[CTN] meeting with Dr. Tweed

[Matthijs van der Meer]

Dear all,

I still have 2 meeting slots available with Dr. Tweed: 1:45pm and
2:30pm. Additionally, I only have 2 people for lunch. If you are
available to meet with Doug, and/or come to lunch, please let me know!

I'll confirm the lunch meeting place later today after I get some more
replies, but we will meet at noon.

Thanks!

- Matt

-------- Original Message --------
Subject: CTN seminar: Dr. Douglas Tweed, Feb 22nd, 3:30pm, PAS 2464
Date: Fri, 18 Feb 2011 11:41:09 -0500
From: Matthijs van der Meer <mvdm at uwaterloo.ca>
To: ctn at ctnsrv.uwaterloo.ca

Dear all,

Please join us for next Tuesday's CTN seminar by Dr. Douglas Tweed, from
the University of Toronto. Title and abstract are below.

Time and place are the usual, 3.30pm on Tuesday in PAS 2464.

Doug will be returning to Toronto shortly after his talk, but if you
would like go to lunch with him, or meet earlier in the day, please let
me know!

Thanks,

- Matt


Title: Neural-network algorithms for near-optimal control

Abstract:

At least since the time of Helmholtz, theorists have proposed that
neural control systems are designed to optimize certain costs, e.g. for
some arm movements the cost might be time to reach a target. This
optimization approach has been very fruitful in neuroscience: it
provides compact descriptions of complex systems, and it has a strong
record of surprising, correct predictions. But how do neural control
systems optimize themselves? I will review a wish-list for an ideal
optimization algorithm, I will make the point that precisely optimal
control isn't feasible for complex systems, and I will discuss a
powerful method of near-optimal control based on generalized
Hamilton-Jacobi-Bellman (GHJB) equations. Then I will show how that
method can be improved: GHJB hinges on learning a vector field called
the gradient of the cost-to-go, delJ, but it works indirectly in the
sense that it doesn’t learn the best approximation to delJ but instead
learns the time derivative dJ/dt and infers delJ from that. I will show
that we can get simpler and lower-cost controllers by learning delJ
directly, and I will compare this direct algorithm with GHJB on test
problems from the control literature.