Policy Fusion for Adaptive and Customizable Reinforcement Learning Agents

This repository is intended as Supplementary Maetrials to the paper Policy Fusion for Adaptive and Customizable Reinforcement Learning Agents.

Prerequisites

• The code was tested with Python v3.6.
• To install all required packages:

  cd Policy-Fusion-RL
  pip install -r requirements.txt
  # Install the modified version of miniworld
  pip install -e gym_miniworld/
  

Reproducing Results

Note that for this experiment we use a modified version of MiniWorld environment. You can find more details in the paper.

Test fusion methods

The repository provides already trained agents to only test the fusion methods.

The agents are trained with rewards:

 R_0: +1 for collecting a box
 R_1: +1 for collecting a ball

You can choose between 'mp', 'pp', 'et' or 'ew'. The names refer to the same names of the paper.

 python test_fusion.py -mn=mini_box,mini_ball -fm=ew

Train agents from scratch

You can also train agents from scratch with either 'ball', 'box' or 'complete' reward.

 python train.py -mn=mini_ball -rt=ball

You can re-use the new trained model with the first script test_fusion.py.

Plot results

You can plot the results obtained with the script test_fusion.py in the same shape of the ones in the paper using:

 python plot_continual.py

Example