(T) I really enjoyed watching this fun video from Professor Sergey Levine on autonomous decision making systems using deep reinforcement learning. Most of the content is somewhat “high level” but require some background in reinforcement learning.
Professor Levine initial premises are that:
- Any component of an autonomous decision making system must improve as more data is provided to the system
- Any component that does not improve will eventually become the bottleneck (so any manual process will become a bottleneck)
- All processes that are in place for the system, to make autonomously decisions must be automated in order to scale
As the result, the fundamental questions that he wants to research toward the insane goal of developing autonomous decision making systems are:
- Can robots learn skills from large amounts of experience that generalize in realistic settings?
- => RL system must transition from learning in closed world settings to open world settings
- How can we make real-world embodied learning fully autonomous?
- => Multi-task and meta-learning could provide the foundation to automate and scale the learning process of the system
- Can reinforcement agents learn from experience without any hand-designed reward function at all?
- => RL system must be able to learn without human specifications of objectives
Note: The picture above is a sculpture in front of the Taipei 101 skyscraper.
Copyright © 2005-2020 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.