Madeleine Udell and Stephen Boyd Stanford University ICME...
Transcript of Madeleine Udell and Stephen Boyd Stanford University ICME...
![Page 1: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/1.jpg)
Sigmoidal Programming
Madeleine Udell and Stephen BoydStanford UniversityICME Colloquium
5/13/13
![Page 2: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/2.jpg)
Motivation
![Page 3: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/3.jpg)
Optimization on the Obama campaign
−30 −20 −10 0 10 20 30Margin
0
10
20
30
40
50
60
70Ele
ctora
l vote
s
![Page 4: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/4.jpg)
Optimization on the Obama campaign
−30 −20 −10 0 10 20 30Margin
0
10
20
30
40
50
60
70Ele
ctora
l vote
s
won by just enough...
wasted effort
![Page 5: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/5.jpg)
Sigmoidal programming
Define the sigmoidal programming problem
maximize∑n
i=1 fi (xi )subject to x ∈ C
I fi are sigmoidal functions
I C is a convex set of constraints
![Page 6: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/6.jpg)
Sigmoidal functions
A continuous function f : [l , u]→ R is called sigmoidal if it is eitherconvex, concave, or convex for x ≤ z ∈ R and concave for x ≥ z .
concave
convex
z
![Page 7: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/7.jpg)
Examples of sigmoidal functionsI logistic function logistic(x) = 1/(1 + exp(x))I profit function f (x) = (v − x)logistic(αx + β)I admittance function (aproximation to step function)I CDF of any quasiconcave PDF
![Page 8: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/8.jpg)
Bid optimization
I Each i corresponds to an auction.
I vi is the value of auction i .
I pi (bi ) is the probability of winning auction i with bid bi .
I To maximize winnings (in expectation), choose b by solving
maximize∑n
i=1 vipi (bi )subject to b ∈ C.
I C represents constraints on our bidding portfolio.
![Page 9: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/9.jpg)
Convex constraints for auctions
I minimum and maximum bids: l ≤ b ≤ u
I budget constraint:∑n
i=1 bi ≤ B
I sector constraint:∑
i∈S bi ≤ BS
I diversification constraint: ∀|S | > k,
∑i∈S
bi ≥ εn∑
i=1
bi
I and more (intersections are ok!)
![Page 10: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/10.jpg)
The politician’s problem
I Each i corresponds to a constituency (e.g. state,demographic, ideological group).
I pi is # votes in constituency i (e.g. electoral, popular, etc).
I Politician chooses actions y .
I Constituency i prefers actions wi .
I fi (wTi y) is the probability of winning constituency i .
I To win the most votes (in expectation), choose y by solving
maximize∑n
i=1 pi fi (wTi y)
subject to y ∈ C.
I C represents constraints on what actions we are willing or ableto take.
![Page 11: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/11.jpg)
Convex constraints for politicians
I min, max position: l ≤ y ≤ u
I max hrs in day:∑n
i=1 yi ≤ B
I don’t annoy any constituency too much: wTi y ≥ −γ
![Page 12: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/12.jpg)
Multiple peaks
−6 −4 −2 0 2 4 6position on issue
0.0
0.1
0.2
0.3
0.4
0.5expect
ed v
ote
state 1state 2
![Page 13: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/13.jpg)
Which way to go?
−6 −4 −2 0 2 4 6position on issue
0.0
0.1
0.2
0.3
0.4
0.5expect
ed v
ote
F(x) = 0.48F(x) = 0.48 totalstate 1state 2
![Page 14: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/14.jpg)
Sigmoidal programming is NP hard
Reduction from integer linear programming:
find xsubject to Ax = b
x ∈ {0, 1}n
Cast as sigmoidal programming:
maximize∑n
i=1 g(xi ) = xi (xi − 1)subject to Ax = b
0 ≤ xi ≤ 1 i = 1, . . . , n
Optimal value of sigmoidal programming problem is 0 ⇐⇒there is an integral solution to Ax = b
(Also NP-hard to approximate, using reduction from MAXCUT)
![Page 15: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/15.jpg)
Global optimization
mission: find all local maxima
I can’t search every point in the space
I willing to search a lot of boxes — which ones to choose?
![Page 16: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/16.jpg)
Branch and bound
Idea of branch-and-bound method (Lawler and Wood, 1968; Balas1968):
I Partition space into smaller regions Q ∈ Q.
I Compute upper and lower bounds U(Q) and L(Q) on optimalfunction value
f ?(Q) = maxx∈Q
n∑i=1
fi (xi )
in region Q:L(Q) ≤ f ?(Q) ≤ U(Q).
I Repeat until we zoom in on global max:
maxQ∈Q
L(Q) ≤ f ? ≤ maxQ∈Q
U(Q).
![Page 17: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/17.jpg)
Ingredients for branch and bound success
We need methods to
I easily compute upper and lower bounds U(Q) and L(Q),
I choose the next region to split, and
I choose how to split it
so that the bounds become provably tight around the true solutionreasonably quickly.
![Page 18: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/18.jpg)
Ingredients for sigmoidal programmingWe can do it!
I Easily compute upper and lower bounds U(Q) and L(Q) usingconcave envelope of the functions fi .
I Choose the region with highest upper bound as the nextregion to split.
I Split it at the solution to the previous problem along thecoordinate with the highest error.
![Page 19: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/19.jpg)
Lower bound
Computing a lower bound is easy, since any feasible point gives alower bound:
n∑i=1
fi (x′i ) ≤ max
x∈Q
n∑i=1
fi (xi ) ∀x ′ ∈ Q.
![Page 20: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/20.jpg)
Upper bound
Upper bound will require solving a maximization problem.
If we have functions f̂i such that
I fi (xi ) ≤ f̂i (xi ) for every x ∈ Q, and
I f̂i are all concave.
Then∑n
i=1 f̂i (xi ) is also concave, so
maximize∑n
i=1 f̂i (xi )subject to x ∈ C
is a convex optimization problem that can be solved efficiently.Let x? be a solution to this relaxed problem.
![Page 21: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/21.jpg)
Bound f ?(Q)Then we have an upper and lower bound on f ?(Q). Since
fi (xi ) ≤ f̂i (xi ) ∀x , i = 1, . . . , nn∑
i=1
fi (xi ) ≤n∑
i=1
f̂i (xi ) ∀x
maxx∈Q
n∑i=1
fi (xi ) ≤ maxx∈Q
n∑i=1
f̂i (xi )
f ?(Q) ≤ maxx∈Q
n∑i=1
f̂i (xi ),
we have that
n∑i=1
fi (x?i )︸ ︷︷ ︸
L(Q)
≤ f ?(Q) ≤n∑
i=1
f̂i (x?i )︸ ︷︷ ︸
U(Q)
.
![Page 22: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/22.jpg)
Concave envelope
The tightest concave approximation to f is obtained by choosing f̂ito be the concave envelope of the function f .
f̂i = −(−f )∗∗ is the (negative) bi-conjugate of −f , where
f ?(y) = supx∈I
(f (x)− yx)
is the conjugate (also called the Fenchel dual) of f .
![Page 23: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/23.jpg)
Concave envelopeThe concave envelope can be computed by first locating the pointw such that f (w) = f (a) + f ′(w)(w − a). Then the concaveenvelope f̂ of f can be written piecewise as
f̂ (x) =
{f (a) + f ′(w)(x − a) a ≤ x ≤ w
f (x) w ≤ x ≤ b
}.
w
z
![Page 24: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/24.jpg)
Branch
Compute tighter approximation by splitting rectangles wisely.
I optimism: split rectangle Q with highest upper bound
I greed: split along coordinate i with greatest error
I hope: split at previous solution x?i
![Page 25: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/25.jpg)
Convergence
The number of concave subproblems that must be solved toachieve accuracy ε is bounded by
n∏i=1
(⌊(hi (zi )− hi (li )) (zi − li )
ε/n
⌋+ 1
),
where
I Qinit = (l1, u1)× · · · × (ln, un),
I fi (x) =∫ xlihi (t) dt, i = 1, . . . , n, and
I zi = arg max[li ,ui ] hi (x), i = 1, . . . , n.
![Page 26: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/26.jpg)
Prune
maxQ∈Q
L(Q) ≤ f ? ≤ maxQ∈Q
U(Q)
I Q ′ is called active if maxQ∈Q L(Q) ≤ U(Q ′).
I Otherwise, the solution cannot lie in Q ′, and we can ignore it.
![Page 27: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/27.jpg)
Example: opposing interests
maximize∑
i=1,2 logistic(xi − 2)
subject to∑
i=1,2 xi = 0
−6 −4 −2 0 2 4 6position on issue
0.0
0.1
0.2
0.3
0.4
0.5
expect
ed v
ote
F(x) = 0.48 F(x) = 0.48totalstate 1state 2
![Page 28: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/28.jpg)
Example: opposing interestsBlack line is feasible set.
issue 1
−4−2
02
4
issue 2
−4
−2
0
24
expect
ed v
ote
0.2
0.4
0.6
0.8
![Page 29: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/29.jpg)
Example: opposing interests
1st iteration:
Best solution so far Active nodes
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
−6 −4 −2 0 2 4 6position on issue 1
−6
−4
−2
0
2
4
6
posi
tion o
n iss
ue 1
∑ifi(xi) =0.12
![Page 30: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/30.jpg)
Example: opposing interests
2nd iteration:
Best solution so far Active nodes
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
0.5
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
−6 −4 −2 0 2 4 6position on issue 1
−6
−4
−2
0
2
4
6
posi
tion o
n iss
ue 1
∑ifi(xi) =0.48
∑ifi(xi) =0.44
![Page 31: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/31.jpg)
Example: opposing interests
3rd iteration:
Best solution so far Active nodes
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
−4 −2 0 2 40.0
0.1
0.2
0.3
0.4
0.5
−6 −4 −2 0 2 4 6position on issue 1
−6
−4
−2
0
2
4
6
posi
tion o
n iss
ue 1
∑ifi(xi) =0.48
∑ifi(xi) =0.48
![Page 32: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/32.jpg)
Bid optimization
maximize∑n
i=1 vi logistic(bi − βi )subject to
∑ni=1 bi ≤ B
b ≥ 0
−6 −4 −2 0 2 4 60.0
0.2
0.4
0.6
0.8
1.0
−6 −4 −2 0 2 4 60.0
0.5
1.0
1.5
2.0
−6 −4 −2 0 2 4 60.00.51.01.52.02.53.0
−6 −4 −2 0 2 4 60.00.51.01.52.02.53.03.54.0
−6 −4 −2 0 2 4 60
1
2
3
4
5
−6 −4 −2 0 2 4 60123456
−6 −4 −2 0 2 4 601234567
−6 −4 −2 0 2 4 6012345678
−6 −4 −2 0 2 4 60123456789
![Page 33: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/33.jpg)
Bid optimization
Simulate for vi ∼ U(0, 1), B = 1/5∑n
i=1 vi , βi = 1. To solve toaccuracy .00001 · n takes very few steps! (1 step = 1 LP solve)
n steps
10 220 430 440 650 760 770 680 690 7
100 6
![Page 34: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/34.jpg)
Integer linear programming
maximize∑n
i=1 |xi − 1/2|subject to Ax = b
0 ≤ xi ≤ 1 i = 1, . . . , n
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0constraints (as % of number of variables)
10
20
30
40
50
60
70
80
90
100
variables
ILP subproblem iterations
200
400
600
800
1000
1200
1400
1600
1800
2000
![Page 35: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/35.jpg)
Political positioning: data
I Data comes from 2008 American National Election Survey(ANES).
I Respondents r rate candidates c as having positionsy rc ∈ [1, 7]m on m issues.
I Respondents say how happy they’d be if the candidate wonhrc ∈ [1, 7].
I We suppose a respondent would vote for a candidate c ifhrc > hrc
′for any other candidate c ′.
I If so, v rc = 1 and otherwise v rc = 0.
![Page 36: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/36.jpg)
Political positioning: model
I For each candidate c and state i , we predict that a respondentr ∈ Si in state i will vote for candidate c with probability
logistic((w ci )T y rc),
depending on the candidate’s perceived positions y rc .
I The parameter vector w ci is found by fitting a logistic
regression model to the ANES data for each candidate andstate pair.
Note: only 34 states, some with only 14 respondents . . .
![Page 37: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/37.jpg)
Political positioning: electoral vote
I Suppose each state i has vi votes, which they allocate entirelyto the winner of the popular vote.
I y denotes the positions the politician takes on the issues.
I Then using our model, the politician’s pandering to state i isgiven by xi = wT
i y , and the expected number of votes fromstate i is
vi1(logistic(xi ) > .5),
where 1(x) is 1 if x is true and 0 otherwise.
I Hence the politician will win the most votes if y is chosen bysolving
maximize∑n
i=1 vi1(logistic(xi ) > .5)subject to xi = wT
i y ∀i1 ≤ y ≤ 7.
![Page 38: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/38.jpg)
Optimal solutions to the politician’s problemObama’s optimal pandering:
AL AZ CA CO CT
DC DE FL GA IL
IN KS LA MA MI
MN MS NC ND NJ
NM NV NY OH OK
OR PA RI SC TN
TX VA WA WI
![Page 39: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/39.jpg)
Optimal positions for Obama
Issue Optimal position Previous position
Spending and Services 1.26 5.30Defense spending 1.27 3.69Liberal conservative 1.00 3.29Govt assistance to blacks 1.00 3.12
Issue 1 7Spending and services provide many fewer services provide many more servicesDefense spending decrease defense spending increase defense spendingLiberal conservative liberal conservativeAssistance to blacks govt should help blacks blacks should help themselves
![Page 40: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/40.jpg)
Conclusion
I Sigmoidal programs are a new, non-convex problem class.
I Sigmoidal objectives are ideally suited to model allocationproblems with sigmoidal utilities.
I The problem class is NP-hard to approximate to arbitraryprecision,
I but solutions on interesting problems are often easy tocompute.
![Page 41: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/41.jpg)
Questions?
![Page 42: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/42.jpg)
Network utility maximization
maximize∑n
i=1 fi (xi )subject to Ax ≤ c
x ≥ 0,
I x represent flows,
I c are edge capacities,
I A is the edge incidence matrix, and
I fi (xi ) is the utility derived from flow i .
![Page 43: Madeleine Udell and Stephen Boyd Stanford University ICME Colloquiumweb.stanford.edu/~boyd/papers/pdf/max_sum_sigmoids... · 2013-06-10 · Political positioning: electoral vote I](https://reader033.fdocuments.in/reader033/viewer/2022042117/5e94e783d0de0f234a03c20b/html5/thumbnails/43.jpg)
Network utility maximizationI 500 flows over 500 edges.I Flows use on average 2.5 edges.I Each edge has capacity 2.5.I fi is an admittance function i = 1, . . . , n.
0 2 4 6 8 10 12 14iteration
186
188
190
192
194
196
198
bound
lower boundupper bound