Re: [sl4] A model of RSI

From: Bryan Bishop (
Date: Wed Sep 24 2008 - 14:53:51 MDT

On Wednesday 17 September 2008, Stuart Armstrong wrote:
> > So how could RSI be defined in a meaningful way?
> By engineering, not mathematics. So a RSI has to be a statement about
> the actual architecture of a program, not about the equivalent Turing
> machine.

I can think of two types of architectures of RSI systems.

(1) Soft RSI. No modification will make a step backwards. However,
running an s-RSI for some finite amount of time will not give you any
guarantees about overall improvement. In this architecture, you can
consider the typical modes of evolutionary search and optimization
being the classic exploitation-versus-exploration. Exploration for the
cool-down cycles, and the exploitation for rapidly bit banging the hell
out of improvement nuggets.

(2) Hard RSI. All modifications make a step forward. This is the one
where everything is magical, the phenotype/genotype problem doesn't
exist and computer scientists can finally sleep at night (the real
night, not the 4 am to 2 pm "bright night").

- Bryan
Engineers: #hplusroadmap

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:03 MDT