Transcript Binh report

What I am doing?
• Working on a specialized solver for ST and
new nonconvex problems.
• What are the options?
– Traditional : like PATH (next slide)
– Novel method : Reformulate the problem as
convex QP (for linearized friction) or convex SQP
(for quadratic friction)
PATH-like solver (1)
• Newton method with FB function
• Need “crash” procedure : to quickly “guess”
Newton’s initial point (very important). I think
we may be able to exploit physical
interpretations to have a better crash
procedure
PATH-like solver (2)
• Need a good sparse solver for KKT condition
every Newton step. Essentially, just a good
sparse solver for Ax=b.
– It looks simple but actually very complicated:
• A could be large ( in the thousands of columns)
• Taking inverse(A) is out of question (numerical,
inverse(A) may not be sparse anymore, etc..)
• So we will have to solve by factorize
• (Next slides)
PATH-like solver (3)
• Solve Ax=b by factorization
– A = L*U : Good for even non PSD A. <- Can try
this to solve ST
– A = L*D*U: Same but numerically better. Can have
rank-revealing (i.e approximate rank(A) so we can
avoid degeneracy pivot: <- This is what Todd
mentioned in his email)
– A = L*L’ or L*D*L’ (Cholesky factorization) : A must
be PD or Positive Indefinite.
Cholesky factorization
• A = L*L’ or L*D*L’
• This is the one I spent most of my time on. It
should be able to solve our “iterative”
approaches if we reformulate as convex QP.
• I’ve been working on a convex QP solver.
• There are two main factors in QP solver:
– Interior point method or Barrier method
– Factorization
Convex QP solver
• Right now I use OOQP (Steve Wright, Wisc)
which use primal-dual interior point algorithm
of Mehrotra-Gondzio method.
• For factorization: HSL MA57, in process of
adding:
– MA77: one of the best factorization code
– CHOLMOD: 2nd best, free and opensource
– SuperLU: opensource and can run in parallel
– LUSOL: The one in PATH.
Convex QP solver
• Hope you get the big picture
• What can we contribute?
– Normally, 90% of computation lies on factorization
– Every step have to factorize (huge) A.
– We can exploit physical interpretation to factorize
A incrementally. So each step, we only need to
update a small portion of A and still get a good
estimate of A factorization.
– We can have a parallel solver for ST and
nonconvex problem.
What next for me?
• QP solver
• Setup ST and nonconvex methods and
experiment QP solver with it.
What else?
• Prof Steve Wright has been very helpful
• I tried my modified QP solver on a problem
with square matrix of 100.000 columns and it
solves in about 1 minute.
Default QP formulation
min f(x) = c_0 + c^T x + 1/2 x^T Q x
subject to A x = b, d <= Cx <= f
l <= x <= u.