Reinforcement Learning: Turning Big Data into Actions

Reinforcement learning, or approximate dynamic programming, is an approach
for learning how to make sequential optimal decisions, in real time, by
interacting with a stochastic environment. The main idea is that if an action
results in an improved situation, then the tendency to produce that action is
strengthened, i.e. reinforced. Reinforcement learning methods have developed
at the confluence of control theory, neuroscience, animal learning, machine
learning and operations research. Notable success stories include world-best
computer game players for Backgammon and Go, helicopter and robotic control,
computer network design, power plant control.

A key problem in reinforcement learning is dealing with big data, in terms of
a very large or infinite number of environment configurations, many possible
actions, or a very fast sampling and decision rate.  In this talk, I will
present temporal abstraction, an approach in which extended actions are used
to control the environment. I will describe the theoretical framework of
``options", which provides well-founded and efficient learning and planning
algorithms. I will then discuss option discovery, the key remaining open
problem in this area.  I will describe new intuitions and approaches that we
are developing for constructing options.

Bio:
Doina Precup is an associate professor in the School of Computer Science of
McGill University, Montreal. She obtained her BEng degree from the Technical
University Cluj-Napoca, Romania (1994), and her MSc (1997) and PhD (2000)
from the University of Massachusetts, Amherst, where she was a Fulbright
fellow. Her research interests are in the areas of artificial intelligence,
machine learning and applications of these methods to real-world problems.