Greeting from Taiwan! Good evening : )
I am intending to apply the admission for KAIST graduate school of artificial intelligence via Korean Government Scholarship program (KGSP).
The following content is my goal of study for the program.
KGSP Statement of purpose -- AI major -- Goal of study and detailed study plan
Please give your thoughts and advice. Thank you very much!
The magic of deep reinforcement learning (DRL) algorithms genuinely attracts me; hence, I wish to study this topic as my master's research. I had taken a graduate course of DRL when I exchanged in SNU and had written a thesis regarding an improved architecture of DDQN. However, in Taiwan, DRL research is not a very popular field. For example, in my school NTHU, only one professor is doing it. In contrast, the AI school of KAIST has quite a few professors doing DRL research. To access an academic environment that fits my interest, I hope to study for my master's degree in KAIST AI school.
My primary goal for the study in KAIST is to publish two papers regarding DRL during my master's studying period. Since I have been experienced in publishing papers and already have a good understanding of DRL, I think this goal is achievable. Regarding the detailed plan, I have conducted some literature survey on the KAIST publications and concluded my intended research field.
I am interested in Prof. Beomjoon Kim's research on task and motion planning (TAMP) problem. For example, in Prof. Kim's paper "Adversarial Actor-Critic Method for Task and Motion Planning Problems Using Planning Experience," they propose a novel loss function that learns the value function from sequences search trees and penalizes the operator instances generated by the policy adversarially. I feel excited when reading this kind of innovation on training DRL agents, and I hope I could help develop ideas like a better form of loss function or an advanced algorithm for plan skeleton searching. Of course, I would conduct the corresponding experiments and summarize them as a manuscript.
Further, I intend to research an alternative solution to the multiple control frequencies problem with Prof. Kee-Eung Kim and his students. Their work "Reinforcement Learning for Control with Multiple Frequencies" has proposed an Action-Persistent Actor-Critic (AP-AC) algorithm that performs the optimization for periodic policy. I think the multiple control frequencies problem is practical in real-world applications; thus, this field's research should be potential and valuable. After I engage in this lab, I would like to devise an algorithm to solve the problem of multiple control frequencies based on DQN or DDPG, and I expect the method could be analogous to the concept of AP-AC. Finally, I will experiment and compare the model performances.
I am also interested in studying the way we improve the DRL model with theoretical inference. The paper "Representation Balancing Offline Model-Based Reinforcement Learning" done by anonymous KAIST authors has motivated me to learn more about AI algorithms' theoretical analysis. This paper has proposed a framework balancing the target policy and behavior policy by adding a specific term on reward evaluation to minimize the off-policy evaluation error, and such regularizer is derived from a theoretical observation. I am amazed to know we could improve DRL algorithms from theoretical indication. I wish I could learn the necessary knowledge to hold theoretical derivations in my research, consequently, attach the theoretical analysis in my paper. I once read a paper using the framework of thermal physics to show the universal approximation theorem of neural networks. As a physics bachelor, I have a strong background in carrying out theoretical derivations; hence, I believe my participation in the theoretical analysis on DRL could bring a substantial contribution.
In summary, my main goal for the study in KAIST AI school is to publish two DRL papers during my master's studying period. My intending field of research includes but is not limited to studying TAMP problem with Prof. Beomjoon Kim, innovating a new solution for the issue of multiple control frequencies with Prof. Kee-Eung Kim, and exploring the theoretical way for improving DRL algorithms with the currently anonymous KAIST authors.
Your review and comment is more than welcome!!