Institution profile

Sungkyunkwan University

Academic institutionasia · kr
Official website
Research library543linked papers
Opportunities0open roles
Selected work

Representative Papers

Self-Predictive Dynamics for Generalization of Vision-based Reinforcement Learning

Jul 01, 2022International Joint Conference on Artificial Intelligence

Vision-based reinforcement learning exhibits poor generalization under unseen image corruptions (e.g., shadows, cloud cover, illumination shifts). To address this, we propose Self-Predictive Dynamics (SPD), the first method to jointly integrate bidirectional (forward and inverse) dynamics prediction with weak/strong dual-path data augmentation for task-agnostic, robust representation learning. SPD employs contrastive self-supervised modeling to explicitly disentangle task-relevant dynamics from corruption-invariant features. Evaluated on MuJoCo vision-based control and CARLA autonomous driving benchmarks, SPD significantly improves policy generalization across corrupted environments—achieving a 23.6% performance gain over state-of-the-art methods under unseen corruptions. The implementation is publicly available.

7 citationsRead paper

Dream to Generalize: Zero-Shot Model-Based Reinforcement Learning for Unseen Visual Distractions

Jun 26, 2023AAAI Conference on Artificial Intelligence

To address the degradation of model-based reinforcement learning (MBRL) generalization under high-dimensional visual observations corrupted by clouds, shadows, and illumination variations, this paper proposes Dr. G—a zero-shot model-based RL framework. Our approach tackles this challenge through three key contributions: (1) a novel dual-contrastive self-supervised learning mechanism that disentangles and encodes task-relevant features from multi-view augmented data; (2) recurrent state-wise inverse dynamics modeling to enhance the world model’s temporal causal understanding; and (3) zero-shot cross-background transfer without fine-tuning. Evaluated on DeepMind Control (with complex video backgrounds) and Robosuite (with randomized environments), Dr. G achieves performance gains of 117% and 14%, respectively, over state-of-the-art methods. The implementation is publicly available.

5 citationsRead paper

Analysis of Distributional Dynamics for Repeated Cross-Sectional and Intra-Period Observations

May 21, 2025

This paper addresses the challenge of modeling the dynamic evolution of state density functions in two distinct data structures: repeated cross-sections (e.g., monthly stock return distributions) and high-frequency intra-period time series (e.g., intraday GBP/USD return distributions). We propose the first unified functional dynamic framework compatible with both. Methodologically, we embed density functions into a Hilbert space and formulate a functional autoregressive (FAR) model, integrating kernel density estimation with asymptotic statistical inference. Theoretically, we establish asymptotic theory for density forecasting and distributional moment dynamics, and prove strong consistency of the estimators. Empirically, our approach significantly outperforms conventional methods on GBP/USD and NYSE datasets, delivering high-accuracy density forecasts. The core innovation lies in overcoming data-type barriers—enabling, for the first time, unified modeling and theoretical analysis of distributional evolution across both inter-period cross-sectional and intra-period temporal dimensions.

4 citationsRead paper

5G LDPC Codes as Root LDPC Codes via Diversity Alignment

Jan 30, 2026

This work addresses the challenge that 5G NR protograph-based QC-LDPC codes fail to guarantee full diversity for information bits over non-ergodic block-fading channels. To overcome this limitation, the authors propose a Diversity Evolution (DivE) approach, which introduces Boolean functions to model the evolution of variable nodes affected by channel fading during belief propagation decoding. A greedy block-mapping search algorithm is designed to optimize the assignment of variable nodes to fading blocks, thereby achieving full diversity for information bits without altering the underlying 5G base graph structure. Experimental results demonstrate that the proposed mapping significantly improves block error rate performance at high signal-to-noise ratios, yielding a steeper error-rate decay slope.

1 citations1 influentialRead paper

Design of Root Protograph LDPC Codes Simultaneously Achieving Full Diversity and High Coding Gain

Feb 02, 2026

This work addresses the challenge of simultaneously achieving full diversity in block-fading channels and near-Shannon-limit coding gain in additive white Gaussian noise (AWGN) channels—a trade-off that has proven difficult to reconcile. To this end, the paper proposes a unified protograph-based LDPC code design framework. Full diversity is guaranteed through a novel generalized root-check structure, while edge connections are optimized via a density-evolution-guided genetic algorithm to enhance AWGN performance. The authors further introduce the DivE Boolean approximation method and a tailored protograph template for dual-block-fading scenarios. This approach uniquely enables the co-design of full diversity and high coding gain within a single LDPC code structure. Simulation results demonstrate that the resulting codes significantly outperform existing designs in both channel conditions.

1 citationsRead paper
Recent publications

Latest Papers