Kukjin Kim

Hello, I’m passionate about reinforcement learning and neural networks. My primary research goal is to develop massively scalable RL algorithms and a general-purpose agent for robotics that is capable of solving a wide range of real-world problems. Ultimately I aim to enable this versatile agent to adapt to new tasks throughout its lifetime. Currently, I work as a research engineer at LG Electronics. Prior to joining LGE, I was a master’s student at Korea University, where I was supervised by Prof. Changhee Joo.

What constitutes a general and adaptive RL agent? I believe it will be based on the integration of various research areas. For example, I think methods for continual reinforcement learning, new types of adaptive neural networks, and innovative learning algorithms are required. I will explore these areas during my future Ph.D. program.

news

Feb 23, 2024	I completed my master’s degree at Korea University.
Jan 29, 2024	I began my career at LG Electronics as a Research Engineer.
Oct 31, 2023	My paper was accepted at NeurIPS 2023 Workshop: Foundation Models for Decision Making.

selected publications

NeurIPS Workshop

Agnostic Architecture for Heterogeneous Multi-Environment Reinforcement Learning

Changhee Joo

NeurIPS 2023, Foundation Models for Decision Making, 2023

Abs HTML

In new environments, training a Reinforcement Learning (RL) agent from scratch can prove to be inefficient. The computational and temporal costs can be significantly reduced if the agent can learn across diverse environments and effectively perform transfer learning. However, achieving learning across multiple environments is challenging due to the varying state and action spaces inherent in different RL problems. Padding or naive parameter-sharing with environment-specific layers for different state-action spaces are possible solutions for multi-environment training. However, they can be less scalable when training for new environments. In this work, we present a flexible and environment-agnostic architecture designed for learning across multiple environments simultaneously without padding or environment-specific layers, while enabling transfer learning for new environments. We also propose training algorithms for this architecture to enable both online and offline RL. Our experiments demonstrate that multi-environment training with one agent is possible in heterogeneous environments and parameter-sharing with environment-specific layers is not effective in transfer learning.