Deep Reinforcement Learning For Agent Control

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

Final Year Thesis Oral Presentation

Title: "Deep Reinforcement Learning For Agent Control"

by 

Mr. ZENG Cancheng 


Abstract: 

In this final year thesis, we have developed deep reinforcement learning 
models that can learn to control in Atari 2600 games, with improvements 
over the previous DQN model [1] in training stability and efficiency. We 
have investigated the sources of the instability of DQN and largely 
improved the training stability with the use of AdaGrad optimization 
method. Apart from that, we have proposed a transfer learning model that 
can improve the initial performance during train- ing. We have also 
investigated other methods that could improve training efficiency, 
including parallel actor-learner training and distributed training with 
Elastic Averaging method.
  

Date                 : 30 April 2016 (Saturday)

Time                 : 11:30am to 12:15pm

Venue                : Room 5510 (lift 25/26)

Advisor              : Prof. D.Y. YEUNG
                              
2nd Reader           : Prof. Nevin ZHANG