What are some URL's of the bad ones?

You'll want to start out by reading the VM specification.  You will end
up implementing your own ClassLoader, so look up java.lang.ClassLoader
in the API docs.  Then take a look at:

   http://java-source.net/open-source/bytecode-libraries

Many of those products almost certainly have tutorials and sample code.  
The only one I've used is BCEL, so I can't give recommendations, but I
know Javassist and cglib are also generally popular.

see http://mindprod.com/jgloss/jasm.html

It is not great, but it is what I have.

I would recommend a different approach, where you dynamically create
sequences of "callable objects", perhaps following the Command or
Strategy pattern. Write (not synthesize) classes for the "atomic
operations", and make a class for a sequence of these atomic
operations. Then have the machine learning bit adjust parameters,
sequences etc. in these callable objects, instead of generating new
classes.

Assuming that you don't want to follow Tor's very sensible suggestion, and
adding to Chris's recommendations:

You will probably find some useful stuff in the GNU Kawa project -- a Scheme
system that dynamicaly compiles to Java bytecodes.

   http://www.gnu.org/software/kawa/

I have seen comments to the effect that the bytecode generation library in Kawa
is easier to use than Apache BCEL, but I haven't used either of them myself.

You might also want to look for the book:

   Programming for the Java Virtual Machine
   Joshua Engel

which talks quite a lot about compiling non-Java to run on the JVM.

   -- chris