[KSE BK21 Seminar Notice] 옥정슬 박사 (Jungseul Ok at UIUC., USA. Post-doc)_Dec 27 (Thu), 16:00, Multimedia room(#2501)
오는 27일, 미국 UIUC, Post-doc. 옥정슬 박사를 초청하여 아래와 같이 세미나를 진행할 예정이오니, 많은 관심과 참여를 부탁드립니다.
Short bio: Jungseul Ok received the B.S. in 2011, and finished Ph.D program in 2016, in School of Electrical Engineering at KAIST. He is currently working as postdoc at University of Illinois at Urbana-Champaign. In this talk, he will present his works on "optimal exploration in structured reinforcement learning", part of which was just presented at Neurips 2018.
Abstract: Real-world reinforcement learning (RL) suffers from slow learning rate mainly due to the extremely large size of dynamic systems to be controlled. However, for such large systems, humans or recent RL algorithms using function approximations are able to quickly learn the optimal policies. This, perhaps, is because they utilize some known structures on the dynamical systems, for example, taking similar actions at similar states produce similar results. In this work, in order to investigate an optimal use of the structure and its gain, we formulate and analyze regret minimization problems with structure. It turns out that a structure can indeed yield scale-free cost of learning. We demonstrate the significant improvement from structures in various experiments.