Page 1
Distributed Arg-max Computation
Jie [email protected]
John MacLaren [email protected]
Adaptive Signal Processing and Information Theory GroupDepartment of Electrical and Computer Engineering
Drexel University, Philadelphia, PA 19104
This research has been supported by the Air Force Research Laboratoryunder agreement number FA9550-12-1-0086.
October 14th, 2015
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 1 / 12
Page 2
Introduction
Outline
1 Introduction
2 Lossless One-way
3 Lossless Interactive
4 Compute Two-way Interactive Rate-regions
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 2 / 12
Page 3
Introduction
Motivation – Resource Allocation in LTE
MS 1
MS 2BS
Encoder
Encoder
Decoder
Subband index
Usersubband
gain1 2 3
Subband index
Usersubband
gain1 2 3
X(1)1 , . . . , X
(M)1
X(1)2 , . . . , X
(M)2
S1
Subband index
Usersubband
gain1 2 3
Z(1), . . . , Z(M)
Z(j) = argmaxn
X(j)1 , X
(j)2
o
Z = g(X1,X2)
Z = f(S1, S2)S2
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 3 / 12
Page 4
Introduction
Problem Model
Z 2 g(XS1 , . . . , XS
N )
limS!1
P (S)e = 0
R1
R2
RN
Enc1
Dec
XS2
XSN
XS1
Enc2
EncN
• Channel capacity: modeled asdiscrete i.i.d. sources
• Assume rateless datatransmission
• The CEO needs to compute{i |Xi = max{Xi : i ∈ [N]}}
• Distortion Measure
dA((X1,s , . . . ,XN,s), ZA(s)) =
{0 if ZA ∈ ZA
ZM(s)− XZA(s),sotherwise
(1)
• Distributed Lossless Computation
E[d((X1, . . . ,XN), Z)
]= 0 (2)
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 4 / 12
Page 5
Lossless One-way
Outline
1 Introduction
2 Lossless One-way
3 Lossless Interactive
4 Compute Two-way Interactive Rate-regions
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 5 / 12
Page 6
Lossless One-way
Lossless One-way Results
• Candidate arg-max functions
RA = minfN∈FA,N
N∑
n=1
mincn∈C(Gn(fN))
H(cn(Xn)) (3)
• Achievability: build f ∗N recursively, graph coloring
• Converse: graph entropy
• i.i.d. sources: do not need the OR-product graph
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 6 / 12
Page 7
Lossless Interactive
Outline
1 Introduction
2 Lossless One-way
3 Lossless Interactive
4 Compute Two-way Interactive Rate-regions
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 7 / 12
Page 8
Lossless Interactive
Problem Model
3 dB
2 dB
2 dB
Ut(�t = 3dB)
X1
X2
X3
V 1t = 1
V 2t = 0
V 3t = 0
Notations
• Xi ∈ Xt = {at , . . . , bt}• Ut Broadcasting message at
round t
• V it Replied message from MS i
at round t
Achievable Interaction Scheme
1: CEO broadcasts a threshold λtat round t
2: User i replies a 1 if Xi ≥ λt and0 otherwise
3: Stops when CEO knows arg-maxreliably
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 8 / 12
Page 9
Lossless Interactive
Analysis
Aggregate rate
Rt(λ) = H(λ|λ1, · · · , λt−1) + Nt + (Ft(λ))NtR∗(Nt , at , λ)
+Nt∑
i=1
(1− Ft(λ))iFt(λ)
Nt−i Nt !
i !(Nt − i)!R∗(i , λ, bt) (4)
Policy Iteration
λ∗t = argminλ
Rt(λ) (5)
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 9 / 12
Page 10
Compute Two-way Interactive Rate-regions
Outline
1 Introduction
2 Lossless One-way
3 Lossless Interactive
4 Compute Two-way Interactive Rate-regions
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 10 / 12
Page 11
Compute Two-way Interactive Rate-regions
BackgroundInteractive Communication
User A User B
X Y
U, Ry
V, Rx
f(x, y)
{(rx, ry) :rx ≥ I(V ; X|U, Y ) ry ≥ I(U ; Y |X)
where U − Y − X and V − (U, X) − Y
with E[d(f(x, y), φ(v, y))] ≤ D}
• Interaction for Lossy SourceReproduction (Kaspi 1985)
• Two-way Interaction FunctionComputation (Orlitsky & Roche2001)
• Interaction for functioncomputation (Ishwar & Ma2011)
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 11 / 12
Page 12
Compute Two-way Interactive Rate-regions
Numerically Compute Rate Regions
BA for user 1
BA for user 2
Update Estimator
Converge?
Init
Update p(u|y)
Update p(u)
Converge
Update p(v|x,u)
Update p(v)
Converge
• Communication order matters
• Interested in minimum sum rate
• Apply Blahut-ArimotoAlgorithm
• Alternating optimization
• Includes Markov chainconstraints
Jie Ren (Drexel ASPITRG) DAC October 14th, 2015 12 / 12