Extending the RLLAB Benchmark of Deep Reinforcement Learning for Continuous Control

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Extending the RLLAB Benchmark of Deep Reinforcement Learning for 
Continuous Control"

by

Liyu CHEN

Abstract:

With the success of deep learning, deep reinforcement learning has 
achieved significant progress in recent years. While reinforcement 
learning algorithms for tasks with discrete state-action space has been 
extensively studied, algo- rithms for continuous state-action space are 
not so well-understood. In this project, we made two contributions on the 
study of reinforcement learning for continuous control. Firstly, we 
introduced new tasks to the recent published platform RLLAB, a benchmark 
for reinforcement learning with continuous control. We tested algorithms 
implemented in RLLAB on our new tasks and reported some novel findings. 
Secondly, we designed a new algorithms based on asynchronous update method 
and intrinsic motivation to tackle the no- torious hierarchical tasks, in 
which all existing algorithms failed to learn a good policy due to 
sparsity of rewarding events. We compared our algorithm with current state 
of the art and reported our understandings on solving hierarchical control 
tasks.

Date                 : 29 April 2017 (Sat)

Time                 : 11:00 - 12:00

Venue                : 2404 (via lifts 17/18)

Advisor              : Prof. Dit-Yan YEUNG

2nd Reader           : Prof. Nevin ZHANG