강화학습연구실

We connect reinforcement learning with control theory, optimization, stochastic approximation, and machine learning. This combination allows us to develop new algorithms, explain why they work, and adapt them to autonomous and industrial systems. We study the mathematical foundations of sequential decision making. Our work develops convergence and finite-time analyses for Q-learning, temporal-difference learning, value iteration, and policy-evaluation methods. We are particularly interested in function approximation, target networks, experience replay, distributed learning, multi-agent games, and alternative Bellman operators. Representative topics: switching-system models of Q-learning; finite-time and sample-complexity analysis; stable learning with function approximation; regularized and target-based algorithms; distributed and game-theoretic reinforcement learning. High-performing learning policies are not sufficient when a physical system must remain stable and safe under disturbances, uncertainty, and limited data. We develop reinforcement-learning and control methods that explicitly account for these requirements. Control-theoretic tools provide both algorithmic structure and mathematical certificates. Representative topics: robust policy gradients; disturbance attenuation; minimax reinforcement learning; Lyapunov-based analysis; safe exploration; stability and convergence certificates; learning-based control of dynamical systems. We translate theoretical ideas into decision-making tools for real systems. Current and recent collaborations include quadrotor control, visual SLAM, autonomous-driving perception and reasoning, offline reinforcement learning for GPU power management, and reinforcement-learning-based industrial temperature control. Representative topics