cocoademon / rl-glue-ext Goto Github PK

http://java.sun.com/j2se/1.5.0/docs/api/java/nio/ByteBuffer.html#get(byte[])

Add uninstall options to manual

Hi all,
I am attaching the corresponfing "setup.py" which is the standar way to
install python modules.

the file should be included into the "src" directory of the distribution
and then run:  "python setup.py install" from within the src directory.

Also an __init__.py is included which sould be placed inside the src/rlglue
directory replacing the current empty __init__.py

bests,
jamh.

Have to update the task spec.

Needs Unit tests for the Java codec

Print out the version for the 3 codecs like we do in the rl_glue code, and
verify that the help is helpful and that those settings work.

I just imported Utils from RL-Glue, same story as issue 8 and issue 7.

Fix it.

Sometimes when modules aren't connected right or something ends in error,
they just stick around.

You might need to do ps -au and look for anything funny like:
btanner  5014  51.7  0.3   334604  13288 s004  R     8:57pm   0:01.62
/usr/bin/java -classpath /Users/btanner/JavaCodec/JavaRLGlueCodec.jar:.
SkeletonExperiment
btanner  4690  47.2  0.3   334508  13128 s004  R     8:38pm  17:31.62
/usr/bin/java -classpath products/JavaRLGlueCodec.jar
org.rlcommunity.rlglue.codec.tests.Test_1_Environment
btanner  4847  46.9  0.3   334508  13204 s004  R     8:52pm   4:05.26
/usr/bin/java -classpath /Users/btanner/JavaCodec/JavaRLGlueCodec.jar:.
SkeletonExperiment
btanner  5015  46.0  0.3   334508  13140 s004  R     8:57pm   0:01.51
/usr/bin/java -classpath /Users/btanner/JavaCodec/JavaRLGlueCodec.jar:.
SkeletonAgent


We should work to address these problems, if possible, and also put in the
docs that people should check for these things.

This is a big deal that we don't have time to do immediately for the codecs.

Should be probably done before the official journal release.

Need to make RL_init return the task spec and RL_episode return how
termination happened.

It might make sense to incorporated the Java RL-Glue codec into RLVizLib.

If we're going to split java from RLVizLib anyway, why not put them together.

Remove RL_Freeze and Agent_Freeze

In RL-Library, the Java Codec is reporting:
RL-Glue Java Agent Codec Version: null (null)
    Connecting to 127.0.0.1 on port 4096...

struct hostent has no member named h_addr

This is from RL_client_agent.c

Think it is a problem with GCC 4.3.

We probably need to go add a check to autoconf.

As per http://code.google.com/p/rl-glue/issues/detail?id=45

We fixed this problem somehow for the competition, need to look up how we
did that.

Need to either use global variables or clean them up after sending to the
network, currently its just leaking.

 C Codec AutoTools should check for the RL-Glue Headers

Basically we don't have task spec parsers that work in any of the
packages.. .or at least not tested.

This needs to be explained.

Update the docs to explain that matlab will try to find teh rl-glue codec,
but yhou should still set the java path.

I changed the behavior of rl_episode a bit and some of the tests will fail now.

Fix them.

I pulled in the examples directory from RL-Glue so we'd have a record of
all of the examples, but they haven't been sorted and don't work.

Make them work.

Do we need codecs for R, Octave, PERL, VB, C#, Ruby, etc?

I think it will be better to use typed arrays instead of lists.

http://docs.python.org/lib/module-array.html

Build 3 libraries, and put the headers in rlglue/codecs

The Java task spec parser will need to give the version back if it can't
parse a task spec because it doesn't recognize the version.

Probably this should be an emptish taskspecdelegate that implements getVersion.

And also update the experiment.

Get SVN Build into matlab code somehow

The network code surely needs to be updated to accomodate the char array.

[deleted issue]

Get SVN Build into python code somehow

To make sure all the charArray stuff works right!

- Make sure no more gpl
- Make sure rl-glue.googlecode is replaced with rl-glue-ext

Put the right names on all the test files!

We need to expose some utility methods to make live easier in
RL_client_experiment et al.  Check out RL_cleanup in there if you want to
see what I mean.

Mention that the type name Reward_observaiton has changed (or just updated
with the new name) and change get/set references to load/save

Make check should work without make install first, does it?

If not, then the test code (scripts included) needs to be written a little
more carefully.

After Matt got both the Java and Matlab sides working, I rewrote the Matlab
side to be more Matlaby.  This was a choice that Matt and I made together.

However, it means that the java code is now quite broken.  We need to fix
it to handle wrongly typed arrays (in case people don't cast their doubles
to ints in Observation.intArray or Action.intArray.  We also need to either
figure out how to pass structs into matlab functions from java, or we need
to write some matlab wrappers to go between Java and matlab.

The latter plan might be easiest, and would allow us to do much of our
checking inside matlab.

Make them configurable.  Look at how the envs and agents do it, see if that
can all be tidied up some.

Make a script that builds python code for Windows, .tar.gz, and also .rpm

It's easy.

http://docs.python.org/distutils/introduction.html#concepts-terminology

Make sure that the docs get distributed with it.

The C/C++ Codec is just an export of RL-Glue, it actually needs to be
cleaned up and made to link against an installed RL-Glue library.

The details from the files copied from rl-glue don't seem to be updating.

http://code.google.com/p/rl-glue/issues/detail?id=47

Hi,

Having trouble getting the java codec to notice the contents of the
RLGLUE_HOST environmental variable. Tried in linux, windows, and
windows/cygwin, all to no avail. In the same shell, I am able to run
programs using the C codec, and they read the variable correctly and
connect the the right host.

- John

Create a non-netbeans ant build script

The C codec pulls things off the network and then passes them back to the
ae^2.  (Agent, Environment, Experiment).

I think it's using some static global variables for things like the message
strings, observations, etc.  We should nail down and LOUDLY document
whether you have to make a copy of these things if you want to store them,
how to make a copy, and how to clean it up.

Get error:
make[3]: *** No rule to make target `test_sanity.sh'.  Stop.


Someone with more experience with automake can probably figure this out.

We should really be able to do:

./configure --with-rl-glue=/some/path and not have to export all sorts of C
and LD flags.

We should figure that out.

http://code.google.com/p/rl-glue/issues/detail?id=54

The matlab tests really need to be automated from the multi-tester.  Need
to write come code to make sure Matlab isn't tested with more than 1
component at a time.

This could probably be done fairly easily if we could figure out how to
start matlab with a command.

Sure it will slow things down, but I think that's ok.

Carry over from the JAva code, not all of the naming and capitalization are
consistent.

Why is the ROAT object .reward but .Observation (note that caps)?

Is it because reward is a primitive?

Maybe that argument will fly, but we should look over things and look for
any similar weirdness.

If you accidentally connect a matlab agent and another type of agent, or
any number of other common accidents, you might freeze all of matlab,
requiring you to force-quite and restart it.

I'm not sure exactly how to resolve this. Maybe by putting timeouts in the
Java codec and the using loops in the matlab code to call it?  Basically,
we need a way of forcing Matlab to come back from Java system calls.