Some people think that Simple AGI is impossible. In order to create a simple
AGI project which would not turn out to be much more than a few narrow AI
methods I need to create a highly constrained testing environment which would
have some potential for elementary complexity. So the solutions to problems,
for example, could not (all) be found with a mathematical range that can be
plugged into an evaluation method of some sort. Some of the complexity of the
test environment has to hidden, some hidden and only revealed through inquiry,
some of the complexity has to be deduced from what the program learns when new
insights are integrated into previously learned knowledge. So I really think a
simple AGI project is feasible as long as the data environment presents
different kinds of complexity which are not (all) solved by simple processes.
A solution to a problem, or the ability to understand a problem will
(typically) require a structured solution that can fit into acquired knowledge
in different ways. A structured solution is more complex than a simple
deduction (for another example). While structures can be fit with some
shaping, the shaping should itself lead to other structures that would fit to
explain the shaping. Reason based reasoning, another thing that I think is
necessary for AGI, is an imperfect art because reasons can be good or not good.
However, if a number of structured reasons are found to be deficient in some
way, while others are found to be good, some commonalities can be identified.
These commonalities are not (usually) going to stand as absolute indicators but
they may help in the shaping process.
What I have been saying recently is that while fundamental Neural Networks and
Bayesian Networks have been shown to be capable of learning to identify certain
characteristics in the data that are presented to it, the simple
implementations of the various kinds of network paradigms are not very
efficient and they do not scale or combine well. I am thinking that if the AGI
program had a little more control over the ways that it learns to identify the
characteristics of different problems, then it should be able to represent the
algorithms that it would use to recognize them more efficiently. This could
lead to a slightly greater range of scalability. For simple problems this
might be more inefficient than a simple Bayesian Net but I am thinking that
once the program discovers some effective procedures it should scale more
efficiently. I will probably need to create some algorithms which can simplify
or generalize a group of recognizer methods that the program would create as it
learned to recognize some kind of situation. I have been thinking of some the
advances made with Bayesian Networks which use supervised learning to train
them. I might start off with some kind of supervised learning to see if I can
demonstrate and develop the potential that I believe my ideas have. Then I
would try to develop the ideas to work in various other situations as well. Jim
Bromer
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
