Reinforced Decision Intelligence Lab

Reliable reinforcement learning and learning-based control with mathematical guarantees for autonomous and industrial systems.

Donghwan Lee, Associate Professor · School of Electrical Engineering · KAIST

Research Identity

How can learning systems make effective decisions while remaining stable, robust, and safe? Our lab develops reinforcement-learning and decision-making methods grounded in control theory, optimization, and applied mathematics. We study fundamental theory and translate it into algorithms for robotics, autonomous driving, energy, and industrial systems

Three Research Pillars

Foundations of Reinforcement Learning
We establish convergence, finite-time, and sample-complexity guarantees for Q-learning, temporal-difference learning, and related algorithms, including function approximation, distributed learning, and Markov games.

Safe and Robust Learning-Based Control
We design learning-based controllers that account for disturbances, model uncertainty, unsafe exploration, and stability requirements, with particular interest in Lyapunov and control-theoretic methods.

Autonomous and Industrial Intelligence
We collaborate on robotics, autonomous driving, GPU power management, and industrial process control, connecting theoretical insight with deployable systems.

Selected Highlightss

ICML 2026 Spotlight
Adaptive Policy Backbone via Shared Network
A shared-network policy architecture for efficient adaptation across tasks.

ICLR 2026
Log-Barrier Methods for LP-Based Markov Decision Problems
A theoretical foundation for solving linear-programming formulations of MDPs through log-barrier optimization.
NeurIPS 2025
Shared Q-Network Pretraining for Offline RL
A transition-predictive backbone that improves data efficiency across offline reinforcement-learning algorithms.
NeurIPS 2024
Regularized Q-Learning
A regularized Q-learning method designed to stabilize learning with linear function approximation.
IEEE TAC 2024
Final-Iteration Convergence of Q-Learning
Finite-time error bounds for the final iterate of Q-learning using a switching-system framework.
IEEE T-RO 2025
Active View Planning for Reliable SLAM
Imitation-learning-based view planning to reduce tracking failures over long routes.

Latest News

2026 “Adaptive Policy Backbone via Shared Network” was selected as an ICML 2026 Spotlight paper.
2026 “Analysis of Approximate Linear Programming Solution to Markov Decision Problem with Log Barrier Function” was accepted to ICLR 2026.
2026 The book “Reinforcement Learning: Mathematical Foundation and Algorithmic Principles” was published.
2026 Donghwan Lee was promoted to Associate Professor at KAIST.
2025 “Pretraining a Shared Q-Network for Data-Efficient Offline Reinforcement Learning” was accepted to NeurIPS 2025.
2025 “SADWA” received a Best Paper Award at an ICCV 2025 Workshop.

Collaboration

Collaborate with Us

We welcome academic and industrial collaborations in reinforcement learning, learning-based control, robotics, autonomous systems, and industrial AI. Our lab can contribute algorithm design, mathematical analysis, simulation, and joint validation on real platforms or operational data.

Projects & Collaboration

We build collaborations around clearly defined technical problems. Our goal is to connect rigorous reinforcement-learning and control methods with platforms, data, and operational needs from academic, public-sector, and industrial partners.

Algorithm Development
Reinforcement learning, offline RL, robust RL, safe learning, multi-agent learning, and learning-based control
Theory and Certification
Convergence, finite-time analysis, stability, robustness, Lyapunov methods, and control-theoretic interpretation.
Simulation and Evaluation
Problem formulation, benchmark design, simulation environments, baselines, ablation studies, and quantitative evaluation.
Application Integration
Adaptation to robotics, autonomous driving, industrial processes, energy systems, and computing infrastructure.

Collaboration Areas

Robotics, drones, and autonomous systems
Autonomous-driving perception, reasoning, and control
Industrial AI, process control, energy, and computing systems
Reliable reinforcement-learning algorithms and theory
Multi-agent decision making and coordination
Vision-language models for sequential decision making

New book published

Amazon