I was trying to set up a simple ground rule case for a partial simulation of AGI. As I started by thinking of the simplest case I could imagine and I found that it was a little more interesting than I thought it would be.
I realized that the old numerical range could be used to test some important ideas. The idea is that a set of narrow AI implementations could be used to develop and test multiple possibility indexing. Suppose that the program has learned various responses to thousands of situations. These responses may be weighted for variations in kinds of cases. So in a typical situation the program might detect hundreds of different characteristics (in the observable input environment) that it had learned to associate with some response so it would have to find strategies to choose the best responses for the situation and strategies to learn from the experience. Even if the situation - which is input to the simple AI program - is made of distinct components, the possible combinations that might be relevant to finding a good response could be very complex. For example, if it learned abc represented a situation that it should respond to (with some kind of response), it might wonder if abxc was a variant of that situation. The problem here is managing and indexing of multiple possible responses that were reasonable for a particular component-of-a-situation when the typical situation might consist of hundreds of situation components. So even a very simple AI simulation might lead to some interesting results. (Of course there would have to be some way to evaluate the responses so the simple simulation would either have to be tied to some game or it would have to be worked out carefully). 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
