Shane Legg wrote: > Why consider just one test problem? > > Doing so you will always be in danger of having a system that > isn't truly general and is built for one specific kind of a task. > > A better idea, I think, would be to test the system on *all* problems > that can be described in n bits or less (or use a large random sample > from this space). Then your system is gauranteed to be completely > general in a computational sense. > > Shane
Shane, Here's the thing... Let S be an intelligent system, and let n, rS and rT be integers Define int(S,n,rS,rT) as the average success of S at solving problems describable in exactly n bits, within rS space resources and rT time resources Then, realistically, we can expect that for each S, int(S,n,rS,T) will be -- generally decreasing as a function of n [though clearly the graph will be a little rugged] -- increasing in rS and rT The way the human brain seems to work is: * some of its architecture is oriented towards increasing the sum over n of int(S,n,rS,rT), where rS is given by brain capacity (enhanced by tools like writing?) and rT is different for different problems, but is a "humanly meaningful time scale." * some of its architecture is oriented towards increasing problem-solving ability for n that are so large that int(S,n,rS,rT) is totally miserable for realistic (rS, rT) For example, everything to do with vision processing and visual cognition pertains to an n that is so large that int(any human brain,n,human brain capacity,1 year) is effectively zero. The same goes for everything to do with human social cognition... human mathematical problem-solving (how good are we going to be at proving randomly generated theorems??) I conjecture that this property of the human brain is going to carry over to any realistic general intelligence, at least until we start talking about a tremendously posthuman computer based on currently inconceivable technology That is: any real-world-useful general intelligence is going to have a mix of "truly general intelligence methods" focused on boosting int(S,n,rS,rT) for small n, and "specialized intelligence methods" that are different from narrow-AI methods in that they specifically leverage a combination of * specialized heuristics * the small-n general intelligence methods Now, a slightly different way of cooking occurs to me... What if we define S(t) = the state of S at a given time E(t) = the state of the system's environment E at a given time We can then define int(S,n,rS,rT|E,t) as the average success of S at solving problems that are describable in exactly n bits *given the background information of S(t) and E(t) *, within rS space resources and rT time resources Now, we're asking not just whether S can solve simple problems, we're asking whether S can solve problems that are simple against the background of its environment and its own mind at a given point in time. In this case, I think the same conclusions as above will hold, but more weakly. I.e. the general intelligence capability will hold robustly for somewhat larger n. But still there will be the dichotomy between small-n general intelligence, and large-n specialized-building-on-general intelligence. Because only some problems relating to self and environment are useful for achieving system goals, and the system S will be specialized for solving *these* problems. Now, moving on, I'll make the following claim: ** achieving "small-n general intelligence" is a basically simple math problem ** In Novamente, we achieve it through using genetic programming and/or bayesian optimization and/or reinforcement learning, over a space of program dags (directed acyclic graphs). The program dags are expresses in terms of certain kinds of nodes and links in Novamente's overall knowledge network. Schmidhuber's OOPS system achieves the same thing in a different way... NARS achieves it using higher-order inference... Big whooptido!! I think that the hard problem of AGI is actually the other part: BUILDING A SYSTEM CAPABLE OF SUPPORTING SPECIALIZED INTELLIGENCES THAT COMBINE NARROW-AI HEURISTICS WITH SMALL-N GENERAL INTELLIGENCE (specialized-intelligence-on-top-of-general-intelligence, or SIOTOGI) I agree, Shane, that algorithmic information theory is useful for the "small-n general intelligence" part. But it's just providing a complicated, sometimes elegant mathematical formalism for what is actually one of the *easier* parts of the practical AGI task. A math theory that could help with the hard part would be a heck of a lot more valuable.... Right now, algorithmic information theory helps us where we don't need help. >From what I understand of James Rogers's AGI approach, he is moving toward SIOTOGI in a way that does specifically build on algorithmic information theory. So my critique may not hold for his variant of alg. info. theory, but it does hold e.g. for Hutter and Schmidhuber's published work, for classic Solomonoff induction, etc. -- Ben Goertzel ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
