Presentation is loading. Please wait.

Presentation is loading. Please wait.

The Price of information in combinatorial optimization

Similar presentations


Presentation on theme: "The Price of information in combinatorial optimization"β€” Presentation transcript:

1 The Price of information in combinatorial optimization
(10th January, 2018) Sahil Singla (Carnegie Mellon University)

2 Motivation: Setting Up an Oil Drill
Set up One Oil Drill: Multiple potential sites Have Estimates on their Values: Location, size, surveys Conduct Inspections to Find Exact Value Pay price per site Which Sites Should you Inspect? Want to Maximize Max[value of inspected site] - Total inspection price Similar Examples Purchasing a company Purchasing a house

3 Pandora’s Box How to generalize? Maximize Expected Utility
Given independent distributions on values: 𝑋 1 , 𝑋 2 , .. , 𝑋 𝑛 Given probing prices: πœ‹ 1 , πœ‹ 2 ,…, πœ‹ 𝑛 Adaptively find π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘βŠ†{1,2,…,𝑛} Maximize Expected Utility max π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ {𝑋 𝑖 } βˆ’ π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Price of Information Price of 1 per-box X1~Unif(0,10) X1= 4 X2~Exp(0.5) X2=πŸ” X3~2 X3=𝟐 X4~Unif(0,10) X4=9 Value = πŸ” Price = πŸ‘ Value = πŸ” Price = 𝟐 Value = 𝟐 Price = 𝟏 Picked How to generalize? Weitzman’s optimal policy

4 Max-/Min-Weight Spanning Tree
Given distributions on edge values/costs: 𝑋 1 , 𝑋 2 , .. , 𝑋 𝑛 Given probing prices: πœ‹ 1 , πœ‹ 2 ,…, πœ‹ 𝑛 Adaptively find π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘βŠ†{1,2,…,𝑛} Utility Maximization Problem Packing forest constraints: β„‘ max π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘ { π‘–βˆˆπ‘† 𝑋 𝑖 } βˆ’ π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Disutility Minimization Problem Covering spanning-tree constraints: β„‘β€² min π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘β€² { π‘–βˆˆπ‘† 𝑋 𝑖 } + π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Price of Information

5 Price of Information (PoI)
Given distributions on values/costs: 𝑋 1 , 𝑋 2 , .. , 𝑋 𝑛 Given probing prices: πœ‹ 1 , πœ‹ 2 ,…, πœ‹ 𝑛 Adaptively find π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘βŠ†{1,2,…,𝑛} Utility Maximization Problem Packing constraints: β„‘ max π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘ { π‘–βˆˆπ‘† 𝑋 𝑖 } βˆ’ π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Disutility Minimization Problem Covering constraints: β„‘β€² min π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘β€² { π‘–βˆˆπ‘† 𝑋 𝑖 } + π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Price of Information Examples Max Matroid Basis Maximum Matching Knapsack Examples Min Matroid Basis Vertex/Set Cover Feedback Vertex Set Facility Location Prize-Collecting Steiner Tree Semiadditive functions

6 Main Result Frugal β‰ˆ β€œGreedy” Theorem 1: For any Packing/Covering Problem, an πœΆβ€“approx Frugal alg in the Free-Info World implies an πœΆβ€“approx strategy in the PoI World. Think of Max Wt Matching with 𝛼=2 Problem Approx Ratio Max/Min Matroid Basis 1 Max Matching 2 Max π‘˜-system π‘˜ Max Knapsack Min Vertex-/Set-Cover min{f,log n} Min Facility Location 1.861 Min Prize Collecting Steiner Tree 3 Feedback Vertex Set O(log n)

7 OUTLINE Pandora’s Box and Price of Information Intuitive Examples
Bounding the Optimum Strategy Using a Frugal Algorithm to Design a Strategy Extensions and Conclusions

8 Optimal Adaptive Strategy
Assume Bernoulli Variables: 𝑿 π’Š = 𝒗 π’Š w.p. 𝒑 π’Š 𝟎 otherwise No Yes 𝑋 1 𝑋 5 𝑋 2 𝑋 3 𝑋 4 𝑋 6 Adaptive Decision Tree Difficult to Find: Can be Exponential Sized Want Simple Optimal/Approximate Strategy Not comparing to hindsight optimum

9 Given Distributions : 𝑋 1 , 𝑋 2 , .. , 𝑋 𝑛
Given Probing Prices: πœ‹ 1 , πœ‹ 2 ,…, πœ‹ 𝑛 Adaptively find π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘βŠ†{1,2,…,𝑛} Maximize Expected Utility max π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ {𝑋 𝑖 } βˆ’ π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Examples Each has price 𝟏 𝑿 𝟏 , 𝑿 𝟐 ,…, 𝑿 𝟏𝟎𝟎𝟎 = 𝟐𝟎𝟎 w.p. 𝟎.𝟎𝟏 𝟎 otherwise 𝑿 𝟏 has price πŸπŸ—πŸ– and others have price 𝟏 𝑿 𝟏 =𝟐𝟎𝟎 w.p. 𝟏 & 𝑿 𝟐 , 𝑿 πŸ‘ ,…, 𝑿 𝟏𝟎𝟎𝟎 = 𝟐𝟎𝟎 w.p. 𝟎.𝟎𝟏 𝟎 otherwise Weitzman’s Index 𝝉 π’Š is solution to 𝐸 𝑋 𝑖 βˆ’ 𝜏 𝑖 + = πœ‹ 𝑖 Open in decreasing index order E[Opt] β‰ˆπŸπŸŽπŸŽ Simple Stopping rule: βˆ€π’Š: 𝐸 π‘€π‘Žπ‘Ÿπ‘”π‘–π‘›π‘Žπ‘™(𝑖) =𝐸 (𝑋 𝑖 βˆ’π‘π‘’π‘Ÿπ‘Ÿπ‘’π‘›π‘‘ + ]≀ πœ‹ 𝑖 NaΓ―ve Greedy order can be bad: π‘Žπ‘Ÿπ‘”π‘šπ‘Ž π‘₯ 𝑖 (𝐸 (𝑋 𝑖 βˆ’π‘π‘’π‘Ÿπ‘Ÿπ‘’π‘›π‘‘ + ]βˆ’ πœ‹ 𝑖 ) How to prove?

10 Proof Idea Theorem 1 (Utility Maximization):
Any πœΆβ€“approx Frugal alg in Free-Info World implies an πœΆβ€“approx strategy in the PoI World. Three Steps: E[Optimal Adap in PoI] ≀ E[Optimal Surrog in Free-Info] Optimal Surrog in Free-Info ≀ 𝜢 Γ— (Frugal Alg in Free-Info) E[Frugal Alg in Free-Info] = E[Frugal Strategy in PoI] ⟹ E[Optimal Adap in PoI] ≀ 𝜢 Γ— E[Frugal Strategy in PoI]

11 OUTLINE Pandora’s Box and Price of Information Intuitive Examples
Bounding the Optimum Strategy Using a Frugal Algorithm to Design a Strategy Extensions and Conclusions

12 The Surrogate Utility Maximization Problem
E[Optimal Adap in PoI] ≀ E[Optimal Surrog in Free-Info] Optimal Surrog in Free-Info ≀ 𝜢 Γ— (Frugal Alg in Free-Info) E[Frugal Alg in Free-Info] = E[Frugal Strategy in PoI] The Surrogate Utility Maximization Problem Index 𝜏 𝑖 is solution to 𝐸 𝑋 𝑖 βˆ’ 𝜏 𝑖 + = πœ‹ 𝑖 Surrogate π‘Œ 𝑖 = min 𝑋 𝑖 , 𝜏 𝑖 Pandora’s Box: 𝐄[𝑂𝑝𝑑]≀𝐄[max{ π‘Œ 1 , π‘Œ 2 ,…, π‘Œ 𝑛 }] Packing Lemma: 𝐄[𝑂𝑝𝑑]≀𝐄[ max & 𝑆 ∈ β„‘ ⁑{ π‘–βˆˆπ‘† 𝑛 π‘Œ 𝑖 }] Example: Single box 𝑋 1 𝐸[𝑂𝑝𝑑]=𝐸 𝑋 1 βˆ’ πœ‹ 1 𝐸 π‘Œ 1 = E[min 𝑋 1 , 𝜏 1 ]=𝐸[ 𝑋 1 βˆ’ 𝑋 1 βˆ’ 𝜏 ] =𝐸 𝑋 1 βˆ’ πœ‹ 1 Similar definitions for disutility minimization Better than E[𝑂𝑝𝑑]≀𝐸[max{ 𝑋 1 , 𝑋 2 ,…, 𝑋 𝑛 }]

13 Packing Lemma Lemma: 𝐄[𝑂𝑝𝑑]≀𝐄[ max & 𝑆 ∈ β„‘ ⁑{ π‘–βˆˆπ‘† π‘Œ 𝑖 }]
E[Optimal Adap in PoI] ≀ E[Optimal Surrog in Free-Info] Optimal Surrog in Free-Info ≀ 𝜢 Γ— (Frugal Alg in Free-Info) E[Frugal Alg in Free-Info] = E[Frugal Strategy in PoI] Packing Lemma Lemma: 𝐄[𝑂𝑝𝑑]≀𝐄[ max & 𝑆 ∈ β„‘ ⁑{ π‘–βˆˆπ‘† π‘Œ 𝑖 }] Proof: Let 𝑨 π’Š and 𝟏 π’Š denote Opt picking and Opt probing, resp. 𝐄 𝑂𝑝𝑑 = 𝐄 𝑖 ( 𝑨 π’Š 𝑋 𝑖 βˆ’ 𝟏 π’Š πœ‹ 𝑖 ) = 𝐄 𝑖 ( 𝑨 π’Š 𝑋 𝑖 βˆ’ 𝟏 π’Š 𝑋 𝑖 βˆ’ 𝜏 𝑖 ) ≀ 𝐄 𝑖 ( 𝑨 π’Š 𝑋 𝑖 βˆ’ 𝑨 π’Š 𝑋 𝑖 βˆ’ 𝜏 𝑖 ) =𝐄[ 𝑖 ( 𝑨 π’Š π‘Œ 𝑖 )] ≀ 𝐸[ max & 𝑆 ∈ β„‘ ⁑{ π‘–βˆˆπ‘† π‘Œ 𝑖 }] 𝐄 𝑋 𝑖 βˆ’ 𝜏 𝑖 + = πœ‹ 𝑖 π‘Œ 𝑖 = min 𝑋 𝑖 , 𝜏 𝑖 since 𝑨 π’Š ≀ 𝟏 π’Š Q.E.D. Note: Proof similar to [Kleinberg et al. EC’16]

14 OUTLINE Pandora’s Box and Price of Information Intuitive Examples
Bounding the Optimum Strategy Using a Frugal Algorithm to Design a Strategy Extensions and Conclusions

15 Frugal Alg An ALG in Free-Info World is Frugal Packing if Examples
E[Optimal Adap in PoI] ≀ E[Optimal Surrog in Free-Info] Optimal Surrog in Free-Info ≀ 𝜢 Γ— (Frugal Alg in Free-Info) E[Frugal Alg in Free-Info] = E[Frugal Strategy in PoI] Frugal Alg An ALG in Free-Info World is Frugal Packing if βˆƒ marginal-value function 𝑔 𝒀 𝑀 , 𝑖, π‘Œ 𝑖 β‰₯0 increasing in π‘Œ 𝑖 In an iteration, let j have largest marginal (β€œbest”) If j can be picked then pick j irrevocably: 𝑀=𝑀βˆͺ𝑗 Else, discard j forever Think of Max Wt Matching: 𝑔( 𝒀 𝑀 ,𝑖, π‘Œ 𝑖 )= π‘Œ 𝑖 Examples Matroids or Matching: 𝑔( 𝒀 𝑀 ,𝑖, π‘Œ 𝑖 )= π‘Œ 𝑖 Set cover: 𝑔( 𝒀 𝑀 ,𝑖, π‘Œ 𝑖 )= 1 π‘Œ 𝑖 ( βˆͺ π‘—βˆˆπ‘€βˆͺ𝑖 𝑆 𝑗 βˆ’ βˆͺ π‘—βˆˆπ‘€ 𝑆 𝑗 ) Observations Marginal 𝑔 𝒀 𝑀 , 𝑖, π‘Œ 𝑖 is independent of unseen elements Captures primal-dual algos without β€œcleanup” phase

16 Frugal Strategy Strategy Proof Idea
E[Optimal Adap in PoI] ≀ E[Optimal Surrog in Free-Info] Optimal Surrog in Free-Info ≀ 𝜢 Γ— (Frugal Alg in Free-Info) E[Frugal Alg in Free-Info] = E[Frugal Strategy in PoI] Frugal Strategy Strategy Use index for unprobed elements to compute marginal In an iteration, let j have largest marginal (β€œbest”) If j cannot be picked, discard j forever Else, if j is unprobed then probe it otherwise, j is already probed and pick it Proof Idea Couple both worlds: Same expected change in value Although no price in Free-Info, they get lower value 𝑔( 𝒀 𝑀 ,𝑖, 𝜏 𝑖 ) π‘Œ 𝑖 = min 𝑋 𝑖 , 𝜏 𝑖

17 Applications Problem Approx Ratio Max/Min Matroid Basis 1 Max Matching
2 Max π‘˜-system π‘˜ Max Knapsack Min Vertex-/Set-Cover min{f,log n} Min Facility Location 1.861 Min Prize Collecting Steiner Tree 3 Feedback Vertex Set O(log n)

18 OUTLINE Pandora’s Box and Price of Information Intuitive Examples
Bounding the Optimum Strategy Using a Frugal Algorithm to Design a Strategy Extensions and Conclusions

19 General Functions Disutility Minimization
Covering constraints β„‘β€²: min π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘β€² { π‘–βˆˆπ‘† 𝑋 𝑖 } + π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 More generally, min π‘†βŠ†π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ & 𝑆 ∈ β„‘β€² {π‘π‘œπ‘ π‘‘(𝑆, 𝑋 )} + π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Semiadditive Functions: π‘π‘œπ‘ π‘‘ 𝑆, 𝑋 = π‘–βˆˆπ‘† 𝑋 𝑖 +β„Ž(𝑆) 1 Facility location: β„Ž 𝑆 = π‘—βˆˆπ‘π‘™π‘–π‘’π‘›π‘‘π‘  min π‘–βˆˆπ‘† 𝑑(𝑖,𝑗) Examples Min Matroid Basis Vertex/Set Cover Feedback Vertex Set Examples Facility Location Prize-Collecting Steiner Tree β„Ž: 2 𝑉 β†’ 𝑅 β‰₯0 independent of 𝑋

20 Constrained Utility Max
Consider Pandora’s Box Besides Prices, Allowed to Probe at most k items How to Adaptively probe a set π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘, s.t. π‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ β‰€π‘˜ Maximize Expected Utility max π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ {𝑋 𝑖 } βˆ’ π‘–βˆˆπ‘ƒπ‘Ÿπ‘œπ‘π‘’π‘‘ πœ‹ 𝑖 Use Small Adaptivity Gap for Stoch Probing [GNS’17] Find NA Soln: Submod max over knapsack Convert NA soln to PoI strategy: Use Pandora’s Box Other applications: Set Probing Problem More generally, any packing constraint

21 Open Problems Problem 1: Any Interesting Algorithms Beyond Frugal?
Min s-t cut in the PoI-World Shortest s-t Path in the PoI-World Problem 2: Any Hardness Results for Max-Matching in the PoI-World? Problem 3: Frugal Algorithms with Better Approx Factors? Feedback Vertex Set Max Matching Problem 4: How to Learn and Optimize from Samples when Probability Distributions Not Known?

22 Summary Questions? What is Price of Information
Pandora’s Box Many packing/covering problems Frugal Algorithms Suffice OPT bounded by a random surrogate in Free-Info world Run Frugal algorithm in PoI world using index Extensions Constrained Utility Max and Set Probing Markov Chains by generalizing Gittins index policies Open Problems Beyond Frugal algos and prove Hardness results Questions?

23 References I. Dumitriu, P. Tetali, and P. Winkler. `OnΒ playing golfΒ withΒ two balls’. SIDMA’03. A. Gupta, H. Jiang, Z. Scully, and S. Singla.Β ` The Markovian Price of Information 'Β . In Preparation. A. Gupta, V. Nagarajan, and S. Singla.Β `Adaptivity Gaps for Stochastic Probing: Submodular and XOS Functions'Β . SODA’17. A. Gupta, V. Nagarajan, and S. Singla.Β `Algorithms and Adaptivity Gaps for Stochastic Probing'Β . SODA’16. R. Kleinberg, B. Waggoner, and E. Glen Weyl.Β `Descending Price Optimally Coordinates Search'Β . EC’16. S. Singla. `The Price of Information in Combinatorial Optimization'. SODA’18. M. L. Weitzman. `Optimal Search for the Best Alternative’. Econometrica’79.


Download ppt "The Price of information in combinatorial optimization"

Similar presentations


Ads by Google