GSoC: Supercompilation

[Alesis Novik]

Greetings. I was wondering about a couple of thing. First of all would
be the general requirements for working on MOSES. Secondly, what are
the expectations and requirements for working on the Supercompilation
of Combo. What programming language should I have knowledge of? Thank
you in advance.

Sincerely,
Alesis Novik

[Moshe Looks]

Alesis,

Thank you for your interest. The baseline requirement for working on
MOSES is a strong grasp of C++, particularly modern C++ (STL & Boost).
Knowledge of and experience with functional programming, basic
probability theory, and machine learning (ideally
estimation-of-distribution algorithms and program evolution) are all
pluses, but are of more or less importance depending on the particular
project chosen. I will let someone else (i.e. Ben Goertzel) elaborate
with regard to supercompilation...

- Moshe

[Alesis Novik]

I had experience using STL, less with Boost, but I have a good
foundation with C++, so getting deeper knowledge wouldn't be much of a
problem, in my opinion. I also have some knowledge of the probability
theory and am currently reading on AI and machine learning. I would
like to get some details on the project itself. Thank you in advance.

Sincerely,
Alesis Novik

[Moshe Looks]

The conceptual background for MOSES is best described in my doctoral
dissertation (metacog.org/main.pdf), which may be considered an
offshoot of the Novamente project (novamente.net). There are also some
wiki docs on the MOSES google code page (code.google.com/p/moses/),
and of course the MOSES codebase itself (which is rather messy at the
moment - cleaner version will be rolled out by the summer). Please
feel free to comment/critique/question! There are a bunch of
individual tasks that could be taken on for GSoc projects (the ideas
page outlines them)...

Regards,
Moshe

>
>
> Sincerely,
> Alesis Novik
> >
>

[Alesis Novik]

I read your dissertation to get a better view of what is MOSES and
read some documents on supercompilers.com to get a view of what is
supercompilation. I looked through the other docs and I was wondering
if there was some sort of reference of Combo. Also, I got interested
in other MOSES ideas, like Procedure learning and wonder if you could
suggest any readings on the Pleasure Algorithm? Thank you in advance.

Sincerely,
Alesis Novik

[Moshe Looks]

Alesis,

I've actually been spending the last couple of months redesigning and
rearchitecting the combo language. Attached are a doc describing the
syntax and some sample (non-AI) code. Its really just a Lisp with some
syntactic sugar - the nifty AI features are not covered in the
documentation yet, sorry (nor are they implemented yet either - maybe
this is part of a SoC project!). As a sample, two of these are
treating probability distributions over functions in a uniform fashion
that allows them to ineroperate with regular functions, and special
runtime macros to implement the rewrite rules used for program
transformation in MOSES (which can themselves be uncertain & learned,
etc.).

A description of pleasure by Ben is available from
http://groups.google.com/group/opencog/files .

- Moshe

[Alesis Novik]

I've read the documentation provided, looked at the example and read
about the Pleasure algorithm, and as Ben Goertzel said, I think it
would be a more interesting project for the Google Summer of Code. The
supercompiler could be made at a later time. As Ben Goertzel said, I
would like to try and integrate the Pleasure Algorithm with MOSES.
Since I have a good c++ and STL foundation and am reading about the
boost library, especially the classes used in MOSES.

Sincerely,
Alesis Novik