[agi] Is "computation" a good concept for describing AI software?

[Ben Goertzel]

Hi all,   I'm away from home (at a bio conference in Australia) so my email access is sporadic, so it's a bad time for me to start list discussions, but I feel like it anyway, so here goes ;-)   I've been talking with Pei Wang about whether "computation" is a good concept for modeling or describing AI systems.  He says it's not.  I tend to agree with him.   Here is how I am thinking about the issue...    The computers on which we build our AI/AGI software are specific, finite physical systems, and the software programs we build are specific, finite sets of code instructions we feed to these specific physical systems.    Then, we can choose to *abstract* these physical systems by considering them as approximations to certain ideal mathematical systems.  For instance, a Turing machine is an ideal mathematical system, and a real computer can be considered as an approximation to a Turing machine (an approximation because a real computer doesn't really have an infinite memory tape).    However, there is a question whether this abstraction/approximation is of any use for studying intelligence.     What if we think about the "amount of memory that can be accessed per second" associated with a given hardware device H, and call this M(H).  Then if we place a practical bound on M(H), and take into account special relativity, we can no longer talk about ordinary computers being equivalent to Turing machines with infinite tapes, and we can no longer talk about bisimulation based equivalency between an arbitrary pair of real computers.      So the question is begged, how do the structures/dynamics required to achieve useful goal-achieving behavior using H depend on M(H).    Then one must argue that they do depend on M(H), in the sense that the useful-for-goal-achieving structures/dynamics for moderate M(H) [such as Novamente or NARS] are quite different from the useful-for-goal-achieving structures/dynamics for extremely large M(H) [such as AIXI or the Godel Machine].    So the problem is that the mapping "intelligent system = universal computer" relies on ignoring bounds on M(H), but it seems almost certain that the optimal ways of achieving useful goal-seeking behavior are strongly dependent on M(H).   (recall my definition of intelligence as "achieving complex goals in complex environments")   thoughts? reactions? insults? ben     To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]