Re: Progams which can create efficient programs

From: Christian Szegedy (
Date: Fri Feb 25 2005 - 04:17:43 MST

Tennessee Leeuwenburg wrote:

> So, kind of. I don't think anyone has tried it though.

I think this is the most basic idea between any reasonably
general AI.
For example the universal search of Hutter (altough only a
theoretically feasible construct) utilizes this idea to give
an algorithm which is provably asymptotically optimal for
any given decision problem. However his algorithm is
infeasible in practice, since it does not use any clever
search strartegy.

I think that every algorithm that adapts itself dynamically
to its environment in some fashion can be viewed as a
program that generates a "new program": the parameter
set that is changed can be regarded as program too.

The LISP programming language has been developed
with this objective in mind. It has a simple internal structure
that can be processed by LISP programs. However LISP
has the disadvantage of statically not typed. Static typing
could reduce the search tree extremely efficiently. In fact,
I think that a good program-generator tool should be proceed
in a primal-dual way: both modifying a program and the
type-system (the constraints) analogously to the efficient
analytical optimization tool which update the dual variables
in every iteration.

> Cheers,
> - -T
> | Just a curious question. I know that people have been very
> | successful at writing programs which can play chess more or less
> | perfectly. Would it be possible to write a program which can write
> | chess playing programs? Has anyone tried this or some equivalent?
> |
> | Thanks, Michael
> |

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT