Giseung Park

Hello! I'm Giseung. [CV]

I recently completed my Ph.D. at the Korea Advanced Institute of Science and Technology (KAIST), Korea, under the guidance of Prof. Youngchul Sung. Currently, I am a researcher at KAIST with the support of AI Hub Project funded by the Korean government.

As a researcher, my goal is to (i) gain a deep understanding of AI fundamentals and explain why AI works effectively, and (ii) develop practical algorithms that are widely accessible and easy to use for everyone. I am eager to collaborate with experts from diverse backgrounds to solve challenging problems!

Contact: gs.park [at] kaist [dot] ac [dot] kr 

Research Interests

Reinforcement Learning (RL)

Ideal RL typically requires (i) states with the Markov property and (ii) a well-designed scalar reward. However, many real-world scenarios do not meet these requirements. My research focuses on partially observable RL and multi-objective RL (MORL) to address situations where one or both of these conditions are not guaranteed.

I believe that advancing the practical application of RL in real-world problems requires a strong focus on sparse-reward RL, multi-agent systems, and multi-modal RL. My research interests include developing data-efficient multi-agent RL algorithms, robust multi-modal RL approaches, and exploring various RL applications.

News

(Feb. 2025) Our paper on sparse-reward RL has been accepted to Neurocomputing! I appreciate the editor-in-chief and 7 reviewers' effort to provide constructive comments.

(Jan. 2025) Our paper on reward dimension reduction in MORL has been accepted to ICLR 2025! I'm happy that all the reviewers unanimously appreciate our work, with an average rating of 7.0. I also thank the support of the AI Hub Project for this work.

(Jan. 2025) Our paper on sparse-reward RL has been conditionally accepted to Neurocomputing

(Dec. 2024) I attended NeurIPS in Vancouver, BC, Canada. 

(Aug. 2024) Our lab has joined a research project called AI Hub Project (funded by the Korean government), where we are participating in the multi-modal RL-based decision-making branch to develop general-purpose foundation models for robotics. I’m excited to collaborate with numerous esteemed research groups on this project! [Official news link from the Korean government with details of the project.]

(Jul. 2024) I presented a poster at ICML in Vienna, Austria. (Check out my photo below! 😊)

(Jun. 2024) I passed my Ph.D. dissertation evaluation! I sincerely thank the committee members—Prof. Leshem, Kim, Park, Ahn, and my advisor, Prof. Sung—for their kind and helpful feedback.

(May. 2024) Our paper has been has been accepted to ICML 2024! I'm very proud of this achievement, which reflects the dedicated effort of our Korea-Israel international research project.

Selected Publications

Reward Dimension Reduction for Scalable Multi-Objective Reinforcement Learning

Giseung Park, Youngchul Sung

International Conference on Learning Representation (ICLR), 2025 (Acceptance rate: 32.0%)

TL;DR: A simple online reward dimension reduction method that enables MORL algorithms to scale to environments with many objectives.

Adaptive Multi-Model Fusion Learning for Sparse-Reward Reinforcement Learning

Giseung Park, Whiyoung Jung, Seungyul Han, Sungho Choi, Youngchul Sung

Aceepted to Neurocomputing

TL;DR: An adaptive fusion algorithm of intrinsic rewards from multiple prediction models in sparse-reward RL based on key axiomatic conditions.

The Max-Min Formulation of Multi-Objective Reinforcement Learning: From Theory to a Model-Free Algorithm

Giseung Park, Woohyeon Byeon, Seongmin Kim, Elad Havakuk, Amir Leshem, Youngchul Sung

International Conference on Machine Learning (ICML), 2024 (Acceptance rate: 27.5%)

TL;DR: An explicit formulation of max-min MORL and a practical model-free algorithm for max-min MORL.

Blockwise Sequential Model Learning for Partially Observable Reinforcement Learning (Oral, Top 4.6%)

Giseung Park, Sungho Choi, Youngchul Sung

AAAI Conference on Artificial Intelligence (AAAI), 2022 (Acceptance rate: 15.0%)

TL;DR: A new architecture based on sequential block inputs in POMDPs, learned through a novel direct gradient estimation using self-normalized importance sampling.

Population-Guided Parallel Policy Search for Reinforcement Learning

Whiyoung Jung, Giseung Park, Youngchul Sung

International Conference on Learning Representations (ICLR), 2020 (Acceptance rate: 26.5%)

TL;DR: Multiple learners share a common replay buffer and collaboratively search for a good policy with guidance from the best policy information.

Personal

I love swimming, yoga, and hiking. In 2011, I won a silver medal in the 100m butterfly at the National University Swimming Competition (Amateur Track) in South Korea. The following year, I served as the president of the KAIST Swimming Team. 😊