IEEE Members: Free
Non-members: FreeDuration: 01:02:16
Yuandong Tian (Meta AI, USA), Abstract: Deep Reinforcement Learning (DRL), as a smart search technique that dynamically improves its policy and value estimation based on observation given previous data, has shown human-level or even super-human performance for games such as Go, chess and StarCraft. On the other hand, when applying DRL in real-world applications, new challenges emerge such as effective integration with current working systems, learning representation of large state and action spaces, or even redefining the temporal structure of sequential decision making. In this talk, I will cover our recent works that include learning initial solutions to the existing solver, learning state representations, or even learning the structure of sequential decision itself.