Hong Je-Gal
SEJONG UNIVERSITY · MAIN LAB

Intelligent decision-making & explainable AI

Hong Je-Gal 제갈홍

Ph.D. Student · Sejong University

I am a second-year Ph.D. student at Sejong University in the Department of Artificial Intelligence and Robotics, and a member of the Machine Intelligence and Networking (MAIN) Lab,
advised by Prof. Hyun-Suk Lee.

I study intelligent decision-making through reinforcement learning and explainable AI. I am broadly interested in understanding how learning-based systems behave and finding ways to improve their decision-making. I believe that advances in AI are blurring the boundaries between fields. I hope to develop AI methods that are easier to understand and more useful in practice across diverse domains.

  • Agentic AI
  • Reinforcement Learning
  • Explainable AI

News

Publications

Google Scholar ↗

A closer look

Explore the papers

Open a paper for its key figure and a one-minute introduction.

Conferences & Workshops 05

2026NeurIPS 2026·Conference · AcceptedLLMs As MDP Designers: A Dependency-Aware Agentic Framework for Automated Robot Policy Generation

H. Je-Gal, J. Kim, C. Yi, and H.-S. Lee

Conference on Neural Information Processing Systems (NeurIPS), 2026

ARPG pipeline from task and robot descriptions through sequential MDP modeling, validation, simulator implementation and policy training.
ARPG framework · Figure 2 in the manuscript. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

From a robot task description to a reinforcement learning pipeline that can actually learn.

ARPG uses LLM agents to design the learning problem before deployment. It builds an MDP in dependency order, checks and repairs intermediate decisions, and then implements the simulator through specialized agents. This keeps states, actions, rewards, and termination conditions aligned. Isaac Sim experiments show improvements in modeling correctness, executable environments, and successful policy generation over monolithic LLM and general agent baselines.

Takeaway Reliable robot learning begins with a consistent problem formulation.

2026ICML 2026·ConferenceFactorized Scheduling Principle: Learning Interpretable and Transferable Policies via Structured Additive Functions

Hong Je-Gal and Hyun-Suk Lee

International Conference on Machine Learning (ICML), 2026

FSP overview: candidate items, condition distribution, additive global scheduling principle, and priority scoring and selection.
FSP framework · Figure 1 in the paper. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Learn a scheduling rule once, understand its decisions, and reuse it at different system scales.

FSP represents the candidate population as a condition distribution and learns a global scheduling principle from experience. It decomposes that principle into additive effects of individual features and feature pairs, making their contributions inspectable. At deployment, the learned principle produces a priority score for each candidate. Experiments on synthetic and realistic tasks demonstrate strong scheduling performance and zero-shot transfer across different system sizes.

Takeaway An interpretable priority rule can also be a transferable learned policy.

BibTeX
@inproceedings{jegal2026factorized,
  title={Factorized Scheduling Principle: Learning Interpretable and Transferable Policies via Structured Additive Functions},
  author={Je-Gal, Hong and Lee, Hyun-Suk},
  booktitle={International Conference on Machine Learning},
  year={2026},
  url={https://arxiv.org/abs/2609.36578}
}
2025NeurIPS Workshop 2025·WorkshopA-LAMP: Agentic LLM-Based Framework for Automated MDP Modeling and Policy Generation

Hong Je-Gal, Chan-Bin Yi, and Hyun-Suk Lee

NeurIPS Workshop on Multi-Turn Interactions in Large Language Models, 2025

Comparison of a human expert workflow with A-LAMP agents for abstract task interpretation, MDP formulation, and coding.
Human and A-LAMP workflows · Figure 1 in the paper. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Turn a natural-language task into an MDP, an executable environment, and a trained policy.

A-LAMP divides this process among agents responsible for modeling, coding, and execution. Each stage produces artifacts that can be checked, helping preserve the meaning of the original task throughout the pipeline. Across classic control and custom reinforcement learning problems, the framework generates policies more reliably than a single LLM. A lightweight version using smaller models also approaches the performance of much larger models.

Takeaway Explicit modeling stages make automated policy generation easier to inspect and verify.

BibTeX
@article{jegal2025alamp,
  title={A-LAMP: Agentic LLM-Based Framework for Automated MDP Modeling and Policy Generation},
  author={Je-Gal, Hong and Yi, Chan-Bin and Lee, Hyun-Suk},
  journal={arXiv preprint arXiv:2512.11270},
  year={2025},
  url={https://arxiv.org/abs/2512.11270}
}
2024IEEE MetaCom 2024·Meta-XP WorkshopA Novel Data-Driven Soft Sensor in Metaverse Provisioning Predictive Credibility Based on Uncertainty Quantification

S.-H. Park, H. Je-Gal, and H.-S. Lee

IEEE MetaCom · Meta-XP Workshop, pp. 285–290, 2024

Multiple physical sensors provide measurements to a soft sensor, which estimates a target sensor value and its predictive credibility.
Soft-sensing system and predictive credibility · Figure 1 in the paper. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

A virtual sensor should report how trustworthy its prediction is, as well as the predicted value.

Soft sensors estimate measurements from data, but their reliability can vary with operating conditions. This work adds uncertainty quantification to a soft-sensor architecture so that each prediction is accompanied by a confidence interval. Experiments use scenarios with different levels of predictive uncertainty to evaluate those intervals. The resulting credibility information helps users assess when a prediction is dependable and when it needs more caution.

Takeaway Prediction uncertainty is useful information for operating a digital twin.

2023IEEE MetaCom 2023·Meta-XP WorkshopEfficient Federated Digital Twin Synchronization in Edge-Cloud Collaborative System

J.-W. Kim, H. Je-Gal, and H.-S. Lee

IEEE MetaCom · Meta-XP Workshop, pp. 714–719, 2023

Digital twin networks split neural inference between edge and cloud, with multiple early-exit points.
Edge–cloud collaboration and early exits · Figure 1 in the paper. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Decide where a digital twin’s neural network should run to reduce synchronization time.

A federated digital twin contains multiple twins whose updates depend on neural-network inference. This work partitions each network between an edge device and the cloud and accounts for early exits that can finish inference before all layers run. It formulates the choice of partition points as an optimization problem to minimize expected total synchronization time, and develops an exhaustive partitioning algorithm for finding the optimal choices.

Takeaway The placement of computation and the chance of an early exit need to be considered together.

International Journals 02

2024JMSE 2024·International journalTime-Series Explanatory Fault Prediction Framework for Marine Main Engine Using Explainable Artificial Intelligence

H. Je-Gal, Y.-S. Park, S.-H. Park, J.-U. Kim, J.-H. Yang, S. Kim, and H.-S. Lee

Journal of Marine Science and Engineering, 12(8), 1296, 2024

Marine engine prediction framework with correlation and SHAP based feature selection, model training and time-specific explanations.
Training and explanation framework · Author’s research presentation. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Predict an engine fault and show which sensor readings and time periods contributed to the prediction.

This framework combines feature selection, time-series prediction, and explanation for marine main-engine monitoring. Statistical correlation and SHAP analysis help select useful sensor inputs. A prediction model estimates future sensor behavior, while temporal explanations identify influential portions of the input history. Evaluation includes simulated fault conditions in operational sensor data, connecting a predicted warning to evidence that an engineer can inspect.

Takeaway A useful fault warning should explain both what may happen and the evidence behind it.

2023JMSE 2023·International journalA Novel Time–Frequency Feature Fusion Approach for Robust Fault Detection in a Marine Main Engine

H. Je-Gal, S.-J. Lee, J.-H. Yoon, H.-S. Lee, J.-H. Yang, and S. Kim

Journal of Marine Science and Engineering, 11(8), 1577, 2023

RPM data split into time-domain and frequency-domain feature extraction, then combined for neural-network fault detection.
Feature-fusion framework · Figure 6 in the paper (CC BY 4.0). Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Keep fault detection reliable even when engine vibration data are sampled less frequently.

Downsampling reduces data volume but can remove information needed for diagnosis. This work combines time-domain statistics with frequency-domain features extracted using EMD–ICA and the Wigner bispectrum. A neural network then classifies engine conditions using the fused features. In the study’s experiments, the combined representation outperforms either domain alone and retains 96.3% accuracy at a 100× downsampling rate.

Takeaway Complementary time and frequency features preserve diagnostic information under reduced sampling.

Domestic Journals 01

2023J-KICS 2023·Domestic journalDigital Role-Model Twin Using Reinforcement Learning Model

J.-W. Kim, H. Je-Gal, S.-J. Lee, and H.-S. Lee

The Journal of Korean Institute of Communications and Information Sciences, 48(2), 282–293, 2023

Role models and contextual decision features feed a shared preference network that offers role-model-specific preferences to users.
Digital role-model twin structure · Figure 1 in the paper. Click to enlarge ↗
THE PAPER IN A MINUTE~1 min read

Learn how a role model makes decisions, then let others explore those preferences in their own context.

The proposed digital twin learns from a target person’s choices in decision-making tasks. Reinforcement learning combines contextual information, role-model features, and a preference network to estimate which actions that person would favor. Users can then query the learned model with new decision problems and contextual information. Experiments examine how well the twin learns and reproduces the target person’s preferences.

Takeaway A digital twin can represent decision preferences as well as physical behavior.

Background

Education

Mar 2025 – Present
Ph.D. Student, Artificial Intelligence and RoboticsSejong University · MAIN Lab
Mar 2023 – Feb 2025
M.S., Intelligent Mechatronics EngineeringSejong University
Mar 2017 – Feb 2023
B.S., Intelligent Mechatronics EngineeringSejong University

Research & Project Experience

Jan – Dec 2024
Video prediction explanations for autonomous twinsGovernment-funded research project

Developed methods for identifying important spatial regions and time frames in video predictions, supporting essential-information extraction for communication between autonomous twins.

Jan – Dec 2023
Explainable marine engine fault predictionIndustry-academic collaboration · Hanwha Ocean

Led feature preprocessing and prediction-explanation work, combining statistical analysis with XAI to identify sensor- and time-specific contributions to fault predictions.

Sep 2021 – Apr 2022
Time–frequency feature fusion for marine engine diagnosisIndustry-academic collaboration · Hanwha Ocean

Developed data preprocessing that combines time- and frequency-domain features to preserve diagnostic information under reduced sampling rates.

Mar – Jun 2022
Personalized news recommendation serviceCapstone team project · Backend development

Integrated a reinforcement-learning recommender, news collection, and a KakaoTalk chatbot in a cloud-based service.