/DVRL

Deep Variational Reinforcement Learning

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

Deep-Variational-Reinforcement-Learning

This is the code accompanying the paper Deep Variational Reinforcement Learning for POMDPs by Maximilian Igl, Luisa Zintgraf, Tuan Anh Le, Frank Wood and Shimon Whiteson.

Disclaimer: I cleaned up the code a bit before release. A few test runs indicate it still works but if you encounter problems please let me know, either here as issue or via email (maximilian.igl@gmail.com). Also, if there are questions or something is unclear, please don't hesitate to approach me - feedback is very welcome!

Running the code

You can either use the provided docker container or install all dependencies.

Using docker

With nvidia-docker installed, first create the container:

cd docker
./build.sh

which builds the docker container (this will take a few minutes). Once that is done, you can run experiments from the main folder in a container using

cd ..
./docker/run.sh <gpu-nr> <name> <command>

for example

./docker/run.sh 0 test-run ./code/main.py -p with environment.config_file=openaiEnv.yaml

the results will be saved in the folder saved_runs using this structure. Please be sure to mount the folder accordingly if you want to access the data after the container finishes.

Without docker

Installing dependencies

You will need

  • Python v3.6 (I used Anaconda but it should work with other distributions as well)
  • Pytorch v0.4.x
  • openai baselines (On MacOS, I needed to install mpi4py using conda beforehand to make the install not fail)
  • pip install 'gym[atari]'

As well as other dependencies by running

pip install -r requirements.txt

in the main folder.

If you're running into an error with matplotlib on MacOS when running the RNN on MountainHike, you can use this simple solution.

Running

From the main folder execute

python ./code/main.py -p with environment.config_file=openaiEnv.yaml

The results will be saved in the saved_runs folder in subfolders with incrementing numbers.

Plotting

I included a very simple plotting script in the main folder:

python plot.py --id <id> [--metric <metric>]

where <id> is the ID of the experiment (created automatically and printed to command line when each run is started). <metric> is which metric you want to plot. result.true is the default and probably what you want, i.e. the unclipped reward.

We use sacred for configuration and saving of results. It fully supports a more elaborat setup with SQL or noSQL databases in the background for storing and retrieving results. I stripped that functionality out for the release for ease of use but can highly recommend using it when working more extensively with the code.

Reproducing results

Below are the commands to reproduce the results in the paper. Plots in the paper are averaged over 5 random seeds, but individual runs should qualitatively show the same results as training was fairly stable. If you run into problems, let me know (maximilian.igl@gmail.com).

Default configuration

The default configuration can be found in code/conf/ in the default.yaml. The environment must be specified in the command line by environment.config_file='<envName>.yaml'. The corresponding yaml file will be loaded as well (and overwrites some values in default.yaml, like for example the encoder/decoder architecture to match the observations space). Everything specified additionally in the command line overwrites the values in both yaml files.

DVRL:

python ./code/main.py -p with environment.config_file=openaiEnv.yaml environment.name=PongNoFrameskip-v0 algorithm.use_particle_filter=True algorithm.model.h_dim=256 algorithm.multiplier_backprop_length=10 algorithm.particle_filter.num_particles=15 opt.lr=2.0e-04 loss_function.encoding_loss_coef=0.1

RNN:

python ./code/main.py -p with environment.config_file=openaiEnv.yaml environment.name=PongNoFrameskip-v0 algorithm.use_particle_filter=False algorithm.model.h_dim=256 algorithm.multiplier_backprop_length=10  opt.lr=1.0e-04

(or with any other Atari environment) Please note that the results printed in the console are the clipped rewards, for the true rewards please check 'result.true' in the metrics.json file or use the plotting script

Credits

The code is based on an older version of https://github.com/ikostrikov/pytorch-a2c-ppo-acktr but heavily modified.