1 / 22

Sublinear FPTASs for Stochastic Optimization Problems

Sublinear FPTASs for Stochastic Optimization Problems. Nir Halman, HUJI. Based on joint works with D. Klabjan, C-L Lee, M. Mostagir, J. Orlin and D. Simchi-Levi. FPTASs.

Download Presentation

Sublinear FPTASs for Stochastic Optimization Problems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Sublinear FPTASs for Stochastic Optimization Problems Nir Halman, HUJI Based on joint works with D. Klabjan, C-L Lee, M. Mostagir, J. Orlin and D. Simchi-Levi

  2. FPTASs Def: An FPTAS(FullyPoly. Time Approximation Scheme) is a (1+ ε)-apx. alg. that runs in time poly. in |x| and1/ε for every instance x and ε> 0 Major techniques for FPTAS: work with dual DP + •rounding/scaling the data •dominance - omitting states/actions dominated or approximately dominated by other states/actions Woeginger’s framework [Wo00] uses these techniques but does not handle stochastic DP, nor exponential action spaces

  3. Talk Outline • The knapsack problem • Motivation for stochastic and oracle settings • Approximating functions in logarithmic space and time • Applications in the design and analysis of approximation algorithms

  4. Knapsack Problem (KNP) 0/1 knapsack: Given object set {a1,…,aT} with volumes vi, profits pi, and knapsack volume B, find a subset whose total volume ≤B and total profit is maximized DP formulation:zt(It)=max{zt+1(Ii), pt + zt+1(It –vt)}= total profit when considering items t,t+1,…,T, starting with knapsack size It. z1(B)=? Can be calculated in O(TB) time, i.e., pseudopolinomial in input size NP-C, admits an FPTAS [IK75] (by scaling a dual DP)

  5. KNP – oracle setting Nonlinear knapsack: Any integer number xi < M of copies of ai can be assigned with profit pi(xi) and volume vi(xi), where pi and vi are non-decreasing functions. Admits FPTAS when pi concave and vi convex[Ho95], and when pi is explicitly-given piecewise linear [KN09]. OPEN when pi, vi are general non-decreasing oracle functions Fact: Some problems stop being polynomially-solvable in oracle setting (input size logM)

  6. KNP - stochastic setting Def.: Given object set {a1,…,aT} with volumes vi+Di, profits pi, and knapsack volume B, find a subset whose total volume ≤B and total expected profit is maximized 1) Order of items considered is now important… 2) Benefit of adaptivity [DGV04]… 3) Fact: Some polynomially-solvable problems stop being such in stochastic setting

  7. KNP - stochastic ordered setting Def.: Given object sequence (a1,…,aT) with volumes vi+Di, profits pi, and knapsack volume B, find a subset whose total volume ≤B and total profit is maximized. Not known to admit an FPTAS DP form.: zi(Ii) = maxx=0,…, B/viEDi{xpi+zi+1(Ii –x(vi+Di))}, (total profit to gain with remaining Ii volume by ai,…,aT) z1(B)=? State space It=0,…,B;Action space At= 0,…, B/vt ,Profit function gt(I, x, Dt) = xpt , Transition function ft(I, x, Dt) = I– x(vt + Dt). Observation: gt(I, x, Dt) non-decreasing in I and x;ft (I, x, Dt) non-decreasing in I, non-increasing in x. Note that At may be exponential in the input size

  8. Talk Outline • The knapsack problem • Motivation for stochastic and oracle settings • Approximating functions in logarithmic space and time • Applications in the design and analysis of approximation algorithms  

  9. A Question Under what conditions an oracle function φ:D→R+ admits an efficient succinctK-approximation?(assume finite DR and K≥1. Input size = log |D|+log φmax) Def: •φ*is aK-apx. of φif φ(x) ≤ φ*(x) ≤ Kφ(x), x (note that if φhas special structure, φ*does not necessarily) •φ*is succinct if stored in logarithmic space and is efficient if built in logarithmic time and # of oracle queries • φ:D→R isunimodal if x* so φ is decreasing until x* and increasing afterwards

  10. K  K 1. K-approximation Sets Definition [H+08]: Let φ:D→Z+ be unimodal. A K-apx. set of φ is a subset W with argminφ, Dmax, DminW Dand the ratio between the values of φ on each two consecutive points in W is at most K Construction: φ* is the apx. ofφinduced byW if: φ W φ* Question: How small a K-apx. ofφ can be? Answer: Logarithmic in the input size.Moreover, it can be constructed in logarithmic time and # of oracle queries

  11. K-approximation Sets – Cont’ Theorem [H+08]: If φ is either monotone, convex, or unimodal with given argmin, then it admits a succinct and efficient K-approximation function that preservesthe structure of φ

  12. 2. Calculus of K-apx. Functions Calculus of K-apx functions [H+08]: α,β≥ 0, φi*aKi-apx. of φi(summation of apx.) α+βφ1*+φ2* is max{K1,K2}-apx. of α+βφ1+φ2 (minimization of apx.) min{φ1*,φ2*} is max{K1,K2}-apx. of min{φ1,φ2} (composition of apx.) φ1*(ψ1) is a K1-apx of φ1(ψ1) (apx. of apx.) if φ2= φ1* then φ2* is a K1K2-apx. of φ1 Optimality equation: zt(I) =min xAt(I){gt (I, x) +zt+1(ft(I, x))} Corollary [H+08]:(mimization of summation of composition) Let gt*, z*t+1 be L1,L2-apx. functions of (unimodal) gt, zt+1 thenzt*(I)=minxAt(I){gt*(I, x)+z*t+1(ft(I, x))}is a max{L1,L2}-apx of zt(I) Theorem [H+08]:Let Wx,1,Wx,2beKi-apx. sets of gt*, z*t+1. Then zt*(x):=minxWx,1Wx,2{g*t (It, x) +z*t+1(ft(It, x))}is a controlled (general) apx. of zt(It)

  13. Subtraction of approximation Theorem [HOS11]:φ1*aK1-apx. of φ1,φ2*aK2-apx. of φ2. If φ2cφ1for c<1/K1K2, then φ1*-φ2* is a controlled apx. of φ1-φ2 Theorem [HOS11]: Let φi*be anLi-apx. of φ1 and Wi,be aKi-apx. set of φi*, i=1,2. If φ2cφ1for c<1/K1L1L2then z*(I) =max xW1W2{φ1*(ψ1(I, x)) -φ2*(ψ2(I, x))} is a controlled apx. of z(I) =max x{φ1(ψ1(I, x)) -φ2(ψ2(I, x))}

  14. 3. Approximating CDFs Let F() be a CDF of integral non-negative r.v. D bounded by M, i.e., F(x)=Prob(Dx). Let where ψis a monotone non-decreasing step function with break points a1,...,an. We decompose ψas the sum of the 2-step functions ψ1,...,ψn, where ψi=0 for x< ai and is the constant ψ(ai)-ψ(ai-1) otherwise. so it is the sum ofnnon-decreasing functions (n is poly., M is not)

  15. 4. Calculusof K-approximation Sets Focuses on the domain of the functions Unimodal functions: α,β≥ 0,WiisKi-apx. set of φi, ψ:DD,(monotonicity of apx.) every superset of W1 is a K1-apx. set of φ1 (composition of apx.) ψ -1(W1) is a K1-apx set of φ1(ψ) (linearity of apx.) W1 is a K1-apx. set of α+βφ1(maximization of apx.)W1W2 is a max{K1,K2}-apx. set of max{φ1,φ2} Monotone functions (of the same kind):Wiis Ki-apx. set of φi, (summation of apx.) W1 W2 is a max{K1, K2}-apx. set of φ1+φ2(minimization of apx.)W1W2 is a max{K1,K2}-apx. set of min{φ1,φ2} Convex functions: Wiis Ki-apx. set of φi, (summation of apx.) W1 W2 is a max{K1, K2}-apx. set of φ1+φ2

  16. 5. Approximated Access to φ Useful when access to φis either impossible of very costly Theorem: efficient succinct approximation of: • a general φ via a unimodal apx. oracle with given argmin • a monotone φ via a general apx. oracle • a convex φ via a general apx. oracle that maintains the structure ofφ.

  17. 6. Discrete Convexity Example:Let f(x, y) = (x – 2y). It is convex over R2 g1(x) =min yRf(x, y) is convex over R g2(x) =min yZf(x, y) is NOT convex over R!

  18. Talk Outline • The knapsack problem • Motivation for stochastic and oracle settings • Approximating functions in logarithmic space and time • Applications in the design and analysis of approximation algorithms   

  19. g5 z6 g3 g1 g4 g2 z3 z4 z2 z1 z5 Optimization OverFinite Horizon monotone/convex DP’s Example:5 periods DP zi(I)=minx{gi (I,x)+zi+1(fi(I,x))} (*) Theorem: stochastic monotone/convex DP admits an FPTAS Proof: Recursively apply DP equation T times with apx. functions and apx. sets with K=1+ /2T. By the inequality (1+x/n)n<1+2x we get that KT<1+ Theorem [H+09]: Let gt*, z*t+1 be L1,L2-apx. functions of gt, zt+1. Then zt*(x):=minxt Wx,1Wx,2{g*t (It, xt) +z*t+1(ft(It, xt))}is a (general) max{L1,L2,min{K1L1, K2L2}-apx of zt(It) Theorem: efficient succinct apx. of monotone/convex z via general apx. z* that maintains the structure of z.

  20. Our Approach • Modular framework for deriving FPTAss • Functional point of view • Using primal DP formulation • Propagation of error via Calculus of K-apx. functions • Compactify the action space via K-apx. sets • Speed up construction of K-apx. sets via Calculus of K-apx. sets

  21. Future Research • generate new rules for the calculi (“open code” approach) • specialize the calculi to new classes of functions • instead of using exact oracles use FPTASs for them • develop other recursive structures to plug into the framework [HLS09]

  22. References [DGV04]= B.C. Dean, M.X. Goemans, and J. Vondrak[Ho95] = D.S. Hochbaum[IK75] = O.H. Ibara and C.E. Kim [KN09] = S. Kameshwaran and Y. Narahari[Wo00] = G.J. Woeginger

More Related