Abstract: In reinforcement learning, an agent tries to learn optimal behavior directly from experience. Reinforcement learning is used in domains such a robotics and games and has allowed a computer to outperform human masters in the games of Go and Chess, using Monte Carlo tree search, and outperform humans in Atari games using model-free reinforcement learning. In this talk, we discuss our recent research work on 1) making Monte Carlo tree search more efficient by controlling mean estimation and 2) framing curriculum reinforcement learning as self-paced learning providing an approach for balancing between learning speed and task difficulty in a principled manner.
Speakers: Joni Pajarinen
Affiliation: Aalto University
Place of Seminar: Zoom