(15a) Reinforcement Learning - A Machine Learning Technology for Making Near Optimal Decisions Based on Experience

Reinforcement
Learning (RL) is a Machine Learning technology in which a computer  agent
learns, by trial and error, the best way to accomplish a task [1].  The power
of RL was recently demonstrated by Google DeepMind, when their RL agent AlphaGo
defeated the world Go grandchampion in a five game match [2].  RL has its roots
in research on animal behavior and optimal control.  In the RL paradigm the
agent learns, through interacting with the environment, which actions will
maximize a long-term sum of future rewards.  Recent advances in RL algorithms, deep
neural networks, and computer hardware have enabled a dramatic increase in the scale
and capability of RL technology, allowing for large-scale complex applications
in games, robotics, self-driving vehicles, process control, and other areas.

In this presentation
we first give a brief tutorial on the fundamentals of RL, ending with an
example application on a bipedal robot.  We then compare RL with Model
Predictive Control technology, focusing on advantages and disadvantages for
process control applications. We then discuss implications for the field of
process control, and we demonstrate how RL can be used to manage the base-level
PID controllers in a plant.  In a final section we outline future research
directions aimed at enabling RL technology to be used more effectively in
process control applications.

References

[1] Richard S. Sutton and Adrew G. Barto,
Reinforcement Learning, Second Edition, in progress.

[2] https://www.theverge.com/2017/5/27/15704088/alphago-ke-jie-game-3-result-retires-future