Amir-massoud Farahmand
Machine
Learning Researcher
Mitsubishi
Electric Research Laboratories (MERL)
Background
PhD in
Computer Science, University
of Alberta (CS)
(Working with Csaba
Szepesvári and Martin
Jägersand), 2011
NSERC
Postdoctoral Fellow, McGill
University (SCS)
(Working with Doina
Precup), 2011-2014
NSERC
Postdoctoral Fellow, Carnegie
Mellon University (RI)
(Working with J.
Andrew Bagnell), 2014
Research Goal
Very Short – Two perspectives:
-
Use
data not only to predict, but also to control [ML
Perspective]. -
Designing
adaptive situated agent [AI Perspective].
Longer:
In the 21st century, we live in a world where data is abundant. We would like to take advantage of this opportunity to make more accurate and data-driven decisions in many areas of life such as industry, healthcare, business, and government. Even though many machine learning and data mining researchers have developed tools to benefit from “big data”, their methods so far have mostly been about the task of prediction.
My goal, however, is to use data to control, that is, taking actions in an uncertain world with a complex dynamics in order to achieve a long-term goal such as maximizing the relief of a patient with a chronic disease, sustainable management of natural resources, or increasing the comfort of a building’s occupants.
Admittedly, we are not there yet. Theoretical foundations should be laid and technologies must be developed. But I believe that data-driven decision making defines a new era in human civilization, and my research moves us toward that era. For more details about data-driven control and decision making and understanding my contributions, refer to my Research Statement or take a look at my Publications. Also if you have any questions, please feel free to contact me.
News
-
-
Two
papers on Random Projection Filter Bank (RPFB) are
accepted: One at NIPS
2017 (short
version; extended
version with proofs and more detail) and another
at PHM
(Prognostics and Health Management conference) 2017.
Joint work with Sepideh
and Daniel.
Summary: To extract features from a time series,
project it onto the span of randomly generated
stable dynamical filters. Similar to Random Kitchen
Sink, but for dynamical systems. -
Value-Aware
Model Function for Model-based Reinforcement
Learning is published at AISTATS 2017. Joint work with
Andre and Daniel.
Summary: A good model for prediction is not
necessarily a good model for model-based RL as it
ignores the decision problem. How can we incorporate
the decision problem?
-
Two
papers on controlling Partial Differential Equations
(PDE) using reinforcement learning: 1) Learning
to Control Partial Differential Equations:
Regularized Fitted Q-Iteration (CDC 2016), and
2) Deep
Reinforcement Learning for Partial Differential
Equation Control (ACC 2017). Joint work with Saleh,
Daniel,
and Piyush.
Regularized
Policy Iteration with Nonparametric Function Spaces
is published at the Journal of Machine Learning Research
(JMLR), 2016. Joint work with Csaba,
Mohammad,
and Shie.
Summary: Regularized Least Squares Temporal
Difference (LSTD) is introduced and analyzed. The
method is minimax optimal in a large class of
nonparametric function spaces.
-
Academic and Non-academic Tweets
Research Interests
-
Machine
Learning and Statistics: statistical learning theory,
nonparametric algorithms, regularization, manifold
learning, non-i.i.d. processes, online learning
-
Reinforcement
Learning, Sequential Decision Making, and Control:
high-dimensional problems, regularized nonparametric
algorithms, inverse optimal control
-
Robotics:
uncalibrated visual servoing, learning from demonstration,
behaviour-based architecture for robot control
-
Industrial
Applications: hybrid vehicle energy management
-
Evolutionary
Computation: cooperative co-evolution, interaction of
evolution and learning
-
Large-scale
Optimization
