In the spirit of AIXItl but more practical, see Juergen's new work on the "Goedel Machine" AGI architecture
http://www.idsia.ch/~juergen/goedelmachine.html I don't think this is really a practical AGI architecture, but I think it's really interesting ... I do like the direction this research program is moving in... " Abstract of 2003 paper: A Goedel machine solves general computational problems in a possibly stochastic and reactive environment. Its initial software includes an axiomatic description of (1) the Goedel machine's hardware, (2) known aspects of the environment, (3) goals and rewards to be achieved, (4) costs of actions and computations, (5) the initial software itself (no circularity involved here). It also includes a possibly sub-optimal initial problem-solving policy and a proof searcher searching the space of computable proof techniques, that is, programs whose outputs are proofs. Unlike previous approaches, the self-referential Goedel machine will rewrite any part of its software (including axioms and proof searcher) as soon as it has found a proof that this will improve its future performance. By definition, it produces optimal self-improvements, given arbitrary formalized problems and typically limited computational resources; its optimality notion is not restricted to the concept of asymptotic optimality. To initialize the proof searcher we may use the recent Optimal Ordered Problem Solver (OOPS). " My problem with this approach is that automated theorem-proving is very hard! Current AI systems cannot prove nontrivial theorems without human intervention at crucial decision points. I'm guessing that OOPS also has this property. The theorems that the Goedel machine must prove in order to do its self-improvement are not simple ones. A related complaint is that it may not be optimal to self-modify only when a provable improvement is found. It may be optimal to self-modify when a probabilistically-likely-to-be-helpful improvement is found, based on heuristic estimates that don't constitute a formal proof. Now, if this is the case, eventually a Goedel machine will discover this and modify itself to make appropriate heuristic probability estimates. But how will it get to this point, if it lacks the ability to control its inferences well enough to do complex proofs??? ;-) -- Ben ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
