Soohyun Choi

PathBridger: Subgoal Bridges for Offline Goal-Conditioned Reinforcement Learning

Soohyun Choi, Seonvin Cho, and Songnam Hong

PathBridger plans across long horizons by selecting an intermediate subgoal, constructing a plausible bridge in state space, decoding the bridge into short executable action chunks with inverse dynamics, and replanning during execution. The method is evaluated on OGBench navigation and manipulation tasks.

Code arXiv

Is One Step Enough for Offline Policy Improvement?

Soohyun Choi^*, Seonvin Cho^*, and Songnam Hong^†

^*Equal contribution · ^†Corresponding author

Multi-step Proximal Policy Improvement (MPI) composes re-centered proximal policy updates. The revised paper distinguishes subdivision of a fixed nominal total horizon T into K stages with local horizon T/K from additional refinement at a common local horizon. It studies how re-centering, geometry, and critic error affect policy improvement, with deterministic TD3+BC and Gaussian IQL extraction experiments.

Code arXiv

Role-Adaptive Policy Optimization for Offline Reinforcement Learning

Seonvin Cho, Soohyun Choi, and Songnam Hong

RAPO learns policy-update coefficients by differentiating through candidate updates and evaluating them according to their role. For TD3+BC, it separates execution and bootstrap actors, using a local improvement surrogate for execution and target stability for bootstrapping. For IQL, it retains the original value updates and adapts only the inverse temperature of advantage-weighted policy extraction. Experiments cover D4RL locomotion and AntMaze tasks.

arXiv

ALARM: Attention and Line Search Based Analog Circuit Optimization on a Riemannian Manifold

Soohyun Choi, Minseok Oh, and Ickhyun Song

ALARM treats circuits with the same designer-selected performance as an equivalence class and combines the induced quotient geometry with an attention-weighted coordinate lift and backtracking line search. In the archived active-inductor VCO study, it met the target in 10/10 runs using 33.2 SPICE evaluations on average.

Technical report Undergraduate capstone project

Extensions of PathBridger

Skill Learning and Composition: Investigating how to learn reusable skills from offline trajectories and compose them for long-horizon goal-conditioned tasks.

Latent-Space Bridge Planning: Exploring learned representations and geometrically structured latent spaces for goal-conditioned bridge generation and short-horizon execution.

REACT: Recursive Algorithm for RC Network Transfer Functions

Soohyun Choi, Joonghoon Lee, Sunggu Lim, Heatbit Park, Songnam Hong, and Jaeduk Han

First-author research conducted through an SK hynix–Hanyang University project. REACT develops a recursive closed-form transfer-function method for arbitrary RC trees and a dominant-pole reduced-order approximation validated against SPICE.