Learning Controllers - From Engineering to AGI

Explore the cutting-edge applications of deep reinforcement learning in autonomous control systems through this 55-minute seminar by Martin Riedmiller at MIT. Delve into the challenges of designing controllers for complex dynamical systems, with a focus on magnetic confinement control of fusion plasma. Discover the "collect & infer" paradigm for reinforcement learning, which offers a novel approach to data collection and exploitation in data-efficient agents. Examine examples of agent designs capable of learning increasingly complex tasks from scratch in both simulated and real-world environments. Gain insights from Riedmiller's extensive experience in machine learning, neuro-informatics, and robotics, including his work with the champion robot soccer team 'Brainstormers'. Learn about the potential of neural reinforcement learning techniques in advancing towards artificial general intelligence (AGI) and their applications in various fields, from fusion energy to locomotion.

Introduction
Mission of DeepMind
Fusion Energy
Control Problem
Challenges
Classical PID Controller
Learning Stages
History
Classical reinforcement learning
Optimize infer
How to collect data
Explore to Offline
Results
Scheduled Auxiliary Control
Sensor Exploration
Locomotion
Examples
Conclusion
Discussion
Question from YouTube
Does anyone have more questions
Latency
Outro