Re: Routing and AI algorthmic efficiency

From: Anders Sandberg (
Date: Tue Feb 12 2002 - 03:21:27 MST

On Mon, Feb 11, 2002 at 07:58:12PM -0500, Dan Clemmensen wrote:
> This leads to two questions:
> 1) Is any part of the AI problem isomorphic to the routing problem?
> 2) What parts of the AI problem, if any, are amenable to brute-force
> over-engineering, and what is a good guesstimate of the "cleverness
> factor" of a clever algorithm versus a brute-force algorithm? The
> cleverness factor for routing is a simple factor of 2.

This is a very interesting set of questions. As a computational
neuroscientist, I think I can answer them for the brain (which after all
is an existence proof of a hardware implemented AI). In the brain
routing does not seem to be very common - there are some ideas how you
could do it, but most signals are simply sent along firmwired paths
where adaptation (i.e. learning) changes where they lead on longer
timescales than the signals are themselves moving. This seems to be a
brute force solution: it grows many more axons and synapses than are
needed, and then the unnecessary are pruned or ignored.

Now, if I was to implement this system on a parallel computer
instead, routing immediately becomes more relevant. If I store different
parts of the network on different nodes, then routing signals
efficiently can affect my performance a lot. The tricky part is that the
brain seem to be very modular on some levels (lobe interconnections,
cortical minicolumns) and highly connected on others (intercortical
connections, inside minicolumns), making division rather hard.

The general AI problem doesn't seem to me to be inherently about
routing, but any implementation beyond a dedicated hardware solution
subjected to scarce computing resources would likely have to deal with
routing to a very high degree. It would mainly be getting data to and
from processors, or accessing remote data/states. Clever routing would
improve the speed but not the smarts, and if cleverness can only help a
factor of 2 then brute force would be the way to go.

Anders Sandberg                                      Towards Ascension!                  
GCS/M/S/O d++ -p+ c++++ !l u+ e++ m++ s+/+ n--- h+/* f+ g+ w++ t+ r+ !y

This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 13:37:39 MST