I constrained mutations to "algebraic identities" -- which constrains the
paths to the search space.  So the evolutionary program synthesis is more
along the lines of a computer algebra system searching for a proof.

Which brings up related idea I proposed to the Maxima (Macsyma) folks back
when DeepMind was making its first inroads into the game of go:

If you gather user interactions with a computer algebra system like Macsyma
the way you gather moves in a game, you might be able to build up enough of
a database to train something like Alpha Go to do algebra in a reasonably
intelligent manner that doesn't get stuck in local minima -- you know --
like sacrificing pieces for strategic advantage.  Of course, there is also
AlphaGo Zero which purportedly required no database of human game play...




On Tue, Mar 23, 2021 at 6:39 PM Ben Goertzel <[email protected]> wrote:

> Hmm, well if you just use mutations then this becomes a greedy
> algorithm which will either get stuck in local optima or take
> close-to-forever
>
> If you use crossover operators or EDA-style probabilistic models then
> things become potentially tractable, but only under appropriate
> assumptions regarding the fitness landscape... right?
>
> On Tue, Mar 23, 2021 at 2:15 PM James Bowery <[email protected]> wrote:
> >
> > Evolutionary program synthesis requires a fitness/cost function which,
> in the case of Solomonoff Induction, can be approximated by the size of the
> program that outputs exactly the currently known observations.  The obvious
> problem with this approach is that of all algorithms, only a disappearingly
> small fraction will output exactly the known observations.
> >
> > Reduce the space by starting with the known observations as an
> executable literal -- say by putting it in quotes for evaluation -- and use
> a reversible programming language with its algebraic identities as
> mutations -- treating the "discarded" bits (inherent in reversible
> algorithms) as needing compression as well.  In the limit, this can be
> represented as a directed cyclic graph of reversible logic gates which will
> tend to configure in such a way as to make the "heat" bits highly
> compressible (and in the limit, all 0s or all 1s).
> >
> > This originally occurred to me prior to the announcement of the Hutter
> Prize back in 2006 but Matt had some argument debunking this approach.
> >
> > PS:  It was rather ironic that one of the first and most vocal critics
> of The Hutter Prize was the inventor of the Kayak reversible programming
> language.
> > Artificial General Intelligence List / AGI / see discussions +
> participants + delivery options Permalink
> 
> --
> Ben Goertzel, PhD
> http://goertzel.org
> 
> “He not busy being born is busy dying" -- Bob Dylan

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tf8bb7754cbb517a4-M9f826fbb20b2d1d9be8916fa
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to