Skip to content

Latest commit

 

History

History
20 lines (16 loc) · 1.12 KB

File metadata and controls

20 lines (16 loc) · 1.12 KB

Locally Constrained Representations

Implementation of the LCR algorithm on the following environments.

Environments

Atari

All Default Hyper-Parameters have been set. Select game with --game and run main.py. The codebase has been adapted from here

Gym-Control

Default Hyper-Paremeters is for the main plot. Choose appropriate hyper-parameters for ablation studies while setting the default ones as constant. Run main.py. Change game between 'CartPole-v1' and 'Acrobot-v1'

Mujoco

This is an extension of the RLZoo library. The LCR code is built on top of RLZoo with only SAC being modified. The code will not work for other algorithms. The default hyper-parameters are set to the default ones for RLZoo SAC. Run train.py to reproduce the results from the paper. Change the max_steps for each environment based on the table provided in the supplementary material.

MiniGrid

The new proposed environments are implemented in env_minigrid.py. Run train.py with default hyper-paramters to recreate the plots from the paper.