Informatics colloquium 5. 12. Mastering games with deep learning
Informatics colloquium 5. 12. 2017, 14:00 lecture hall D2
doc. RNDr. Tomáš Brázdil, Ph.D., FI MU
Mastering games with deep learning
Abstract: The main theme of the talk is how much a deep learning agent could
learn by playing games against itself. I will start with a historical overview
of TD-Gammon, a famous algorithm for playing backgammon at expert level. In the
main part I will concentrate on the very recent algorithms for playing Go above
master level. I will give an overview of fundamental methods, namely deep
reinforcement learning and Monte Carlo tree search. Finally, I will comment on
new directions in the area, especially the Starcraft II challenge in which the
current algorithms still fail to deliver convincing results.