Machine Learning in Optimization ?

Hi,

I've read the paper "Using Machine Learning to Focus Iterative
" where the authors use machine learning to learn "good"
sequences of compiler transformations for particular classes of
programs. They then use the learned data as start sequences for new
programs to run an focused iterative compiler optimization based on a
genetic algorithm. That way they don't waste time with "worse"
sequences.

Just out of curiosity I was wondering how the machine learning system
is integrated into a compiler. Is this done by having a data base
which is extended during learning or are the results stored in a file?
And what free tools are there which can be integrated in a C++
project? Any powerful C++ libraries?

I'd be grateful to hear your experiences.

Thank you.

Best regards,
Stephan

Hi Bjorn,

thank you for your answer.

You may want to look at GCC ICI [1] (= Interactive Compilation
Interface). Basically, GCC ICI is an initiative to open up the
internal heuristics of GCC and allow an external tool to make the
decisions for GCC whenever it has to decide on whether or not to apply
a transformation or to choose a transformation parameter. So, GCC ICI
is a GCC framework with handles to an external decision making tool
which may or may not be based on machine learning.

I will have a look at the software.

In terms of machine learning packages WEKA [2] may be useful to
you. As WEKA is written in Java you may need to develop your own C++
wrappers, though. Some information on bridging WEKA and .NET can be
found in [3].


I'm still interested in 2 other issue:
1) how many lines of code (or benchmarks) were required in your
paper to achieve good results for machine learning, i.e. how much
code had to be passed in the training phase before the learned
data base provided reliable results for the classification of
new benchmarks?
2) The selection of static features that describe a particular C
construct (as is used in your paper to learn "good" optimizations for
different classes of programs) seems to be a crucial part. Can you
give me some advices how such "good" features for learning can
be found? Is it promising to specify as many static features of
the C programs in the training phase as possible (intuitively more
features should allow one to describe a program more precisely) or
might a too large set of features have a negative impact of machine
learning aiming at pattern recognition?

Best regards,
Stephan