Loading...

학술행사

Korean AI Association

학술행사
2019년 강화학습 (Reinforcement learning) 하계단기강좌

  >   학술행사   >   국내학술대회

국내학술대회

연사 및 강의소개
▶ Introduction to Reinforcement Learning
 
Abstract:
강화 학습은 기계학습의 하위분야이지만 자동화 된 의사 결정 및 AI를 위해 범용으로 적용할 수 있는 분야입니다. 이 강의에서는 agent가 명시적으로 행동하고 세계와 상호 작용하는 통계 학습 기술을 소개합니다. 사회가 점차적으로 상호작용하는 로봇이나 챗봇과 같이 상호작용하는 agent 그리고 지능적 의사결정에 더 많은 관심이 모아지면서, 학습하는 agent의 중요성과 난제들을 이해하는 하는 것은 오늘날 매우 중요합니다.  이 강에서는 강화학습의 초석이 될 수 있는 Markov Decision Process (MDP), exploration/exploitation tradeoff/decision-making through value function 을 소개 합니다.
유창동 교수(KAIST) Homepage: http://slsp.kaist.ac.kr
▶ Reinforcement Learning with Generalization
 
Abstract:

The theoretical foundation of reinforcement learning rests on solving a Markov Decision Process. This foundation is in deep tension with applications of reinforcement learning which rely heavy on generalization---the ability to successfully learn what to do given never-before-seen circumstances. We've pursued an agenda of adding generalization to reinforcement learning for well over a decade now, resulting in:
  Contextual Bandits: Addressing learning of immediate rewards with generalization.  This is now a service ( http://aka.ms/personalizer) winning the AI system of the year award at IJCAI 2019.

  Learning to Search: An efficient approach to improving on existing policies, with or without simulators.

  Contextual Decision Processes: A new theory directly combining strategic exploration, generalization, and temporal credit assignment which we've used to solve 2^100-sparse problems.
John Langford Homepage: http://hunch.net/~jl/
▶ Model-based reinforcement learning
 
Abstract:
Reinforcement learning (RL) algorithms, a class of iterative methods that solve optimal control problems through self-play, have demonstrated an ability to succeed in a few arduous tasks, emerging as a general framework for decision making in robotics and neuroscience. Recent studies have improved their design for the sake of adaptation and task generalization. The first part of the course provides a concise introduction of model-based RL theory and algorithms. The second part outlines a new approach to model-based RL design with a human-like intelligence, called neuroscience-inspired AI.
이상완 교수(KAIST) Homepage: http://aibrain.kaist.ac.kr
▶ Model Predictive Control
 
Abstract:
Model predictive control (MPC), which is a powerful control method, is nowadays getting more popular with the advent of powerful GPUs. MPC’s popularity is evidenced by the fact that MPC was used by top-ranked teams in the 2019 AlphaPilot Innovation Challenge Qualifiers organized by Lockheed Martin. In this course, we will cover not only  MPC but also dynamic programming and linear-quadratic regulators that are precursors to MPC. We will see a lot of similarity between MPC and  reinforcement learning, which will undoubtedly help us understand reinforcement learning more and better.
장동의 교수(KAIST) Homepage: http://control.kaist.ac.kr
▶ Robot Learning: When Machine Learning Meets Robotics
 
Abstract:
With recent advances in hardware, sensing, and algorithms, we are witnessing the emergence of a new robotics industry. I will present a few examples of new services provided by upcoming service robots, assisting us in the near future in places, such as offices, malls, and homes. But, for a robot to coexist with humans and operate successfully in crowded and dynamic environments, a robot must be able to learn from experiences to act safely and harmoniously with human participants. I will discuss research challenges for service robots and our attempts to address those challenges. In particular, I will present our recent work in foundations in robot learning: nested sparse networks for allowing a single deep neural network to perform multiple tasks in a resource-aware manner and Tsallis reinforcement learning, a unifying framework for maximum entropy reinforcement learning. If time permits, I will describe other research activities in our lab.
오성회 교수(서울대학교) Homepage: http://rllab.snu.ac.kr/