An approach to Hierarchical and Explainable Reinforcement Learning using Reward Decomposition and Options-based Learning for sequentially composable subtasks.

The following instructions have been validated on Ubuntu 18.04 only.

1. Install MuJoCo (if not already installed). See Mujoco Installation

2. Clone this repository

    git clone https://github.com/SomeshDaga/seqopt.git
   

3. Install Python 3.6 and Virtualenv

     sudo apt install python3.6
     sudo apt install python3-venv
   

4. Initialize a Python 3.6 Virtual Environment

     cd seqopt
     virtualenv -p /usr/bin/python3.6 .
   

5. Activate the virtual environment

     source bin/activate
   

6. Install python dependencies

     pip install -r requirements.txt
   

1. Download MuJoCo

2. Unzip and install

     unzip mujoco200_linux.zip
     mkdir -p ~/.mujoco
     mv mujoco200_linux ~/.mujoco/mujoco200
   

3. Obtain a MuJoCo license. Install the license mjkey.txt to ~/.mujoco

Since our method requires expert-defined aspects for any training environments, current implementation only supports

1. A custom variation of the manipulator domain and bring_ball task from the DeepMind Control Suite
2. A custom variation of the door task with the Jaco manipulator from robosuite

Our custom variations do not change the environment dynamics in any way, but rather introduce changes to things like initialization conditions, hand-crafted features etc.

Support for additional environments may be added by defining configuration files for newly added tasks such as those for the door and manipulatortasks in seqopt/scripts/door and seqopt/scripts/manipulator respectively.

An Options-Based Soft-Actor Critic algorithm is implemented for learning. A single option may be specified for the task (through the configuration files) to allow for benchmarking using a conventional (i.e. not options-based) SAC agent.

Note: A PPO-based implementation was also attempted but presented with poor learning across the task, and hence wasn't pursued further

Run all scripts from the root of the repository

Execute seqopt/scripts/train.py with desired arguments/flags

     usage: train.py [-h] [--option-critic] [--continue-training CONTINUE_TRAINING]
            [--eval-log-name EVAL_LOG_NAME] [--verbose]
            {ppo,sac}
            {door,door_benchmark,manipulator,manipulator_benchmark}

Example usage:

     python -m seqopt.scripts.train sac door --eval-log-name experiment_1 --verbose

Required (or Positional Arguments):

• Algorithm: ppo or sac (Best to avoid ppo as it may have bugs)
• Environment: For each environment, we define an options-based and a benchmark configuration e.g. door and door_benchmark. Use the *_benchmark environment for a conventional SAC agent configuration

Optional:

• --continue-training: Specify path to a checkpoint model zip file to resume training from
• --eval-log-name: Specify a folder name to enable checkpoint and tensorboard logging
• --option-critic: Use flag to enable the Option-Critic execution model. Otherwise, defaults to the Option Chain model
• --verbose: Use flag to enable logging training updates to terminal

Execute seqopt/scripts/evaluate.py with desired arguments/flags

     usage: evaluate.py [-h] [--stochastic-actions] [--stochastic-terminations]
               [--n-eval-episodes N_EVAL_EPISODES] [--seed SEED]
               [--no-render] [--device {cpu,cuda,auto}]
               {ppo,sac}
               {door,door_benchmark,manipulator,manipulator_benchmark}
               model

Example usage:

     python -m seqopt.scripts.evaluate sac door experiments/door/seqsac/option_critic.zip --stochastic-terminations --n-eval-episodes 5 --device cpu

Note: Please use the --device cpu flag on all runs (minor bugs with cuda implementation may exist)

Obtain fully-trained models (for supported environments) under the Options Chain and Option-Critic models, and a conventional SAC agent at this link.

More documentation to come soon...