that common knowledge for interacting with the real world is hard to come by, nearly undocumented, let alone in any statistical size.

My money is on inference, which can derive much from little knowledge, optionally enhanced with some trial-and-error and statistical enforcement.
Title: Re: program developing off a memory (simple and well known knowledge)
Post by: ranch vermin on May 14, 2015, 02:31:59 pm
I dont know how good watson is, im just speculating, making an average sentence sounds interesting.