Self Programming AI Structure

[Black Rose]

This is an idea i have: what if there were a computer program that creates new programs to solve problems. Now these problems would be composed as components that can be combined into larger structures. It would remember every problem it has ever solved and this would be added to its structure. To solve the combinatorial explosion problem it would compress what it knows into a hierarchy where as it only remembers what it has solved and forgets (prunes) the solutions that were not used. So where as it makes many combinations it has a meta evaluator. It combines many things it has learned into a finite set of plausible solutions and the meta evaluator by experience learned when selecting a particular solution how that selection turned out. It then evaluates all the sets of solutions it has generated and picks the best one. This solution is added to the memory that is hierarchically compressed and it forgets the other sets. The hierarchy is contains all that it has done so that it knows things such as what has happened and how patterns lead to other patterns. Its goal set is chosen by the ethical considerations such that it is taught by language. This could be done by downloading conversations on the web such that it learns how to be like the person by reading what they said and taking it as if the AI said what the user actual said in the web conversation. It would then learn programming in the same way by seeing the results of what it wrote on a 2D screen. It would start with a library of selected materials for it to read and programs it could write. It would then create new materials by having real conversations on the web such like a website where it could interact teaching people what it has learned. Signifiers would be used in its database so it could learn who knows what, this is found in the conversation "hi i am jeremy, lets see what the program does". All this is in a loop structure the creates new loops or incorporates small programs into the larger main one that changes how it programs depending on what it encounters.

This is my idea but i don't know programming, the AI would know programming from its learn materials. I know the program of this system would be simple, I have the flow chart map in my head. Please tell me if you know how it could be programmed in the way i described or if you have any ideas. I want to program it in python because its the easiest for me to understand but i don't know all the steps to take. If anyone could help me we could work on it together. Thanks

 

[computerhxr]

Have you seen this?

http://www.wsj.com/video/scientists...bot/EF23D2FF-8292-4F47-A8E8-089DAE76410A.html

 

[Aetherius Rimor]

You would enjoy this book: Programming the Human Biocomputer: John C. Lilly, Beverly A. Potter: Amazon.com: Books@@AMEPARAM@@http://ecx.images-amazon.com/images/I/51DbAlTFN-L.@@AMEPARAM@@51DbAlTFN-L

It talks about what you've described by what I can tell.

He refers to routines for self-reprogramming as Meta-programs, and even mentions Meta-meta-programs for reprogramming the reprogrammers (think stuff a therapist teaches a person to do in cognitive behavioral therapy).

 

[Architect]

This is how artificial neural nets work, except the node structure doesn't change. Which doesn't matter, it's all in the weights which the algorithm re-orders.

The human brain is the only example we have of self programming. We'll reverse engineer it before too long. There's some thought that it's a system of hierarchical hidden markov models, which makes complete sense to me and would be simultaneously mind blowing.

 

[Aetherius Rimor]

We'll reverse engineer it before too long. There's some thought that it's a system of hierarchical hidden markov models, which makes complete sense to me and would be simultaneously mind blowing.

This is the reason I want to go back to college for cognitive science, with the hopes of contributing to that effort.

Do you have any citations for the hidden Markov model idea by chance? Used them in my AI class, so it would be really interesting if we stumbled upon the algorithm our intelligence is based on without realizing it o_O

 

[Architect]

Do you have any citations for the hidden Markov model idea by chance? Used them in my AI class, so it would be really interesting if we stumbled upon the algorithm our intelligence is based on without realizing it

Kurzweil is working at Google on the idea. Read "How to Create a Mind". But it's "Hierarchical HMM's". Ordinary HMM's are used many places, especially in voice recognition.

 

[Aetherius Rimor]

Kurzweil is working at Google on the idea. Read "How to Create a Mind". But it's "Hierarchical HMM's". Ordinary HMM's are used many places, especially in voice recognition.

Hiearchical supports previous understanding of the way the brain processes data and identify's patterns.

Can't remember where I was reading it, but the gist was that all brain cells are identical; their function is determined by what they're connected to. And they are responsible for identifying patterns of increasingly larger scope the further the data gets processed. Example using sound, first the phoneme, then the word, then the sentence, etc.

Sounds like they're on the right track.