The problem with computing with quantities instead of relationships is that it is unidirectional, whereas there is plenty of biological and mathematical evidence that AGI absolutely MUST be bidirectional in nature.
Of course it is possible to use analog computational methods (like our biological brains apparently do), but that doesn't work well with present-day computers. I believe a better alternative is to compute with relationships - basically with an algebra engine, where the entire set of solutions to many simultaneous relationships is sought. The general case of this is obviously impossible, but by casting out small terms it appears to become tractable, though some of the individual solutions in the set of solutions may be lost - a little like losing track of some of the potential alternative universes. There have been a number of programs written that were capable of manipulating algebraic formulas, and some to solve isolated equations. There have also been programs to solve simultaneous *linear* equations via array operations. In the engine I envision, every "formula" would be presumed to equal zero, and hence would be an equation, despite its missing = sign, similar to array operator implementations to solving simultaneous linear equations. Instead of operators like "addition" or "subtraction", there would be new operators like "union" that would solve a system of simultaneous NONlinear equations, etc. WAY down at the VERY bottom, akin to the atomic "Load Address" operation on IBM mainframes that harnasses its advanced address computation system, would be an "evaluate" operator that would plug in the present values of the various independent variables, to find a solution to an identified argument variable. Which solution was found might be indicated by an operator, e.g. maximum, minimum, mode, etc. The actual additions, subtractions, etc., would all be buried down inside of "evaluate". This moves the job of "learning" to where it belongs, adjusting relationships to include observations, and does so in the purest and most manipulable form. "Programming" would be almost trivial in such an environment, with individual array operators then having the power to describe entire logical layers and columns in a self-organizing system. Sure, algebra is slower than arithmetic, but it regains the speed advantage and more by eliminating the need to numerically recompute every time slice. I strongly suspect that once such a "virtual algebra machine" has been implemented in software, it won't take much to move that into hardware that lacks the Von Neumann constraints of software, whereupon practical AGIs may soon. become possible. I would think that Ben, et al, would simply LOVE to program in such a highly powered environment, where most of their presently challenging "details" are taken care of "by electric", without the necessity of programmer attention beyond the initial programming of the algebra engine. Apparently they just haven't yet grokked this concept. I await their enlightenment, as probably indicated by carefully attacking the details of this proposal. Steve ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
