Skip to content

Latest commit

 

History

History
51 lines (44 loc) · 2.23 KB

README.md

File metadata and controls

51 lines (44 loc) · 2.23 KB
Raw

MEPOL

This repository contains the implementation of the MEPOL algorithm, presented in A Policy Gradient Method for Task-Agnostic Exploration.

Installation

In order to use this codebase you need to work with a Python version >= 3.6. Moreover, you need to have a working setup of Mujoco with a valid Mujco license. To setup Mujoco, have a look here. To use MEPOL, just clone this repository and install the required libraries:

git clone https://github.com/muttimirco/mepol.git && \
cd mepol/ && \
python -m pip install -r requirements.txt

Usage

Before launching any script, add to the PYTHONPATH the root folder (mepol/):

export PYTHONPATH=$(pwd)

Task-Agnostic Exploration Learning

To reproduce the maximum entropy experiments in the paper, run:

./scripts/tae/[mountain_car.sh | grid_world.sh | ant.sh | humanoid.sh | hand_reach.sh | higher_lvl_ant.sh | higher_lvl_humanoid.sh]

It should be straightforward to run MEPOL on your custom gym-like environments. For this purpose, you can have a look at the main training script.

Goal-Based Reinforcement Learning

To reproduce the goal-based RL experiments, run:

./scripts/goal_rl/[grid_goal1.sh | grid_goal2.sh | grid_goal3.sh | humanoid_up.sh | ant_escape.sh | ant_navigate.sh | ant_jump.sh]

By default, this will launch TRPO with MEPOL initialization. To launch TRPO with a random initialization, simply omit the policy_init argument in the scripts. For further modifications, you can check the main training script.

Results visualization

Once launched, each experiment will log statistics in the results folder. You can visualize everything by launching tensorboard targeting that directory:

python -m tensorboard --logdir=./results --port 8080

and visiting the board at http://localhost:8080.

Citing

To cite the MEPOL paper:

@misc{mutti2020policy,
    title={A Policy Gradient Method for Task-Agnostic Exploration},
    author={Mirco Mutti and Lorenzo Pratissoli and Marcello Restelli},
    year={2020},
    eprint={2007.04640},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}