/softlearning

Softlearning is a reinforcement learning framework for training maximum entropy policies in continuous domains. Includes the official implementation of the Soft Actor-Critic algorithm.

Primary LanguagePythonOtherNOASSERTION

Softlearning

Softlearning is a deep reinforcement learning toolbox for training maximum entropy policies in continuous domains. The implementation is fairly thin and primarily optimized for our own development purposes. It utilizes the tf.keras modules for most of the model classes (e.g. policies and value functions). We use Ray for the experiment orchestration. Ray Tune and Autoscaler implement several neat features that enable us to seamlessly run the same experiment scripts that we use for local prototyping to launch large-scale experiments on any chosen cloud service (e.g. GCP or AWS), and intelligently parallelize and distribute training for effective resource allocation.

This implementation uses Tensorflow. For a PyTorch implementation of soft actor-critic, take a look at rlkit.

Getting Started

Prerequisites

The environment can be run either locally using conda or inside a docker container. For conda installation, you need to have Conda installed. For docker installation you will need to have Docker and Docker Compose installed. Also, most of our environments currently require a MuJoCo license.

Conda Installation

  1. Download and install MuJoCo 1.50 and 2.00 from the MuJoCo website. We assume that the MuJoCo files are extracted to the default location (~/.mujoco/mjpro150 and ~/.mujoco/mujoco200_{platform}). Unfortunately, gym and dm_control expect different paths for MuJoCo 2.00 installation, which is why you will need to have it installed both in ~/.mujoco/mujoco200_{platform} and ~/.mujoco/mujoco200. The easiest way is to create a symlink from ~/.mujoco/mujoco200_{plaftorm} -> ~/.mujoco/mujoco200 with: ln -s ~/.mujoco/mujoco200_{platform} ~/.mujoco/mujoco200.

  2. Copy your MuJoCo license key (mjkey.txt) to ~/.mujoco/mjkey.txt:

  3. Clone softlearning

git clone https://github.com/rail-berkeley/softlearning.git ${SOFTLEARNING_PATH}
  1. Create and activate conda environment, install softlearning to enable command line interface.
cd ${SOFTLEARNING_PATH}
conda env create -f environment.yml
conda activate softlearning
pip install -e ${SOFTLEARNING_PATH}

The environment should be ready to run. See examples section for examples of how to train and simulate the agents.

Finally, to deactivate and remove the conda environment:

conda deactivate
conda remove --name softlearning --all

Docker Installation

docker-compose

To build the image and run the container:

export MJKEY="$(cat ~/.mujoco/mjkey.txt)" \
    && docker-compose \
        -f ./docker/docker-compose.dev.cpu.yml \
        up \
        -d \
        --force-recreate

You can access the container with the typical Docker exec-command, i.e.

docker exec -it softlearning bash

See examples section for examples of how to train and simulate the agents.

Finally, to clean up the docker setup:

docker-compose \
    -f ./docker/docker-compose.dev.cpu.yml \
    down \
    --rmi all \
    --volumes

Examples

Training and simulating an agent

  1. To train the agent
softlearning run_example_local examples.development \
    --universe=gym \
    --domain=HalfCheetah \
    --task=v3 \
    --exp-name=my-sac-experiment-1 \
    --checkpoint-frequency=1000  # Save the checkpoint to resume training later
  1. To simulate the resulting policy: First, find the path that the checkpoint is saved to. By default (i.e. without specifying the log-dir argument to the previous script), the data is saved under ~/ray_results/<universe>/<domain>/<task>/<datatimestamp>-<exp-name>/<trial-id>/<checkpoint-id>. For example: ~/ray_results/gym/HalfCheetah/v3/2018-12-12T16-48-37-my-sac-experiment-1-0/mujoco-runner_0_seed=7585_2018-12-12_16-48-37xuadh9vd/checkpoint_1000/. The next command assumes that this path is found from ${SAC_CHECKPOINT_DIR} environment variable.
python -m examples.development.simulate_policy \
    ${SAC_CHECKPOINT_DIR} \
    --max-path-length=1000 \
    --num-rollouts=1 \
    --render-mode=human

examples.development.main contains several different environments and there are more example scripts available in the /examples folder. For more information about the agents and configurations, run the scripts with --help flag: python ./examples/development/main.py --help

optional arguments:
  -h, --help            show this help message and exit
  --universe {gym}
  --domain {...}
  --task {...}
  --num-samples NUM_SAMPLES
  --resources RESOURCES
                        Resources to allocate to ray process. Passed to
                        `ray.init`.
  --cpus CPUS           Cpus to allocate to ray process. Passed to `ray.init`.
  --gpus GPUS           Gpus to allocate to ray process. Passed to `ray.init`.
  --trial-resources TRIAL_RESOURCES
                        Resources to allocate for each trial. Passed to
                        `tune.run`.
  --trial-cpus TRIAL_CPUS
                        Resources to allocate for each trial. Passed to
                        `tune.run`.
  --trial-gpus TRIAL_GPUS
                        Resources to allocate for each trial. Passed to
                        `tune.run`.
  --trial-extra-cpus TRIAL_EXTRA_CPUS
                        Extra CPUs to reserve in case the trials need to
                        launch additional Ray actors that use CPUs.
  --trial-extra-gpus TRIAL_EXTRA_GPUS
                        Extra GPUs to reserve in case the trials need to
                        launch additional Ray actors that use GPUs.
  --checkpoint-frequency CHECKPOINT_FREQUENCY
                        Save the training checkpoint every this many epochs.
                        If set, takes precedence over
                        variant['run_params']['checkpoint_frequency'].
  --checkpoint-at-end CHECKPOINT_AT_END
                        Whether a checkpoint should be saved at the end of
                        training. If set, takes precedence over
                        variant['run_params']['checkpoint_at_end'].
  --restore RESTORE     Path to checkpoint. Only makes sense to set if running
                        1 trial. Defaults to None.
  --policy {gaussian}
  --env ENV
  --exp-name EXP_NAME
  --log-dir LOG_DIR
  --upload-dir UPLOAD_DIR
                        Optional URI to sync training results to (e.g.
                        s3://<bucket> or gs://<bucket>).
  --confirm-remote [CONFIRM_REMOTE]
                        Whether or not to query yes/no on remote run.

Resume training from a saved checkpoint

In order to resume training from previous checkpoint, run the original example main-script, with an additional --restore flag. For example, the previous example can be resumed as follows:

softlearning run_example_local examples.development \
    --universe=gym \
    --domain=HalfCheetah \
    --task=v3 \
    --exp-name=my-sac-experiment-1 \
    --checkpoint-frequency=1000 \
    --restore=${SAC_CHECKPOINT_PATH}

References

The algorithms are based on the following papers:

Soft Actor-Critic Algorithms and Applications.
Tuomas Haarnoja*, Aurick Zhou*, Kristian Hartikainen*, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, and Sergey Levine. arXiv preprint, 2018.
paper | videos

Latent Space Policies for Hierarchical Reinforcement Learning.
Tuomas Haarnoja*, Kristian Hartikainen*, Pieter Abbeel, and Sergey Levine. International Conference on Machine Learning (ICML), 2018.
paper | videos

Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor.
Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. International Conference on Machine Learning (ICML), 2018.
paper | videos

Composable Deep Reinforcement Learning for Robotic Manipulation.
Tuomas Haarnoja, Vitchyr Pong, Aurick Zhou, Murtaza Dalal, Pieter Abbeel, Sergey Levine. International Conference on Robotics and Automation (ICRA), 2018.
paper | videos

Reinforcement Learning with Deep Energy-Based Policies.
Tuomas Haarnoja*, Haoran Tang*, Pieter Abbeel, Sergey Levine. International Conference on Machine Learning (ICML), 2017.
paper | videos

If Softlearning helps you in your academic research, you are encouraged to cite our paper. Here is an example bibtex:

@techreport{haarnoja2018sacapps,
  title={Soft Actor-Critic Algorithms and Applications},
  author={Tuomas Haarnoja and Aurick Zhou and Kristian Hartikainen and George Tucker and Sehoon Ha and Jie Tan and Vikash Kumar and Henry Zhu and Abhishek Gupta and Pieter Abbeel and Sergey Levine},
  journal={arXiv preprint arXiv:1812.05905},
  year={2018}
}