GELIELEO/Sparse-Latent-Policy-Search

GrouPS algorithm for latent space policy search. GrouPS combines reinforcement learning and dimensionality reduction, while also including prior structural knowledge about the task

GrouPS

GrouPS algorithm - Implementation of Sparse Latent Policy Search. It combines reinforcement learning and dimensionality reduction, while also including prior structural knowledge about the task.

Dependencies

• All code is written in Python 3.
• Please install 'numpy' and 'scipy' libraries.

Description of files

Files that should NOT be edited:

filename	description
main.py	Starts the program and has GrouPS algorithm.
update_equations.py	Contains the update equations required by GrouPS algorithm.

Files that can be edited:

filename	description
configuration.py	Contains parameters required by GrouPS algorithm.
get_samples.py	Code connecting the simulators and GrouPS algorithm.

Usage

python main.py

It loads up the simulator and starts the training. Displays Iteration deatails on the terminal. Stores 'checkpoint.npy' for every iteration. It contains the distributions learned.

Configuration

In configuration.py , please check the following variable.

load_the_latest_state = True  ## Loads the 'checkpoint.npy'
load_the_latest_state = False ## Does not load the saved state

To begin training the task from the start please set the above variable to false.