MVE: Model-based Value Estimation

Using short-horizon nonlinear dynamics for on-policy simulation to improve value estimation. See the paper for algorithmic details.

Requirements

Here are setup-specific requirements that you really, really have to do yourself:

• MuJoCo 1.50, with the appropriate key available - MuJoCo downloads. On Linux, you can run ~/install-mujoco.sh with the key mjkey.txt in the CWD.
• Python 3.5 (scripts/ assume this is the python and pip in PATH)
• If you get any error messages relating to glfw3, then reinstall everything in a shell where the following environment variables are set (and for good measure in the shell where you're launching experiments):

    export LD_LIBRARY_PATH=~/.mujoco/mjpro150/bin
    export LIBRARY_PATH=~/.mujoco/mjpro150/bin

Other system dependencies:

• System dependencies for gym - gym README
• System dependencies for gym2 - gym2 README

Example installation:

# GPU version
# ... you install system packages here
conda create -y -n gpu-py3.5 python=3.5
source activate gpu-py3.5
pip install -r <( sed 's/tensorflow/tensorflow-gpu/' requirements.txt )

# CPU version
# ... you install system packages here
conda create -y -n cpu-py3.5 python=3.5
source activate cpu-py3.5
pip install -r requirements.txt

# Lazy version (defaults to CPU)
conda create -y -n cpu-py3.5 python=3.5
./scripts/install-mujoco.sh
./scripts/ubuntu-install.sh

# Lazy version (GPU)
conda create -y -n gpu-py3.5 python=3.5
./scripts/install-mujoco.sh
sed -i 's/tensorflow/tensorflow-gpu/' requirements.txt
./scripts/ubuntu-install.sh

Scripts

All scripts are available in scripts/, and should be run from the repo root.

Running experiments

To run experiments locally, use main_ray.py (note the resources requirements specified here are only used for scheduling trials; the actual processes are free to create as many threads as they want if you'd like to oversubscribe the machines by setting tf_parallelism within the YAML config to larger than the number of guaranteed CPUs). To multiplex the GPUs locally, set self_host to the total number of virtual GPUs.

python mve/main_ray.py --experiment_name hc0 --config experiments/hc0.yaml --ncpus 1 --ngpus 0 --self_host 1

Experiments can also be run in a distributed manner by connecting to a live ray cluster by changing --self_host to --port RAY_REDIS_PORT above. Multiple experiments can be run by the same driver. See python mve/main_ray.py --help for details.

Parallelism

Multiple components of this code run in parallel.

• Evaluation on environments is parallelized by multiple processes or threads (which one depends on the environment), and uses --env_parallelism workers. I have found that peak performance is reached when there is some batching; i.e., the number of workers is less than half the number of environments used for evaluation (default 8).
• If running on CPUs, TensorFlow internal thread pool parallelism (for each of intra-operation and inter-operation thread pools) is set by the value of --tf_parallelism (default nproc).
• If running on GPUs, then TensorFlow GPU count used is automatically set to those GPUs specified by the environment variable CUDA_VISIBLE_DEVICES, which if left empty uses the first GPU available on the machine. There is currently no support for actually using multiple GPUs in a single experiment.
• The OMP_NUM_THREADS is overriden; there's no need to set it.

Adding MuJoCo key to CI securely

Just use the manual encryption instructions here. .travis.yml is already configured to securely unencrypt mjkey.txt.gpg.

Adding New Fully Observable Environment

1. Create a new FullyObservable<env> under mve/envs. See FullyObservableHalfCheetah.py for a link to the Open AI Gym commit that contains the source code you should adapt.
2. Expose the new environment by importing it in mve/envs/__init__.py
3. Make it accessible through mve/env_info.py's _env_class function.

Make sure to test that your environment works and amend the tests in scripts/tests.sh to include a check that it runs correctly.