Learning with training wheels: Speeding up training with a simple controller for Deep Reinforcement Learning

By Linhai Xie, Sen Wang, Stefano Rosa, Niki trigoni, Andrew Markham.

The tensorflow implmentation for the paper: Learning with training wheels: Speeding up training with a simple controller for Deep Reinforcement Learning

Contents

0. Introduction
1. Prerequisite
2. Instruction
3. Citation

Introduction

In this project we proposed a switching machanism to let the agent learn from another simple controller, e.g. PID, during training instead of purely random exploration and speed up the training of DDPG.

For details please see the paper

The implementation of DDPG is based on Emami's work.

Prerequisites

Tensorflow > 1.1

ROS Kinetic

ros stage

matplotlib

cv2

Instruction (Edited)

roscore

rosrun stage_ros stageros worlds/Blank.world

rosrun stage_ros stageros worlds/Obstacles.world

rosrun stage_ros stageros worlds/Obstacles3.world

python DDPG.py

tensorboard --logdir [directory name]

Citation

If you use this method in your research, please cite:

@INPROCEEDINGS{8461203, 
	author={L. Xie and S. Wang and S. Rosa and A. Markham and N. Trigoni}, 
	booktitle={2018 IEEE International Conference on Robotics and Automation (ICRA)}, 
	title={Learning with Training Wheels: Speeding up Training with a Simple Controller for Deep Reinforcement Learning}, 
	year={2018}, 
	volume={}, 
	number={}, 
	pages={6276-6283}, 
	doi={10.1109/ICRA.2018.8461203}, 
	ISSN={2577-087X}, 
	month={May},}

Note to self

Use 'git push -u origin' to push changes I've made to this repository