Norms! #465

joehuchette · 2015-06-14T02:45:44Z

Still need to plug the solvers in.

mlubin · 2015-06-14T05:23:18Z

src/JuMP.jl

+# Container for expressions containing Norms and AffExprs
+type NormExpr
+    norm::Norm
+    coeff::Float64


What is coeff? Needs more comments

coeff*||norm.terms|| + aff

Okay, just add that as a comment :)

mlubin · 2015-06-14T05:25:00Z

Nice! What do you think about switching to the conic interface if norms are present? I guess this would need the SDP branch merged in first.

joehuchette · 2015-06-14T05:32:44Z

I was thinking this gets merged before SDP, so I can add the conic stuff for SOC there. The default solver stuff for conic problems needs to be rethought at some point.

mlubin · 2015-06-14T05:36:47Z

Sounds good.

mlubin · 2015-06-14T05:37:23Z

Also, do we need to store the coeff term? Why not just distribute it into the norm?

joehuchette · 2015-06-14T05:40:29Z

If it's negative, we want to recognize that something like -||x^2+y^2|| >= -1 is SOC representable.

mlubin · 2015-06-14T05:47:12Z

Or we could not and let users write it with a positive coefficient. We're not Convex.jl after all

joehuchette · 2015-06-14T05:49:35Z

With the way we parse expressions, I think something like 1 <= norm([x,y]) might get rewritten that way regardless

mlubin · 2015-06-14T05:54:47Z

Fair enough

IainNZ · 2015-06-14T17:12:49Z

Any hope for other norms? Or are we committing ourselves to L2?

joehuchette · 2015-06-14T17:24:09Z

It should be pretty easy to make Norm parametric and handle arbitrary norms.

tkelman · 2015-06-14T18:02:58Z

1 and inf norms would be super useful (yes I know it's easy to reformulate those but the code looks prettier when I don't have to)

joehuchette · 2015-06-14T18:13:12Z

Agreed. What would nice syntax look like? norminf and norm1?

tkelman · 2015-06-14T18:17:22Z

That would probably be good enough, yeah. Extra args inside norm{foo, p} might be ambiguous with other ways you use commas inside braces. 1, 2, and inf norms should cover most use cases to start with.

mlubin · 2015-06-14T20:46:32Z

I would make the syntax something like norm2{} instead of norm{}.

IainNZ · 2015-06-14T22:39:00Z

This is some back-seat designing, but could be good to have NormConstraint{T} be a constraint on a Norm{T}, and then its m.normconstrs etc.
This is going to be a nice addition.

joehuchette · 2015-06-14T23:13:36Z

Good point, that doesn't generalize nicely with arbitrary norms.

IainNZ · 2015-06-15T02:47:38Z

src/JuMP.jl

@@ -474,6 +478,40 @@ end
 getValue(arr::Array{QuadExpr}) = map(getValue, arr)

 ##########################################################################
+# Norm
+# Container for √(∑ expr)
+type Norm{T}


Actually, can we make this parametric on the expression:

type Norm{T,E<:GenericAffExpr} terms::Vector{E} end

So I can use it for JuMPeR?

iamed2 · 2015-06-25T19:16:22Z

+1

tkelman · 2015-07-12T22:11:40Z

cc @smukoehler

mlubin · 2015-07-14T13:02:16Z

@joehuchette, could you rebase?

IainNZ · 2015-07-20T20:25:31Z

src/macros.jl

 for (mac,sym) in [(:LinearConstraints, symbol("@LinearConstraint")),
-                  (:QuadConstraints,   symbol("@QuadConstraint"))]
+                  (:QuadConstraints,   symbol("@QuadConstraint")),
+                  (:SOCConstraints,   symbol("@SOCConstraint"))]


Alignment, if you care

mlubin · 2015-07-27T15:18:56Z

I think this needs an update for the copy changes?
I just reran travis and got a failure: https://travis-ci.org/JuliaOpt/JuMP.jl/jobs/71783825#L429

mlubin · 2015-07-27T15:19:45Z

Nvm, master is broken

joehuchette · 2015-07-27T18:02:27Z

OK, I'm gonna merge this in the next day or so if there aren't any outstanding concerns. Operator coverage should be 100%, and the printing code is ugly but functional.

joehuchette · 2015-07-27T18:03:30Z

Conic duals don't work either, but I think that'll take more work throughout the stack that's sufficiently orthogonal from this PR.

mlubin · 2015-07-27T23:50:40Z

test/sdp.jl

-        @defVar(m, L1[1:d])
-        @addConstraint(m, L1 .== (μ-μhat))
-        @addConstraint(m, sum{L1[i]^2, i=1:d} <= t1^2)
-        @addConstraint(m, t1 <= Γ1(𝛿/2,N))


Could we leave these in as a separate test? Always good to test multiple ways of doing the same thing

mlubin · 2015-07-27T23:52:05Z

LGTM.
I do think we should have duals for the 0.10 release, at least for the conic solvers which support it.

mlubin · 2015-07-28T01:24:14Z

Printing tests are failing on travis

joehuchette · 2015-07-28T01:31:06Z

Yeah, it's gonna be a mess to fix...

mlubin · 2015-07-28T01:31:30Z

What's wrong?

joehuchette · 2015-07-28T01:36:09Z

It looks like p[end] and last(p) don't work for p::Pair. We'll probably have to rewrite JuMPContainer iteration to match, as well.

mlubin · 2015-07-28T01:36:59Z

Is this a recent change in master or just this branch?

joehuchette · 2015-07-28T01:37:36Z

Master as well, I'm pretty sure

mlubin · 2015-07-28T01:51:49Z

Oh fun. Maybe we can propose to add these definitions for Pair?

mlubin · 2015-07-28T01:52:43Z

Pair doesn't really fit with our iteration syntax either.

joehuchette · 2015-07-28T01:53:05Z

JuliaLang/julia#12335

joehuchette · 2015-07-28T01:53:21Z

I can add endof too

joehuchette · 2015-07-28T16:52:51Z

OS X release is failing Travis because of an old build, I ran it locally and everything passes.

Norms!

mlubin · 2015-07-28T17:02:14Z

💯

IainNZ · 2015-07-28T17:28:52Z

Good stuff Joey!

mlubin · 2015-07-28T17:30:50Z

doc/refmodel.rst

@@ -105,6 +105,17 @@ you would a linear objective::

    status = solve(m)

+Second-order cone constraints


Shouldn't this be in the "Expressions and Constraints" section?

Yeah, that'd make sense

mlubin reviewed Jun 14, 2015
View reviewed changes

IainNZ reviewed Jun 15, 2015
View reviewed changes

joehuchette force-pushed the norms! branch 2 times, most recently from 73f3f68 to 0623f06 Compare June 30, 2015 17:25

joehuchette force-pushed the norms! branch from 845c1dd to 85eab81 Compare July 14, 2015 13:54

IainNZ reviewed Jul 20, 2015
View reviewed changes

joehuchette force-pushed the norms! branch from c7c6069 to b2015a4 Compare July 27, 2015 18:39

mlubin reviewed Jul 27, 2015
View reviewed changes

initial implementation for norms

64a0e7d

joehuchette force-pushed the norms! branch from 82d854a to 64a0e7d Compare July 28, 2015 16:05

joehuchette added a commit that referenced this pull request Jul 28, 2015

Merge pull request #465 from JuliaOpt/norms!

b2adb58

Norms!

joehuchette merged commit b2adb58 into master Jul 28, 2015

mlubin reviewed Jul 28, 2015
View reviewed changes

mlubin mentioned this pull request Jul 28, 2015

implement norm{} syntax for SOC constraints #434

Closed

IainNZ deleted the norms! branch September 5, 2015 15:51

YDHsieh mentioned this pull request Apr 18, 2018

Errror when trying to use NLexpression #1239

Closed

		@@ -105,6 +105,17 @@ you would a linear objective::

		status = solve(m)

		Second-order cone constraints

Norms! #465

Norms! #465

Conversation

joehuchette commented Jun 14, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mlubin commented Jun 14, 2015

joehuchette commented Jun 14, 2015

mlubin commented Jun 14, 2015

mlubin commented Jun 14, 2015

joehuchette commented Jun 14, 2015

mlubin commented Jun 14, 2015

joehuchette commented Jun 14, 2015

mlubin commented Jun 14, 2015

IainNZ commented Jun 14, 2015

joehuchette commented Jun 14, 2015

tkelman commented Jun 14, 2015

joehuchette commented Jun 14, 2015

tkelman commented Jun 14, 2015

mlubin commented Jun 14, 2015

IainNZ commented Jun 14, 2015

joehuchette commented Jun 14, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

iamed2 commented Jun 25, 2015

tkelman commented Jul 12, 2015

mlubin commented Jul 14, 2015

Choose a reason for hiding this comment

mlubin commented Jul 27, 2015

mlubin commented Jul 27, 2015

joehuchette commented Jul 27, 2015

joehuchette commented Jul 27, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mlubin commented Jul 27, 2015

mlubin commented Jul 28, 2015

joehuchette commented Jul 28, 2015

mlubin commented Jul 28, 2015

joehuchette commented Jul 28, 2015

mlubin commented Jul 28, 2015

joehuchette commented Jul 28, 2015

mlubin commented Jul 28, 2015

mlubin commented Jul 28, 2015

joehuchette commented Jul 28, 2015

joehuchette commented Jul 28, 2015

joehuchette commented Jul 28, 2015

mlubin commented Jul 28, 2015

IainNZ commented Jul 28, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment